University of Southern California Dissertations and Theses

Knowledge‐driven decision support for assessing radiation therapy dose constraints

(USC Thesis Other)

Knowledge‐driven decision support for assessing radiation therapy dose constraints

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content
KNOWLEDGE-DRIVEN DECISION
SUPPORT FOR ASSESSING
RADIATION THERAPY DOSE
CONSTRAINTS
By Ruchi Deshpande

A dissertation presented to the faculty of the USC Graduate School at the
University of Southern California in partial fulfillment of the requirements for the
degree DOCTOR OF PHILOSOPHY (Biomedical Engineering)

MAY 2016

i
Dedication

To my wonderful parents, Deepa and Rajiv, for inspiring me with their words and actions, and
for giving me their selfless support, unshakeable faith and unconditional love;
my husband Rohan for being my muse, anchor, compass, abettor, jester and champion;
my sister Ramya for the joy and exuberance she brings to my life;
my grandparents Prabhakar, Manda and Rekha for nurturing me with their tender loving care
and fantastic story-telling abilities;
and my best friends Ambika and Krupa for blues successfully vanquished, and advice cannily
given.

ii
Acknowledgements

This work has been done with the partial contribution and collaboration of the following
colleagues and friends:

Dr. Brent Liu PhD Advisor
Dr. John DeMarco Clinical Collaborator and Advisor
Dr. James Sayre Statistical Analysis
Dr. Kerstin Kessel Contribution of treatment planning data
Jing Wang Web-based DICOM Viewer
Wanwara Thuptimdang Client-Side DICOM Parser
Sneha Verma Web-based DICOM Viewer
Dr. Anh Le Initial database design
Jonathan Wong Python DICOM Parser

In addition, I would like to express my immense gratitude to my advisor Dr. Brent Liu for his
constant support, valuable insight and unfailing faith in me; and to Dr. John DeMarco for his
enthusiasm, committed collaboration and endless patience. My warmest thanks go out to Dr.
Bernie Huang for his wisdom and counsel; Anh Le and Jorge Documet for their mentorship in the
early stages of my PhD; Sneha Verma for being a wonderful listener and an even better critique;
and to all my other lab-mates at IPILab, past and present, for their camaraderie, support and
assistance. Last, but certainly not least, a very special thanks to my family and friends without
whose love and support I would never have made it this far.

iii
TABLE OF CONTENTS
Dedication ……………………………………………………………………………………………………………. i
Acknowledgements …………………………………………………………………………………………….. ii
List of Tables ………………………………………………………………………………………………………… v
List of Figures ……………………………………………………………………………………………………….. vi
Abbreviations ……………………………………………………………………………………………………….. vii
Abstract ……………………………………………………………………………………………………………….. viii
Chapter 1 Introduction …………………………………………………………………………………………
1.1 Medical Imaging Informatics for Decision Support in Healthcare ……………..
1.2 Imaging Informatics in Radiation Therapy ………………………………………………..
1.3 Previous Work in Radiation Therapy Informatics ……………………………………..
1.4 Challenges in Radiation Therapy ……………………………………………………………...
1.5 Research Aims and Scope …………………………………………………………………….....
1.6 Chapter Scheme of Dissertation ……………………………………………………………....
1
1
2
3
6
8
9
Chapter 2 Background ………………………………………………………………………………………….
2.1 Head and Neck Cancer ……………………………………………………………………………..
2.2 Radiation Therapy …………………………………………………………………………………….
2.3 Intensity Modulated Radiation Therapy (IMRT) ………………………………………..
2.4 Inverse Treatment Planning (ITP) Workflow in IMRT …………………………………
2.5 Challenges in IMRT Treatment Planning …………………………………………………..
11
11
13
15
16
22
Chapter 3 System Design ………………………………………………………………………………………
3.1 The DICOM Standard ………………………………………………………………………………..
3.2 Data Model ………………………………………………………………………………………………
3.3 System Architecture …………………………………………………………………………………
3.4 Flow of data in the decision support system …………………………………………….
25
26
29
30
34
Chapter 4 Decision Support Module …………………………………………………………………….
4.1 Previous Work in Knowledge Driven Decision Support ……………………………..
4.2 Similarity Features ……………………………………………………………………………………
4.2.1 Overlap Volume Histogram (OVH) ………………………………………………….
4.2.2 Spatial Target Signature (STS) ……………………………………………………......
4.2.3 Target Dose (TD) …………………………………………………………………………….
4.3 Similarity Matching …………………………………………………………………………………..
4.4 System Recommendations ……………………………………………………………………….
4.5 Decision Support Workflow ………………………………………………………………………
4.6 Use Case Scenarios …………………………………………………………………………………..
36
37
39
39
43
47
48
49
51
54
Chapter 5 Evaluation …………………………………………………………………………………………….
5.1 Data Collection …………………………………………………………………………………………
5.2 Validation of the Anatomical Similarity Matching Algorithm …………………….
5.3 Evaluation of the decision support module ………………………………………………
60
60
62
65
iv
5.3.1 Evaluation Per Iteration ………………………………………………………………….
5.3.2 Evaluation for the entire dataset …………………………………………………….
67
69
Chapter 6 Multi-Site Collaborative Data Aggregation ………………………………………….
6.1 Challenges in Implementing Collaborative Data Sharing …………………………..
6.2 RT Data Sharing Research Initiatives …………………………………………………………
6.3 HIPAA-Compliant Data Sharing Protocol …………………………………………………..
6.4 Integration of Multi-Institutional Data ……………………………………………………..
73
73
74
75
77
Chapter 7 Discussion & Conclusion ………………………………………………………………………
7.1 Targeted User Community ………………………………………………………………………..
7.2 Expanding the Scope of Research ……………………………………………………………..
7.3 Challenges Encountered …………………………………………………………………………..
7.4 Suggested System Improvements for Workflow Enhancement …………………
7.5 Scope and Direction of Future Research …………………………………………………..
7.6 Final Thoughts ………………………………………………………………………………………….
79
80
81
81
82
83
84
Bibliography …………………………………………………………………………………………………………. 86
Author Publications and Presentations …………………………………………………………………. 89

v
List of Tables

Table 5-1 Data objects contained within each dataset ……………………………………...... 61
Table 5-2 Characteristics of the dataset used for development and testing …………. 61
Table 5-3 Correlation between artificially introduced dissimilarity represented by
the translation coefficient, and the dissimilarity calculated by the
decision support algorithm ……………………………………………………………………

65
Table 5-4 Success rate of the decision support algorithm in identifying relevant
database cases for comparison ……………………………………………………………..

69
Table 5-5 Tolerance dose compliance rates of Organs At Risk ……………………………… 71
Table 5-6 Trends of volumetric overlap with the PTV for the various Organs At
Risk ……………………………………………………………………………………………………….

71

vi
List of Figures

Figure 1-1
Effect of PTV proximity on risk of OAR overdose ………………………………………………
6
Figure 2-1 Multi-Leaf Collimators ………………………………………………………………………………….. 15
Figure 2-2 Inverse Treatment Planning Workflow in Radiation Therapy …………………. 17
Figure 2-3 Target Volume Definitions …………………………………………………………………………… 18
Figure 2-4 Example of ROI Contours ……………………………………………………………………………… 18
Figure 2-5 2-Dimensional dose color wash …………………………………………………………………… 21
Figure 2-6 Examples of Dose Volume Histograms ………………………………………………………. 21
Figure 3-1 Example of meta-data from a DICOM-RT Plan object ……………………………….. 28
Figure 3-2 A conceptual data model of the decision support system ……………………… 29
Figure 3-3 Architecture of the decision support system ……………………………………………… 31
Figure 3-4 List of modules contained within the DICOM-RT Structure Set object …… 32
Figure 4-1 Steps involved in constructing an OVH for a two-dimensional CT slice ..... 41
Figure 4-2 Examples of Overlap Volume Histograms ………………………………………………..... 42
Figure 4-3 The three spherical coordinates utilized in the Spatial Target Signature .. 43
Figure 4-4 Significance of the Spatial Target Signature (STS) in retrieving cases with
similar dose distributions ……………………………………………………………………….

46
Figure 4-5 (A) DVH Spread of all Database Cases (B) DVH Spread of Similar
Database Cases (C) 95% DVH Confidence Interval of the query case ………

50
Figure 4-6 Contribution of the Decision Support System to the Clinical Workflow …. 52
Figure 4-7 Interactive Plan Assessment Tool used for selecting IMRT dose
constraints ……………………………………………………………………………………………………..
56
Figure 4-8 Interactive Plan Assessment Tool for treatment plan evaluation …………… 57
Figure 4-9 Interactive Plan Assessment Tool for re-adjusting dose constraints ………. 59
Figure 5-1 The three transformations applied in order to introduce artificial
dissimilarity to the test case. ROIs: Parotid glands in dark blue,
brainstem in green and the PTV in red ………………………………………………………..

64
Figure 5-2 Relationship between anatomical similarity and DVH similarity ……………. 66
Figure 5-3 Leave-one-out cross validation ……………………………………………………………………. 66
Figure 5-4 Dose Volume Histogram comparison of anatomically similar and
dissimilar cases ………………………………………………………………………………………………

67
Figure 6-1 A HIPAA-Compliant Data Sharing Protocol …………………………………………………. 76

vii
Abbreviations

3D-CRT 3-Dimensional Conformal Radiation Therapy
ACR American College of Radiology
ASTRO American Society for Radiation Oncology
CT Computed Tomography
CTV Clinical Target Volume
DB Database
DICOM Digital Imaging and Communications in Medicine
DICOM-RT Digital Imaging and Communications in Medicine - Radiation Therapy
DVH Dose Volume Histogram
EMD Earth Mover’s Distance
EMR Electronic Medical Record
ePR Electronic Patient Record
GTV Gross Tumor Volume
GUI Graphical User Interface
HIPAA Health Insurance Portability and Accountability Act
IMRT Intensity Modulated Radiation Therapy
ITP Inverse Treatment Planning
MLC Multi-Leaf Collimator
MRI Magnetic Resonance Imaging
NROR National Registry of Radiation Oncology
OAR Organ At Risk
OVH Overlap Volume Histogram
PACS Picture Archiving and Communication System
PET Positron Emission Tomography
PHI Protected Health Information
PTV Planning Target Volume
QUANTEC Quantitative Analyses of Normal Tissue Effects in the Clinic
ROI Region Of Interest
RT Radiation Therapy
RTOG Radiation Therapy Oncology Group
SIB Simultaneous Integrated Boost
STS Spatial Target Signature
TPS Treatment Planning System
TRIAD Transfer of Images and Data
VMAT Volumetric Arc Therapy

viii
Abstract

Treatment planning in radiation therapy (RT) is a partly subjective process, relying considerably
on the associated clinicians’ personal experience and judgment. This dissertation presents an
informatics system that aims to assist RT treatment planners through quantified knowledge
gained by analyzing retrospective treatment plans. Decision support is provided in the form of
treatment parameter estimates that are likely to help reduce radiation dose to healthy tissue
around the cancer tumor. In order to compute good recommendations, the system learns from
the dose profiles of retrospective patients that are anatomically similar to a new patient. The
pattern recognition features, metrics and techniques employed for this purpose have been
described in this manuscript, along with the results of evaluating their effectiveness. Informatics
applications such as this are often incomplete without a scheme for assimilating the developed
programs into the clinical environment. As a result, this work also focuses on clinical workflow
analysis to help identify weaknesses in the operational workflow, and corresponding informatics
tools and techniques that can provide a resolution through effective data management and
integration. The clinical workflow, data standardization through the adoption of DICOM, the
system architecture and specifications of major components are discussed. Since the knowledge
base forms the crux of the system, the manuscript highlights mechanisms to facilitate data
collection and promote data diversity by enabling multi-institutional collaborations. Four major
innovations presented herein include prototypes of: the Spatial Target Signature (STS), which is
a novel feature for assessing anatomical similarity; a similarity calculation methodology; an
interactive plan assessment tool; and a HIPAA-compliant protocol for facilitating multi-
institutional collaboration.
1
Chapter 1 Introduction
1.1 Medical Imaging Informatics for Decision Support in Healthcare
Advanced medical imaging techniques such as Computed Tomography (CT) and
Magnetic Resonance Imaging (MRI) have contributed to an unprecedented rise in the
production of electronic health data. As a result, there is a growing demand for
solutions that provide effective storage, management and distribution of digital
healthcare data. The field of medical imaging informatics grew in response to this
demand, and gave rise to industry standards, information systems and protocols that
form the backbone of today’s healthcare IT industry. Radiology, being the medical
specialty that depended most heavily on digital health data, provided the first firm
foothold for this emerging discipline. Early research in imaging informatics led to the
evolution of ‘Digital Imaging and Communications in Medicine’ (DICOM) [1] and ‘Picture
Archiving and Communications Systems’ (PACS) [2]. DICOM standardized storage and
communication of medical images, while PACS streamlined clinical workflow in
radiology. Since the unanimous adoption and consequent refinement of PACS and other
information systems, one of the new avenues of imaging informatics research involves
leveraging the vast amounts of data that are routinely stored in these systems. One way
to successfully exploit this data is to search for patterns and develop mathematical
models that can assist clinicians in their decision making process, ultimately improving
patient outcomes. Computer based applications that can learn from a knowledge source
using reasoning or inferential mechanisms, in order to provide constructive guidance to
2
clinicians, are known as decision support systems. My work falls under the domain of
imaging informatics based clinical decision support, applied to the field of Radiation
Therapy.
1.2 Imaging Informatics in Radiation Therapy
As mentioned previously, the traditional domain of imaging informatics research is
Radiology, since it is more heavily dependent on imaging than any other medical
specialty. However, imaging informatics is now expanding to many other fields,
supported by the increasing reliance of other medical specialties on imaging. Some
examples of such fields include pathology, cardiology, surgery, ophthalmology, dentistry
and even physical therapy and rehabilitation engineering. Radiation Therapy is one such
specialty that can benefit from the introduction of imaging informatics tools and
techniques. This is because of its increasing reliance on medical images to provide an
accurate three-dimensional model of the patient’s internal anatomy in order to
ascertain the size, location, spread and spatial distribution of tumors with respect to its
surrounding anatomical structures. This is very important in determining the most
optimal treatment parameters that will ensure delivery of a large and uniform radiation
dose to the tumor, while avoiding radiation exposure of surrounding healthy tissue as
much as possible. A high target dose is necessary for tumor eradication. However,
normal tissue has a low tolerance for radiation, and excessive exposure of surrounding
organs may result in harmful side effects that could drastically lower the patient’s
quality of life post-treatment. Medical imaging provides precise knowledge of every
patient’s individual anatomy, and is hence the key to an effective treatment strategy.
3

The informatics infrastructure in radiation oncology does not yet measure up to the
seamless system integration, workflow efficiency and data management provided by
Radiology information systems. Radiation Therapy data extends beyond images to
objects such as radiation dose distributions, Region Of Interest Contours, histograms,
beam intensities, etc. Although standards do exist for these data objects, most vendors
still use proprietary formats for storage and communication. As a result, it is difficult to
build universal imaging informatics applications that can directly interface and
communicate with products from multiple vendors. Consequently, many research
groups prefer to build applications that are specific to the treatment systems employed
at their institutions. This is an obstacle to integrating data not only from different
countries and institutions, but also from different treatment planning systems in the
same institution.
Despite the lack of infrastructure for systems integration, there is potential for deriving
knowledge from pre-existing treatment planning data and using that knowledge to
power decision support systems that can aid clinicians in developing treatment plans.
Unlike radiology, where the aim of computer-aided systems is to help physicians in
diagnosing problems, the aim of decision support systems in radiation therapy is to help
clinicians plan their treatments, and hopefully promote better outcomes.
1.3 Previous Work in Radiation Therapy Informatics
Imaging Informatics in Radiation Therapy: Previous work by one of IPILab’s long-
standing collaborators, Maria Law, involved a comprehensive analysis of the radiation
4
therapy workflow in an informatics context, thus laying the foundation for future
imaging informatics applications in Radiation Therapy [3]. Further, it dealt with the
derivation of a practical data model for Radiation Therapy objects using DICOM RT and
the design of an electronic patient record system for storage and distribution of
Radiation Therapy data [4, 5]. One such electronic patient record was prototyped and
tested at the IPILab by Anh Le, in collaboration with the Loma Linda University Medical
Center [6, 7]. The goal of this ePR was to integrate data related to proton therapy from
various stand-alone clinical information systems, and to demonstrate the utility of such
an infrastructure in improving patient outcomes by minimizing radiation toxicity in
prostate cancer treatment. My goal was to modify and expand the infrastructure of the
prototype proton therapy ePR system in order to make it more robust, comprehensive,
up-to-date and flexible for eventual collaboration with multiple clinical sites. The new
infrastructure is utilized to serve an entirely new clinical application – assisting clinicians
in picking dose constraints for Radiation Therapy of Head and Neck Cancer. Head and
neck cancer was chosen due to the complexity of disease manifestation in terms of the
number of organs at risk, their proximity to the cancer tumors, as well as the variation in
tumor location. A novel decision support module capable of intelligent image
processing, pattern recognition and data analysis was designed, prototyped and
evaluated. The system infrastructure was also expanded to allow the integration of
radiation therapy data from multiple institutions to facilitate collaboration, and a new
strategy was developed to deal with sharing of protected health information (PHI).
5
Collaborative Data Sharing in Radiation Therapy: There are other research groups
pursuing similar research goals of building infrastructure to facilitate sharing of clinical
RT treatment planning data for the purpose of establishing a knowledge base. Two
examples that stand out are the ‘Transfer of Images and Data’ (TRIAD) platform [8]
developed by the American College of Radiology (ACR) and the ‘National Registry of
Radiation Oncology’ (NROR) [9, 10] sponsored by the American Society of Radiation
Oncology (ASTRO). TRIAD is a medical data-sharing platform that enables exchange of
DICOM and non-DICOM data from various participating sites for the purposes of clinical
trials, accreditations, etc. The Radiation Therapy and Oncology Group (RTOG) has
chosen to use TRIAD in collecting data for its clinical trials. However, TRIAD is intended
to facilitate the flow of data involved in sharing and collecting files and does not address
the need to catalogue any meta-data into a database for easy query and retrieval. This is
one of the key components of my decision support system. NROR, on the other hand,
aims to build a large-scale national registry of treatment delivery and health outcomes
data to fuel a broad range of research and quality assurance activities. However, NROR
may take years to build and implement fully for multiple disease sites and moreover, the
business model that will eventually be used is not yet clear. The aim of my imaging
informatics infrastructure is to build a database on a smaller scale to collect enough
data to test and evaluate decision support algorithms that can later be used by systems
like NROR. Furthermore, the client-side DICOM metadata extraction tool will be made
open source once it is fully developed, so that applications such as TRIAD and NROR may
utilize it to expand their privacy and data sharing business models.
6
Decision Support in Radiation Therapy: The topic of knowledge-driven treatment
planning is currently very popular in Radiation Oncology. However, most existing
applications focus on data analytics, but lack a web-based, vendor-neutral informatics
infrastructure that provides systems and data integration. Chapter 4 reviews previous
work dealing with IMRT decision support, and discusses the unique contributions
provided by my algorithms.
1.4 Challenges in Radiation Therapy
The primary goal in radiation therapy is to target the tumor with maximum possible
radiation dose, while limiting dose to any healthy tissue surrounding the tumor. This is
challenging because tumors are often located in close proximity healthy organs, as
shown in figure 1-1. Part (a) shows an example where
the right parotid (solid green) is at high risk of
radiation exposure since it is situated close to the PTV
(red). Part (b) shows an example where the right
parotid is at lesser risk of over-exposure since it is
situated further from the PTV, and is easier to spare.
The three-dimensional spatial relationship between the tumor and a neighboring
healthy structure (Organ At Risk, or OAR) often limits the probability and extent of being
able to spare that OAR. If the geometrical configuration of the tumor with respect to an
OAR is particularly unfavorable, then it might become counter-productive to comply
with tissue tolerance and safety guidelines. In such cases, eradication of cancer cells
must be given priority over OAR sparing. On the other hand, the spatial location of the
Figure 1-1 Effect of PTV proximity on
risk of OAR overdose
7
tumor may be such that significant OAR sparing is possible. However, it is difficult for
clinicians to estimate exactly how much sparing is achievable. Additionally, practical
considerations often limit the number of trial-and-error iterations that may be
performed to arrive at good estimates.
Training, expertise, personal experience and judgment on the part of clinicians can often
compensate for these uncertainties in treatment planning. However, computer-based
tools that provide relevant evidence-based inputs during key decision-making points in
the workflow may facilitate the process, and help relieve some of the burdens
associated with treatment planning. I have designed and developed a system with
decision support tools that provide clinicians with knowledge regarding the possible
dose profiles that a new case may attain, based on the dose profiles of anatomically
similar retrospective cases. This knowledge assists clinicians in selecting dose constraints
for the new case. Dose constraints form a part of the initial input parameters that are
provided to the treatment planning system, and heavily influence the quality of the
resulting treatment plans. My premise is that patients with similar tumor anatomies
often exhibit similar radiation dose profiles. Based on this hypothesis, dose parameters
of anatomically similar retrospective cases can be used to determine an empirically
probable range of dose parameters for a prospective patient. An interactive computer
interface will then allow clinicians to visually pick the most promising dose constraints
from the range of possible dose parameters provided by the system. This kind of
decision support could be particularly helpful for residents and new physicians, since it
compensates for their relatively lower experience. The most basic benefit provided by
8
such a system would be to increase the clinicians’ confidence in their decisions. A much
greater advantage would be an actual improvement in patient outcomes as a result of
utilizing this system.
Since the performance of knowledge-driven applications such as these are determined
by the size and heterogeneity of the database, it is essential to incorporate data from
different sources as much as possible. This includes collecting data from different
treatment planning systems, different institutions and even different countries. My
decision support system will be able to integrate data from different institutions, and
will also include a HIPAA compliant data sharing protocol that facilitates easy sharing of
Protected Health Information (PHI).
1.5 Research Aims and Scope
My research objectives are to design, prototype and evaluate a knowledge-driven
decision support system to help clinicians select dose constraints for Intensity
Modulated Radiation Therapy of head and neck cancer. In order to evaluate my
work, I have tested the hypothesis that the dose profile of a prospective patient is
likely to match the dose profiles of retrospective patients with similar tumor
anatomies more than it will match the dose profiles of retrospective patients with
dissimilar tumor anatomies. If this is true, the dose profiles of patients with similar
tumor anatomies may be used as empirical evidence to guide the selection of dose
constraints for the prospective patient. Clinicians can pick the most promising
constraints from a probable range provided by the system.
9
The scope of my work is limited to external beam Intensity Modulated Radiation
Therapy (IMRT) for head and neck cancer. My project goals are listed below.
• Design a data model and build a database for DICOM RT treatment planning data
and post-processed knowledge extracted from this raw DICOM data
• Design and implement a robust program to parse DICOM RT objects from various
treatment planning systems
• Design, prototype and evaluate a knowledge-driven decision support algorithm
for facilitating selection of dose constraints in IMRT of head and neck cancer
• Design a HIPAA compliant data sharing protocol for facilitating multi-institutional
collaboration
• Evaluate the system’s ability to integrate data from at least two different
institutions
• Demonstrate the clinical impact and significance of the decision support system
1.6 Chapter Scheme of Dissertation
Following this introduction to my research topic and goals, the next chapter provides
clinical context to the problem being studied, by explaining the relevance of decision
support for RT of head and neck cancer, the various advancements and innovations in
Radiation Therapy, the clinical workflow of treatment planning in Radiation Therapy and
the challenges that lie therein. Chapter 3 outlines the defining features of the decision
support system and describes its conceptual data model, architecture and components
from an informatics perspective. Chapter 4 describes the decision support module,
including the similarity-matching algorithm, features used for calculating anatomical
10
similarity and generation of dose constraint recommendations for the user. It also
demonstrates the clinical impact of the decision support system by illustrating possible
workflows that result from integrating the system into the clinical environment. Chapter
5 presents the results of evaluating the decision support module by testing the system’s
ability to retrieve relevant retrospective treatment plans for guiding dose constraint
selection of a new patient. Chapter 6 explains the need for multi-institutional
collaboration, outlines the design for a HIPAA-compliant data sharing protocol and
reports the results of integrating data from two different institutions – University of
California, Los Angeles and the Technical University at Munich. Chapter 7 concludes this
work by discussing its benefits and limitations, potential implications and
recommendations for future work.
11
Chapter 2 Background
2.1 Head and Neck Cancer
The clinical challenges outlined in the previous chapter are most relevant in cancer sites
which present the most complex anatomical arrangement in terms of the number and
size of critical organs as well as variability in tumor location. For instance, the location of
the tumor in prostate cancer does not vary significantly from patient to patient,
especially with respect to its two most vital Organs At Risk (OARs) – the bladder and the
rectum. In contrast, the location of tumors in head and neck cancers may vary vastly
from patient to patient. The internal geometric connections between OARs and the
tumor are complicated in head and neck cancer, and warrant the use of decision
support tools and algorithms to provide additional quantitative input to clinicians in
order to dispel some of the uncertainties involved in treatment planning. The
complexity of treatment planning in head and neck cancer influenced my decision to
focus on this disease site for my dissertation.
Head and neck cancer or Squamous Cell Carcinoma of the Head and Neck refers to
epithelial malignancies that arise in the head and neck region, specifically in the oral
cavity, nasal cavity, paranasal sinuses, pharynx and larynx [11]. Treatment may be
carried out either via surgery, chemotherapy, radiation therapy or a combination
thereof. The critical structures involved in radiation therapy of head and neck cancer are
numerous and rather small in size. Because of the complex anatomy of the head and
12
neck region, variations in the tumor location lead to considerably different treatment
plans and dose distributions across patients.
The large variations in treatment plans for head and neck cancer compound the
challenge of limiting radiation exposure of critical structures surrounding the tumor.
Excessive exposure of critical structures leads to harmful changes in the tissue, causing
side effects known as ‘radiation toxicity’ that often have a far-reaching impact on the
patient’s quality of life after treatment, irrespective of tumor eradication. For example,
exposure of the salivary glands to radiation dose levels over the tolerance limit leads to
salivary gland hypofunction or xerostomia (dry mouth). Xerostomia has an adverse
effect on the patient’s quality of life by interfering with daily functions such as speaking,
chewing, swallowing and tasting [12, 13]. Decreased saliva production also increases the
risk of oral infection and dental cavities. Salivary hypofunction often manifests itself in
other varied symptoms such as lip dryness, difficulty wearing dentures, sensitivity to
strong flavors and increased thirst. As a result, patients are susceptible to considerable
discomfort, malnutrition, limited social interactions, constraints in daily activities, and
reduced levels of mental and physical well-being. Radiation Toxicity effects for the
mandible include osteoradionecrosis, which manifests as exposure of devitalized bone
[14]. If this condition either progresses steadily or worsens, it may lead to fistulas and
infections with severe necrosis, necessitating intervention through surgery. Brainstem
injury may lead to symptoms of motor, sensory and cerebellar dysfunctions leading to
bulbar palsy, ataxia, hearing loss, etc. [15]
13
Many research studies have been conducted to understand the effects of different
radiation dose exposures on patient outcomes, in order to determine tolerance levels
for various tissue types [16]. QUANTEC, or ‘Quantitative Analyses of Normal Tissue
Effects in the Clinic’ is a research initiative that has compiled a set of normal tissue
tolerance and recommended safety guidelines based on recently conducted studies
dealing with outcomes research [17, 18]. According to QUANTEC recommendations, in
order to maintain a level of salivary function that is at least 25% of the baseline, one of
the following two conditions must be met: (1) The mean dose to both the parotid glands
should not exceed 25Gy. (2) At least one of the parotid glands should receive less than
20Gy dose. The radiation dose exposure to the brainstem should not exceed 64Gy.
Although QUANTEC does not specify tolerance for the mandible, RTOG guidelines
mandate that the exposure to the mandible should not exceed 66Gy.
Since my topic focuses on treatment of head and neck cancer through Radiation
Therapy, the following sections are devoted to providing a better understanding of
Radiation Therapy.
2.2 Radiation Therapy
Radiation Therapy (RT) involves the use of high-energy ionizing radiation to destroy
cancer cells by damaging their DNA, thus interrupting their ability to replicate. Since the
radiosensitivity of cancer cells is higher than that of most normal tissue, this DNA
damage causes differentially more harm to cancer cells. The radiation source may
sometimes be placed within the patient’s body. This type of treatment is called
brachytherapy. However, the more common form of treatment is External Beam
14
Radiation Therapy, where the radiation source located outside the patient’s body
produces beams directed at the tumor. The scope of this project does not include
brachytherapy, and from this point onwards, any references to ‘radiation therapy’ imply
‘external beam radiation therapy’. In external beam RT, radiation beams may be
sculpted to conform specifically to the tumor shape and avoid critical structures such as
the brainstem, spinal cord, parotid glands, etc. This provides further protection for
normal tissue. ‘Conformal Radiation Therapy’, which is now the accepted standard of
treatment delivery in most institutions, was made possible by advanced medical imaging
techniques such as Computed Tomography (CT).
Before the emergence of CT, radiation therapy treatment planning was based on two-
dimensional X-ray images. Beam setup arrangements were also relatively simple, often
consisting of rectangular beams emerging from one to four different directions, creating
a box-shaped radiation field at their intersection. This inevitably led to the exposure of
normal tissue situated in the four-field treatment box. Three-dimensional Conformal
Radiation Therapy (3D-CRT) provided considerably more freedom for beam setup due to
greater knowledge regarding the position and shape of tumors. This included the ability
to vary beam intensities using physical devices such as filters and wedges, allowing for
greater conformity of the treatment field to irregularly shaped tumors. The next major
break-through development in RT was ‘Intensity Modulated Radiation Therapy’ (IMRT).
15

2.3 Intensity Modulated Radiation Therapy (IMRT)
Intensity Modulated Radiation Therapy (IMRT) is a type of treatment planning and
delivery methodology that significantly improves conformity of dose to the target
volume as well as sparing of organs at risk [19]. Since this reduces the volume of normal
tissue that is irradiated, dose escalation becomes more viable,
thus improving the therapeutic ratio. This is made possible by
the ability to shape radiation beams to conform to the tumor,
and to vary or modulate the intensity of each beam throughout a
treatment session. The shaping and intensity modulation of
beams is achieved through the use of multi-leaf collimators
(MLCs), shown in the top portion of figure 2-1. MLCs are
computer-controlled devices that consist of many individual fingers or leaves placed at
the aperture of the radiation beam. MLCs can be moved back and forth in order to block
certain parts of the beam, while letting other parts emerge. Varying the MLC
configuration produces different beam shapes, as shown in the bottom portion of figure
2-1. The configuration of the MLCs is often changed in real time during each treatment
session.
Another unique feature associated with IMRT is Inverse Treatment Planning (ITP). In
conventional forward 3DCRT treatment planning, the planner specifies beam direction,
shapes and static beam intensities to the TPS, which then merely calculates the resulting
dose distribution. In inverse treatment planning however, the planner specifies the
Figure 2-1 Multi-Leaf
Collimators
16
desired dose distribution, and the TPS calculates the beam intensities and shapes that
best meet the specified dose distribution. ITP is discussed further in the next section.
There are two primary scenarios that warrant the use of IMRT despite its considerably
higher costs and planning complexity [19]. The first is where conventional RT is currently
not very successful due to target dose restrictions, in order to spare surrounding OARs.
IMRT allows the target dose to be increased significantly while protecting OARs at the
same time, resulting in much better tumor control. The second instance is where
conventional techniques already provide good tumor control, but where reduction in
OAR dose may significantly improve a patient’s quality of life after treatment. Examples
of this include pediatric cancers, whole abdomen radiation (WAR), breast cancer, etc.
Recent advancements in IMRT include Volumetric Arc Radiation Therapy (VMAT) and
Tomotherapy. VMAT aims to reduce treatment time by using single or multiple radiation
arcs in a single 360 degree gantry rotation. Tomotherapy uses a slice-by-slice treatment
delivery mechanism that greatly improves dose conformity but may result in increased
treatment times.
IMRT is often a good choice for treating head and neck cancer due to the complex
arrangements of the various organs at risk around the tumor. Studies have shown the
benefits of IMRT in reducing Organ At Risk (OAR) dose for head and neck cancer,
especially the parotid glands, as well as improving outcomes [20].
2.4 Inverse Treatment Planning (ITP) Workflow in IMRT
Treatment planning in radiation therapy is carried out on workstations known as
Treatment Planning Systems (TPSs). Some widely used popular treatment planning
17
systems are Varian’s ‘Eclipse’ and Philips’ ‘Pinnacle’. The six major steps involved in
Radiation Therapy treatment planning are summarized in figure 2-2.
1. Image Acquisition
Radiation Therapy treatment plans are
developed based on a three-dimensional
model of every patient’s individual
anatomy, which is obtained through
medical images. The most common
modality used for treatment planning is
Computed Tomography (CT). At times,
Magnetic Resonance Imaging (MRI) and
Positron Emission Tomography (PET) may
also be used. The scope of this thesis is
limited to treatment planning based on CT.
2. Delineation of the target and critical organs
After acquisition of the planning image set, the radiation oncologist and dosimetrist
draw contours of specific Regions of Interest (ROIs) on each CT slice. This may be done
using special software on treatment planning workstations, or in some cases may even
be automated to some extent by segmentation algorithms. Computer algorithms use
these contours directly to identify objects of interest for dose optimization, generation
of evaluation metrics for specific structures, etc.

Figure 2-2 Inverse Treatment Planning (ITP)
workflow in Radiation Therapy
18
There are two main types of ROIs:
• Treatment Volumes
The cancer tumor makes up most of the treatment volume, and is the region that is
targeted for maximum irradiation during treatment planning. There are three types
of treatment volumes [21]:
a. The Gross Tumor Volume (GTV) is the visible tumor as
seen on CT or other images, and is depicted in red in
figure 2-3.
b. The Clinical Tumor Volume (CTV) includes the GTV as well
as other regions of risk such as microscopic spread of the
cancer, nodal chains, etc. and is highlighted in green in figure 2-3.
c. The Planning Target Volume (PTV) is obtained by expanding the CTV to
account for setup errors, patient motion, alignment errors, etc. For all intents
and purposes, this is the true target for radiation therapy. It is shown both in
figure 2-3 (blue) and figure 2-4 (red).
• Organs At Risk (OARs) or critical structures
These are structures situated in close proximity to the PTV, and
are likely to be affected by the treatment dose. My study focuses
on four OARs: the two parotid glands, the brainstem and
mandible. Some of these structures are highlighted in figure 2-4.
Other OARs for head and neck cancer include the spinal cord, the
eyes, the cochlea, larynx pharynx, etc.
Figure 2-3 Target
Volume definitions:
GTV (red), CTV (green)
and PTV (blue)
Figure 2-4 Example of
ROI Contours:
Mandible (pink),
Parotid Glands
(green), Tongue
(yellow), PTV (red)
19
3. Selection of Treatment Parameters and Dose Constraints
Development of the treatment plan begins with the dosimetrist and radiation oncologist
specifying initial parameters such as the number of treatment beams, the direction of
every beam, the maximum field size and beam energies. In conventional treatment
forward planning, clinicians would then vary the intensity profiles of each beam
manually, using filters, compensators and manual adjustments to multi-leaf collimators.
This would form the input to the treatment planning algorithm, which would simply
calculate the resulting dose distribution.
In inverse treatment planning, however, clinicians do not concern themselves with the
beam intensities at all. The treatment planning system incorporates an optimization
algorithm that calculates the best beam intensity profiles based on user defined dose
limits for certain ROIs. The dose constraints specified by the clinicians are described
below.
• Dose limits for the PTV: Clinicians may define the minimum dose that must be
delivered to the PTV.
e.g. Minimum dose to the PTV = 65Gy
• Dose limits for critical structures: Clinicians may specify the maximum dose limits for
organs at risk surrounding the PTV.
e.g. Maximum dose to the brainstem = 54y
• Dose Volume Effects: Instead of setting constraints for the entire structure, clinicians
may set constraints for part of a structure.
20
e.g. PTV D
80
> 55Gy, i.e. 80% of the volume of the PTV must receive at
least 55Gy
• Weights and penalties: Clinicians may prioritize fulfillment of dose constraints for a
particular structure over another. For instance, they may decide that limiting dose to
the brainstem is more important than maximum dose exposure for a PTV. Every
structure may have a set of dose constraints associated with it, where each
constraint has a different priority.
The optimization algorithm performs multiple iterations until it reaches a tolerable
difference between the dose parameters of the calculated distribution and the specified
dose constraints. Selection of good dose constraints is a very important step in
producing an acceptable treatment plan in IMRT. Overly lax constraints may lead to
inferior plans, and overly stringent constraints may lead to unexpected negative
consequences. The proposed decision support system facilitates selection of good dose
constraints by providing quantified, evidence-based inputs based on retrospective
treatment planning data.
4. Dose Optimization
The optimization algorithm of the TPS then computes the best beam intensity profiles
that result in accomplishing the specified dose constraints. The TPS will also generate a
three-dimensional dose distribution that corresponds with the calculated beam
intensities.

21
5. Plan Evaluation
This step is very important in determining whether the dose distribution produced by
the optimization algorithm is acceptable or not. Clinicians must ensure that their original
dose constraints were implemented within acceptable tolerance limits, without any
unintended negative consequences. Commonly used evaluation tools include the
following:
• 2-dimensional isodose contours or dose color washes: Dose
distributions are visually rendered either as contours or
color-maps overlaid on every CT slice. Different colors
represent varying magnitudes of dose. An example of a 2-
dimensional dose color wash is shown in figure 2-5.
• 3-dimensional color wash: The dose distribution is rendered as a three-dimensional
color wash on a three-dimensional model of the PTVs and critical structures.
• Dose Volume Histograms: The Dose
Volume Histogram encodes a
comprehensive depiction of the three-
dimensional dose distribution of an ROI
in a simple two-dimensional curve. It
summarizes the amount of dose
received by various volumetric fractions
of a particular ROI. The DVH is the most widely used tool in the evaluation of IMRT
using ITP, and is frequently utilized in daily clinical practice. Figure 2-5 provides some
Figure 2-5 Examples of Dose Volume Histograms
Figure 2-5 2-Dimensional
dose color wash
22
examples of DVHs. Point A in figure 2-5 shows that 100% of the PTV (i.e., the entire
PTV) receives at least 68Gy of radiation dose. Area under the curve should be
maximized for target volumes in order to maximize dose exposure, and thus tumor
eradication. Point B shows that 50% (half) of the volume of the right parotid receives
at least 22Gy dose. This is in compliance with guidelines, which recommend a safe
dose limit of 25Gy for the parotid glands. Clinicians can thus utilize DVHs to
determine over-exposure or under-exposure of various ROIs. For Organs At Risk,
minimum area under the curve in desired. Point C shows that no part of (0%) the left
parotid receives 62Gy, i.e. the maximum left parotid dose is 62Gy. Since this is
where the curve ends, it indicates that the maximum dose delivered to the left
parotid is 62Gy.
6. Re-adjustments
If a treatment plan is deemed unacceptable after evaluation, changes must be made
either to the IMRT dose constraints, ROI contours or other initial treatment parameters
such as number and direction of beams. The TPS then re-initiates plan optimization
based on these new input parameters. If the resulting plan is found acceptable,
treatment planning is concluded and the patient is scheduled for treatment delivery.
2.5 Challenges in IMRT Treatment Planning and the need for decision support
While selecting dose constraints for critical structures, clinicians have no concrete way
of knowing the least possible dose limit that would allow for a feasible treatment plan.
However, experimenting with rigidly low dose constraints could lead to dose
distributions with undesirable hot or cold spots, and other inconsistencies. It would
23
require a considerable number of iterations to try various combinations of low OAR
dose constraints, which would be extremely impractical. In everyday practice, clinicians
often rely on their own judgment and experience, as well as institutional guidelines to
prescribe good dose constraints.
The possibility of sparing critical structures is usually limited by every patient’s unique
tumor characteristics such as size, shape and location. This means that the estimation of
good dose constraints is ultimately a very subjective judgment call, since the
manifestation of head and neck tumors varies considerably from patient to patient.
There are no “right” or “wrong” dose constraints, but there are often subtle margins of
error and scope for possible improvement, given the right information at the right time
in the planning workflow. For instance, if it were possible to retrieve the dose
distributions of previous cases that resembled the current case at hand, one could
extrapolate realistically feasible dose constraints for the current case from the DVHs of
those previous cases. Showing clinicians an accurate picture of what was feasibly
accomplished in the past for similar cases could help them identify if the dose
constraints for their current case are too optimistic or simply inadequate. This would
not only increase their degree of confidence in the decision-making process, but could
also help lower the dose delivered to critical structures in some cases, thus potentially
improving patient outcomes for those cases. This kind of timely influx of relevant
information that could positively influence the decision making process is termed
‘decision support’. With respect to the IMRT inverse treatment planning workflow, a
24
decision support system could benefit the plan evaluation loop in certain cases, and at
the very least, increase clinician confidence in others.
This chapter provided clinical context to the work described in this manuscript,
beginning with an introduction to Intensity Modulated Radiation Therapy. This was
followed by a detailed discussion of workflow in radiation therapy treatment planning
and and an explanation of key concepts such as Region Of Interest classification, Dose
Volume Histograms, IMRT dose constraints and treatment plan optimization.
Bottlenecks in the workflow were highlighted and a solution was offered in the form of
a knowledge driven decision support system. The next chapter describes the features
and components of this system.
25
Chapter 3 System Design
I have developed a computer based, knowledge-driven decision support system for
IMRT treatment planning of head and neck cancer. This system performs three major
functions:
1. Builds and maintains a database of diverse retrospective treatment planning
data for head and neck cancer patients
2. Provides tools for knowledge discovery of cases with tumor characteristics
similar to those of a prospective “new” patient
3. Provides decision support for selection of dose constraints for the new patient
based on the dose volume histograms of similar database patients.
This system incorporates features listed below, in order to accept real-world clinical
data, provide modularity of design and easily integrate back into the clinical
environment. These features address the challenges of system integration and
incorporation into the clinical environment for use by clinicians in daily practice.
1. Support for multi-institutional collaboration: The success of a knowledge driven
decision support system depends on the availability of a large and
heterogeneous knowledge base. Since most institutions often follow standard
treatment protocols, including data from different institutions can increase
diversity. This would involve building a system robust and flexible enough to
accept complex treatment planning data from various sources, and devising a
26
strategy to incorporate HIPAA compliance since treatment planning data is
considered Protected Health Information (PHI).
2. Vendor Neutrality: Although most treatment planning system (TPS) vendors
claim conformance with the DICOM standard, most systems often store their
data objects in proprietary formats for practical purposes. To develop a decision
support system that aggregates data from various treatment planning systems
and institutions, I have chosen to work with the DICOM standard. Most
treatment planning systems feature the ability to export their data objects to
DICOM. If the TPS does not have DICOM export features, our system is able to
convert the treatment data objects into the DICOM format.
3. Robustness: The database design and DICOM parser must be robust enough to
successfully integrate data from multiple treatment planning systems and
different institutions.
4. Modularity: The decision support system must include the ability to integrate
data from larger databases such as cancer registries, or decision support
modules or tools from other systems. Conversely, it must also be designed such
that its pieces may easily be integrated into a larger information system such as
a hospital electronic medical record (EMR) system.
3.1 The DICOM Standard
In order to facilitate integration of treatment planning data from multiple clinical
sources into a single knowledge base, it is essential to adopt a common data standard
that is known and accepted across all manufacturers. The ‘Digital Imaging and
27
Communications in Medicine’ (DICOM) format [1], which is the de-facto standard for
clinical storage and transmission of medical data, is now a recognized industry standard
for all medical equipment manufacturers. Although treatment planning systems do not
directly employ DICOM for native storage, most vendors include a DICOM export
functionality that packages TPS data objects into a standardized DICOM format.
When the DICOM standard was first published in 1993, it was mainly geared towards
standardizing storage and communication of medical images in radiology departments.
However, it has been revised and expanded since then to include several non-traditional
medical specialties such as radiation therapy, cardiology, surgery, ophthalmology and
pathology, to name a few. Radiation Therapy was the first specialty to be included in the
DICOM standard due its heavy reliance on medical imaging. Four DICOM Radiation
Therapy (DICOM-RT) objects were introduced in 1997, followed by three more in 1999
[22].
DICOM utilizes object-oriented data structures to store medical data in the form of
attribute-value pairs. This information includes image pixel values, patient demographic
information, medical equipment details, annotations, contours, etc. Figure 3-1 shows an
example of metadata from a DICOM RT Treatment Plan object. The ‘Field Name’ column
describes attributes or various characteristics of DICOM-RT plan objects, while the
‘Content’ column provides the values of those attributes for this particular instance of
the DICOM-RT plan object. For instance, it can be observed that this plan belongs to a
patient whose patient ID is ‘ANON1’.
28

Figure 3-1 An example of meta-data from a DICOM-RT Plan object
The four major DICOM objects relevant to this decision support system are listed below.
1. CT Images allow clinicians to visualize the patient’s anatomy and tumor location.
2. The DICOM RT Structure Set contains labels and contours identifying Regions Of
Interest (ROIs) such as the Planning Target Volume (PTV) or the tumor, and critical
structures or Organs At Risk (OARs) surrounding the PTV.
3. The DICOM RT Dose object provides a detailed radiation dose distribution matrix for
the treatment area. Most treatment planning systems allow Dose Volume
Histograms (DVHs) to be exported as a part of the DICOM RT dose object.
4. The DICOM RT Plan contains technical treatment plan details including patient
positioning information, prescription details that include IMRT dose constraints, a
fractionation scheme that describes the number of fractions and prescription of
each fraction, tolerance tables that specify allowable error margins of treatment
29
parameters and detailed beam parameters including collimator settings, beam
modification device settings, gantry angles and isocenter coordinates.
3.2 Data Model
Since the goal of this decision support system is to perform data mining on a knowledge
base, it is essential to provide efficient and timely access to required data elements.
A conceptual data model provides a high-level abstract representation of the system’s
data objects by identifying all the major information entities, their attributes and the
relationships between them. It lays the groundwork for preparing the data schema,
which is a detailed blueprint of the database containing table names, field names, data
types, primary keys, etc. The development of a robust data model and schema ensures
Figure 3-2 A conceptual data model of the decision support system
30
ease of query-retrieve operations. Figure 3-2 shows a conceptual data model of the
system.
Each patient in the system may be associated with multiple RT treatments. Each
treatment forms a study, and may contain multiple CT and DICOM RT series. The CT
pixel data, ROI coordinates and the dose matrix may all be registered to the same
patient coordinate system by using their respective patient position and orientation
attributes. This provides a three-dimensional model of the system that may be used to
derive the Overlap Volume Histogram (OVH) and Spatial Target Signature (STS), which
form the decision support system’s similarity features, and will be described in more
detail in Chapter 4. All attributes are obtained directly from DICOM objects except for
the Overlap Volume Histogram and the Spatial Target Signature, which are derived
computationally.
3.3 System Architecture
The decision support system is comprised of a number of software services and tailored
scripts, running on a Linux rack server that is housed in a secure data center at the
University of Southern California. This section describes the five major system
components, as shown in figure 3-3.
1. DATA GATEWAY: All clinical data objects entering the decision support system must
pass through the data gateway module to ensure proper processing and redirection.
Since this data is mainly in DICOM format, the DICOM parser forms the biggest portion
of the data gateway module. The parser’s main job is to read incoming DICOM files and
extract desired meta-data. The rest of the data gateway functions include assignment of
31
physical storage locations for the uploaded DICOM files, re-organization of directory
structures and insertion of extracted metadata into the database. There are two
versions of the DICOM parser.
Server-Side DICOM Parser: The server-side parser runs on the decision support system
server, and accepts files that have been uploaded to the server over the web. In order to
ensure HIPAA compliance, the user must anonymize all DICOM files before uploading.
This server-side parser is written and implemented in Python, and has a modular design
that enables its integration into future RT-based applications and other information
systems. The uploaded files may either be stored or deleted after metadata extraction,
based on user preference. The user may prefer to use the client-based version of the
parser for data upload to avoid the need for manual anonymization.
Client-side DICOM Parser: This novel DICOM parser is written in JavaScript, and allows
extraction of DICOM metadata on the client’s machine, without any need for the files to
leave the institution’s network. It forms the backbone of the HIPAA compliant data
sharing protocol, and is described further in Chapter 6.
Figure 3-3 Architecture of the decision support system
32

2. DATABASE: The raw metadata that is extracted by the parser from the DICOM files is
stored in a set of 53 database tables, which correspond to DICOM modules belonging to
the various DICOM data objects. Figure 3-4 shows an example of all the modules
contained in the DICOM RT Structure Set object. Each module is assigned a separate
table in the database schema.
Each module is assigned a separate database table, and nested sequences within each
module are further assigned their own database tables. The database also stores the
similarity features extracted from the raw metadata by the image processing engine,
user accounts and decision support results that the user may want to save. The
database schema follows the DICOM model of the real world, and is organized to
facilitate quick retrieval of desired data on querying. The system uses MySQL, which is a
robust and open-source database management system.

Figure 3-4 List of modules contained within the DICOM-RT Structure Set object
33
3. IMAGE PROCESSING ENGINE: The image processing engine is responsible for
extracting similarity features from the CT images and ROI contours, using computational
techniques described in chapter 4. These functions were prototyped and tested in
MATLAB, with a direct MATLAB-MySQL connection facilitating efficient data
management. A Python implementation should ultimately replace all MATLAB functions.
4. ANALYTICS ENGINE: The analytics engine is responsible for identification of
anatomically similar database patients, based on the features extracted by the image
processing engine. The analytics functions were also developed and tested using
MATLAB, and will be ported to Python once ready for deployment. Chapter 4 describes
this module in greater detail.
5. GRAPHICAL USER INTERFACE (GUI): The GUI serves two main purposes. It facilitates
decision support by providing tools to select IMRT dose constraints and to help in
evaluating existing dose distributions. It also lets users upload retrospective treatment
planning data in order to expand the knowledge base. The GUI incorporates a DICOM
Viewer that enables users to examine CT image series, with ROI contours and isodose
contours overlaid on top of each corresponding image. Users can scroll through images,
select specific ROIs and isodose curves for display and view Dose Volume Histograms.
One component of the GUI is the interactive plan assessment tool, which helps clinicians
to pick IMRT dose constraints and evaluate dose volume histograms. It is discussed in
more detail in chapter 4.
34

3.4 Workflow of the decision support system

The decision support system may operate in one of two modes, as described here.
1. Data Collection
Users contribute retrospective data to the system on a regular basis, in order to
expand and update the knowledge base. The DICOM gateway provides an interface
for uploading DICOM files, using either the server-side or the client-side DICOM
parser.
(i) Server Side Processing: Once the server has received the anonymized DICOM
files, the DICOM gateway activates the Python-based DICOM parser to
extract metadata. The files are then stored in their designated directories,
and pointers to those files are recorded in the database. However, users
must anonymize data on their end before uploading to the server. This
requires the use of either institutional or third-party software applications.
The anonymization step may be avoided by utilizing the client-side parser.
(ii) Client Side Processing: The gateway activates the JavaScript client-side
DICOM parser in the user’s browser. The extracted metadata is transferred
back to the server, where the gateway receives and catalogs it in the
database.
Next, the DICOM gateway activates the Image Processing engine to extract similarity
features using raw metadata from the database, and to save the resulting features
back into the database.
35
2. Clinical Decision Support
When users access the decision support tools from the system for a new patient,
they must first upload the patient’s CT images and ROI contours. The first two steps
involve extraction of raw metadata and computation of similarity features, as
described in the previous section. Next, the system triggers the analytics engine to
search the database for anatomically similar retrospective cases, based on the
system’s similarity features. Once the most similar retrospective cases have been
identified, the analytics engine utilizes their Dose Volume Histograms to compute
95% Confidence Intervals for four critical structures – the parotid glands, mandible
and brainstem. The Dose Volume Histograms and the 95% Confidence Interval
estimates are supplied to the Graphical User Interface, which incorporates them into
the plan assessment tool. Users may then employ the plan assessment tool to select
IMRT dose constraints or to evaluate treatment plans.
This chapter discussed the major functions and features of the decision support system,
its architecture, data model and flow of operations. The decision support module was
briefly introduced, and will be described in greater detail in the next chapter.
36
Chapter 4 Decision Support Module

The goal of the decision support module is to utilize knowledge gained from
retrospective treatment planning data to identify good radiation dose end-points. This is
accomplished by assisting clinicians in picking IMRT dose constraints and evaluating
existing dose distributions. The key to doing this is providing adequate empirical
evidence to construct a reference scale for assessing the feasibility of a range of possible
dose parameters that the clinician may choose from. This empirical evidence is in the
form of relevant retrospective dose distributions from previous treatment plans
performed in recent years for similar patients. The relevance or applicability of
retrospective dose distributions to the current case is defined by patient-specific factors
that influence radiation dose distributions. Studies have shown that one of the primary
factors that affects the resulting dose distribution is the spatial relationship of the
Planning Target Volume (PTV) to the surrounding Organs At Risk (OARs) [23, 24]. This
includes properties such as volume, distance, overlap, location, orientation and shape.
Since it is difficult to conform radiation beams precisely to the target during treatment,
structures closest to the PTV are at the greatest risk of dose exposure, especially those
that overlap with the PTV. Placement of the PTV with respect to surrounding organs at
risk controls the angle of entry of radiation beams, which in turn influences the dose
distribution. Finally, concave PTV shapes make it difficult to spare any healthy tissue
that may be enclosed in the target concavity. The second factor that dictates the
resulting dose distribution is the intended target dose. The decision support module
37
incorporates these characteristics into three comprehensive features. Database cases
that exhibit features similar to a prospective case are called ‘anatomically similar
matches’. Since the decision support features have a bearing on the resulting dose
distribution, the anatomically similar database matches are likely to have dose
distributions that are relevant to the current case.
To summarize, the decision support module is responsible for identifying retrospective
database cases that have tumor characteristics similar to the prospective patient. This is
done in order to utilize the dose distributions of the selected database cases as
reference templates or guidelines to assess the merit of various dose parameters for the
prospective patient. Some important naming conventions that have been used in this
chapter are listed below.
Query Case, or ‘Q’ The new, prospective patient for whom decision
support is sought
Database Case, or ‘DB’ Retrospective database case
Q-DB Pair The query case and database case that are currently
being compared
4.1 Previous Work in Knowledge Driven Decision Support
Automation of treatment plan development and evaluation is currently a popular topic
of research in Radiation Therapy. Different methods of comparing and quantifying
relevant patient anatomy have been developed to assist in quality control. Moore et al
initially developed a simple mathematical model to study the effects of PTV-OAR
overlap on OAR mean dose [23]. Although this was limited to a single feature, PTV-OAR
overlap, it was successfully utilized to perform quality control by flagging treatment
plans where the planned OAR dose exceeded the dose predicted by the model.
38
Appenzoller et al. continued this work to develop a more comprehensive mathematical
model that predicts dose volume histograms (DVHs) by calculating the dose delivered to
each voxel based on its distance to the closest PTV surface [25]. This model however,
cannot effectively be used for Simultaneous Integrated Boost (SIB) plans, where multiple
targets are prescribed differing doses. Schreibmann and Fox developed a method to
identify tumor similarity for prostate cancer by co-registering the query-database pair
and then calculating the mean distance from the query patient’s prostate volume to the
nearest prostate surface of the database patient [26]. However, this method is difficult
to apply to cancer sites that involve greater PTV variations. Yuan et al performed a study
that conclusively identified factors that affect OAR dose sparing [24]. A majority of these
factors are related to the distance of volumetric OAR fractions from the PTV, and have
been included in this work. Following these publications, a new feature was developed
that condensed the various features mentioned in some of the previous studies into a
more comprehensive feature – the Overlap Volume Histogram (OVH) [27]. Simari et al
and Wu et al have published several studies that utilize single-valued points and
statistics from the OVH to identify similar retrospective cases for decision support [28,
29, 30].
The decision support module in this manuscript has some important distinctions from
the previous studies. Although it has been developed for head and neck cancer, it is
flexible enough to allow expansion to other disease sites. The OVH is utilized in
conjunction with a completely new feature – the Spatial Target Signature (STS). The
decision support module supports treatment plans with multiple targets, as in
39
Simultaneous Integrated Boost (SIB) cases, and provides holistic rather than point-by-
point comparison of feature histograms. Moreover, it places an emphasis on building a
diverse and continuous expanding knowledge base through multi-institutional data
sharing, made possible with data standardization, vendor neutrality and HIPAA-
compliant data sharing protocols. Collaboration is also prioritized by allowing plug-and-
play integration with similar knowledge repositories or decision support applications.
4.2 Similarity Features

As mentioned previously, the features utilized for determining the relevance of
retrospective database cases as knowledge for the prospective query case, must
have a bearing on the final OAR dose profile. This is because the selected database
cases will serve as reference templates for determining a good dose distribution for
the query case. This decision support module uses three main features to evaluate
anatomical similarity – the Overlap Volume Histogram (OVH), the Spatial Target
Signature (STS) and the Target Dose. It should be noted that two of these features are
histograms that capture a combination of detailed spatial information. They provide
greater specificity than using multiple single-valued features, which are analogous to
picking limited points or measures of central tendency and variability from the
histogram distributions.
4.2.1 Overlap Volume Histogram (OVH)

The Overlap Volume Histogram (OVH), developed by Kazhdan, et al. [23] is a feature
that describes the volumetric spread of an Organ At Risk (OAR) from the closest surface
40
of the Planning Target Volume (PTV), described in terms of distance and overlap. The
dose that is delivered to a point in space depends on its proximity to the PTV surface. It
is thus useful to know the distance of various parts of an OAR to the closest PTV surface.
Sometimes, parts of an OAR may even intersect with the PTV. Such overlapping regions
inevitably receive high dose exposure. All of this information is included in the OVH of a
particular OAR. The OVH is always calculated with respect to two structures – the PTV
and an OAR.
Consider an overlap volume histogram defined with respect to an organ ‘O’ and a tumor
‘T’. The value of the OVH at a distance t is defined as the volume of that subset of O,
which is at a distance of t or less from the tumor T.
!"#
$,&
' =
*+! | -(*,/)≤'
!
[4-1]

* : Voxels within the organ
! : The volume of organ O
-(2,/) : The distance of point 2 from the boundary of the tumor T
Figure 4-1 demonstrates a simplified version of the procedure involved in generating an
overlap volume histogram for the right parotid and the PTV, from a two-dimensional CT
slice. It is important to note that although the figure shows a 2-dimensional example,
the actual OVH feature used in the system is inherently three-dimensional in nature and
represents the 3D spatial relationship between the PTV and surrounds OARs. The steps
shown in figure 4-1 are as follows:
(1) Extract contours for the right parotid (green) and the PTV (purple).
41
(2) For each voxel, calculate the distance between the voxel and the closest surface of
the PTV. This is done by calculating the distance of the voxel to every point on the PTV
surface, and then selecting the minimum distance.
(3) Label each voxel with its distance score.
(4) Split the range of voxel-PTV distances into equally spaced intervals to generate
histogram bins.
(5) Count the number of voxels that belong to each bin, and generate a histogram.
(6) Normalize the voxel count to a percentage, and convert the histogram to a
cumulative form.
Figure 4-1 Steps involved in constructing an OVH for a two-dimensional CT slice
42
Figure 4-2 provides an example of three OVH curves for different OARs, belonging to the
same patient. Consider the green curve, which shows the relationship between the right
parotid gland and the PTV. The OVH plots volume fractions (in %) of the right parotid
versus the distance of those volume fractions to the PTV. The OVH implicitly integrates a
number of simpler features such as volumetric overlap and mean distance between the
PTV and the OAR.
For instance, point A in figure 4-2 represents the volumetric overlap feature, and shows
that 4% of the voxels in the right parotid are at a distance of 0mm or less from the PTV.
That is, 4% of the right parotid overlaps with the PTV. Similarly, point B tells us that 50%
of the voxels of the right parotid lie at a distance of 20mm or less from the PTV, i.e., half
of the right parotid lies at a distance of at most 20mm from the PTV.
Figure 4-2 Examples of Overlap Volume Histograms
43

4.2.2 Spatial Target Signature (STS)

The Overlap Volume Histogram alone is not enough to describe all the factors that
impact the treatment parameters and dose distribution. It represents the relationship of
an OAR to the most proximal PTV surface, but provides no information about the
distance, location, shape or orientation of the more distal parts of the PTV. These
factors often impact the number, direction and shape of treatment beams, which in turn
have a direct impact on the dose
distribution. It is thus essential to
account for such factors when
utilizing anatomical similarity to
estimate the potential similarity of
dose distributions. The second
feature of the decision support
module that fulfills this purpose is
the Spatial Target Signature (STS). It
quantifies the volumetric
distribution of the Planning Target Volume in three-dimensional space, with respect to
an Organ At Risk as the reference point. This is in contrast to the Overlap Volume
Histogram, which is defined for the OAR with respect to the PTV as a reference point.
Figure 4-3 The three spherical coordinates used in the
Spatial Target Signature (STS)
44
The STS is defined as a function of each PTV voxel’s spherical coordinates with respect
to the OAR centroid as the origin. The three spherical coordinates are illustrated in
figure 4-3. The radial distance is the three-dimensional Euclidean distance of the voxel
from the OAR centroid O
C
, and ranges from 0 to the distance of the furthest voxel. The
decision support module uses increments of 5mm to define radial distance histogram
bin intervals I
r
, which are thus defined by the series {(0,5) (5, 10) (10,15) ...}. This
increment may be reduced for greater resolution in the future, but provides a practical
balance between speed and accuracy at the moment. The azimuth and elevation are
angles defined for the vector !
3
* that starts at the OAR centroid ‘O
C
’ and ends at the
PTV voxel ‘v’. The azimuth is the angle formed by the orthogonal projection of !
3
* on
the x-y plane, with the positive x-axis, and is always measured in the counter clockwise
direction. The elevation is the angle formed by !
3
* with the x-y plane, and provides the
inclination angle of the voxel, above or below the x-y plane. Both the azimuth and the
elevation range from -180
o
to +180
o
. The histogram bin intervals for azimuth and
elevation (I
az
and I
el
) use increments of 45
o
, and are as follows: {(-180
o
, -135
o
) (-135
o
, -
90
o
) (-90
o
, -45
o
) (-45
o
, 0
o
) (0
o
, 45
o
) (45
o
, 90
o
) (90
o
, 135
o
) (135
o
, 180
o
)}. The STS
function given by equation 4-2 counts the number of PTV voxels that are associated
with each combination of histogram bin intervals. The voxel counts are then normalized
by the total number of PTV voxels to yield volumetric fractions corresponding to various
localized regions in space defined by the histogram bin intervals.

45
4/4 5
6
,5
78
,5
9:
=
*:<(*,!
=
)∈5
6
,?@(*,!
=
)∈5
78
,AB(*,!
=
)∈5
9:
C/"
[4-2]

DℎA*:*JNAB OABJILFIL 'J 'ℎA C/" !
3
:KAI'C/" = *JBPQA JH 'ℎA C/" <:3

?@:?@FQP'ℎ JH *AK'J< !
3
* AB:ABA*?'FJI JH *AK'J< !
3
*
5
6
,5
78
,5
9:
= #FG'JL*:< *,!
=
∈5
6
,?@ *,!
=
∈5
78
,AB *,!
=
∈5
9:
: "JBPQA JH ?BB 'ℎA *JNABG 'ℎ?'
JKKP< FI FI'A<*?BG 5
6
,5
78
?I- 5
9:

The STS effectively partitions the three-dimensional space around the OAR into a virtual
three-dimensional grid where each grid cell is uniquely identified by a triplet of intervals
– (I
r
, I
az
, I
el
). Setting the OAR centroid as the origin ensures standardization of the
spherical coordinate space across patients, and eliminates the need for costly image
registration procedures. The STS is thus a four-dimensional histogram that is
constructed for a PTV with respect to a single OAR. As a result, multiple Spatial Target
Signatures may be defined for one patient, each with respect to a different OAR.
46
Figure 4-4 demonstrates the need to incorporate the STS in addition to the OVH.
The left parotid Overlap Volume Histograms of cases #1 and #2 show a similarity of
99.89%. However, it is apparent from the 3-dimensional model shown in the figure that
the two cases exhibit marked differences in the shape, size, location and volumetric
spread of their respective PTVs (yellow). This is because the OVH quantifies the distance
of the OAR (left parotid in this case) to the closest PTV surface, without considering the
location of the PTV. A bold black circle is used to highlight the PTV surface voxels closest
to the left parotid. A double headed black arrow shows the distance between the left
parotid and the closest PTV surface. It can be observed that this distance is the same in
both cases. However, there are differences in other PTV characteristics such as location
and shape, which are not quantified by the OVH. This fact is corroborated by the DVH
dissimilarity shown in the bottom half of the figure, where the left parotid in case #1
(orange) receives a lot less dose than the left parotid in case #2 (green). So although the
Figure 4-4 Significance of the Spatial Target Signature (STS) in retrieving cases with similar dose
distributions
47
OVH similarity is high, the dose profiles are not similar. The low STS similarity of 69% can
thus account for the differences in the DVHs, which could not be explained by the OVH.
4.2.3 Target Dose

The dose delivered to critical structures depends greatly on the prescription dose
for the Planning Target Volume. While it is always possible to lower the target dose
enough to significantly spare critical structures, this is rarely implemented in
clinical practice since it is counter-productive to the primary goal of radiation
therapy. However, OAR sparing may sometimes be prioritized over tumor coverage
to a considerable extent, due to a number of reasons. The most common reasons
are:
(i) The treatment is palliative rather than curative, attempting to relieve symptoms
and maintain the patient’s quality of life rather than to eradicate the entire tumor.
This involves a greater focus on OAR sparing compared to target exposure, thus
influencing the dose distribution.
(ii) Sequential boost treatments are forthcoming
(ii) The potential damage to the OARs outweighs the risk of cancer retention
When clinicians are presented with anatomically similar retrospective cases, it is
essential to select database cases that exhibit the target dose coverage expected by
the clinician for the prospective case. For instance, if a database case is very similar
in anatomy to the new patient and has a very low OAR dose, it could be a very useful
asset in developing a good treatment plan for the new patient. However, if the low
OAR dose is due to an irregularly low target dose, this result is misleading since the
clinician needs reference database cases that have low OAR doses as a result of good
48
treatment planning despite high target dose coverage. In summary, the mean PTV
dose of an anatomically similar database case must match the intended target
prescription dose of the query case to qualify as a useful reference template.
4.3 Similarity Matching
The similarity between a query case and a database case is determined by first
calculating the dissimilarity between individual features. The target dose dissimilarity
(D
TD
) for every query-database pair is calculated as a simple difference between the
prescribed target dose for the query case and the mean delivered PTV dose of the query
case, as shown in equation 4-3.
S
&T
= C/" CU
−C/" WA?I SJGA
TX
[4-3]
Since the next two features are histograms, dissimilarity is represented by the distance
between the histogram distribution of the query case and that of the database case.
Some histogram dissimilarity metrics that may be used include bin-by-bin metrics such
as the Minkowski-form distance and histogram intersection, as well as cross-bin metrics
such as the Quadratic form distance and the Kolmogorov-Smirnov distance. This
decision support module however, employs the Earth Mover’s Distance (EMD) [31] to
measure histogram distances. The EMD measures the minimum cost that must be paid
to transform one distribution into another, and is based on a solution to the
transportation problem from linear algebra. It has been shown to perform better than
other histogram dissimilarity metrics for applications related to content-based image
retrieval and computer vision. Equations 4-4 and 4-5 are used to calculate OVH and STS
dissimilarities (D
OVH
and D
STS
) respectively.
49
S
$YZ
= [WS !"#
U
,!"#
TX
[4-4]
S
\&\
= [WS 4/4,4/4
TX
[4-5]
The three individual dissimilarities are combined into a single similarity score, as defined
by 45W
]&Y,$^_
in equation 4-6. Each feature dissimilarity score is first normalized to
range from 0 to 1, and then assigned a weight that may be modified according to user
preference, or with the potential addition of new features.
45W
]&Y,$^_
= 1− D
&T
∗
S
&T
Q?N S
&T
+D
$YZ
∗
S
$YZ
Q?N S
cYZ
+D
\&\
∗
S
\&\
Q?N S
\&\
[4-6]
GPKℎ 'ℎ?' D
&T
+D
$YZ
+D
\&\
= 1
The weighted feature scores are then added and inverted to yield a score for similarity.
The final similarity score from equation 4-6 defines the query-database OAR similarity
for a single PTV. Some cases have more than one PTV, in which case the algorithm first
calculates a separate similarity score for each PTV, and then takes the average over all
PTVs.
4.4 System Recommendations

The main objective of the decision support module is to pick a set of retrospective OAR
dose profiles to establish benchmarks or standards of reference for assessing potential
dose parameters for the query patient. These reference cases are selected from the
database on the basis of anatomical similarity as described in the previous sections. The
dose profile of an anatomically similar database case can be considered a realistically
achievable objective for the query case. Depending on the size and diversity of the
database, the set of similar cases picked by the system represent a range of possible
50
dose profiles that the query case may be able to
reproduce. The decision support module uses Dose
Volume Histograms (DVHs) to represent OAR dose
profiles, since they are commonly used in clinical
practice to summarize the distribution of radiation
dose within OARs.
Once the decision support module has calculated an
OAR similarity score for the query case with respect to
every database case, a similarity score threshold is
applied to filter out the set of most similar database
cases. Figure 4-5(A) shows a super-position of the left
parotid DVHs of all the database cases, while figure 4-
5(B) shows the DVHs of the most similar database
cases. The DVHs of the similar cases in figure 4-5(B) are
clustered together more tightly than figure 4-5(A),
since anatomical similarity often signifies DVH
similarity. In figure 4-5(C), the red curves show the
95% confidence interval for the mean of all the reference DVHs. These confidence
interval curves are obtained by constructing 95% Confidence Intervals for pre-defined
points on each DVH. They represent the region where the query DVH is most likely to
occur. The reference dose profiles and confidence intervals are transferred to the plan
Figure 4-5 (A) DVH Spread of all
Database Cases (B) DVH Spread of
Similar Database Cases (C) 95% DVH
Confidence Interval of the query case
51
assessment tool (described in section 4.5), which provides a web interface for users to
navigate and interact with the presented results.
4.5 Decision Support Workflow
This section highlights the potential impact of integrating the decision support system
into clinical practice, by analyzing workflow changes and providing step-by-step use-
case scenarios. The use case of this decision support system was determined based on
knowledge gained from the literature, expertise of mentors and my personal
observations. I spent a week at a UCLA Radiation Oncology satellite facility in Santa
Monica, observing and interviewing a dosimetrist and radiation oncologist at their
workstations. The knowledge gained from this experience was very valuable in defining
the goal of my work.
Figure 4-6 shows the regular clinical workflow in grey, while contributions from the
decision support system are highlighted in orange. Steps along the regular workflow
pathway are marked with the letter ‘R’, while steps along the decision support pathway
are marked with the letter ‘D’. The decision support system provides quantified,
knowledge-driven assistance to the clinician at three important junctions in the planning
workflow, as discussed below.
52

Figure 4-6 Contribution of the Decision Support System to the Clinical Workflow
1. Selecting IMRT dose constraints
In regular practice, radiation oncologists and dosimetrists pick IMRT dose constraints
based on prior experience and judgment (R1 in figure 4-6). However, with the decision
support system, they can now refer to retrospective treatment plans of anatomically
similar cases and draw upon best-practice techniques specific to the new patient’s
anatomy (D1 in figure 4-6). In cases where the tumor size and location is too restrictive
and unfavorable, the system is likely to show that other anatomically similar database
cases did not exhibit good OAR dose sparing. This may reinforce the clinician’s own
choice of high OAR dose parameters, since it suggests that an attempt to provide
increased dose sparing will likely not succeed. The decision support system thus saves
considerable time and effort that would otherwise be needed in the subsequent re-
planning loop (feedback from R3 to R2 in figure 4-6). Or the other hand, the system
53
might encourage clinicians to attempt greater OAR dose sparing than thought possible,
by showing similar retrospective cases with low OAR dose. This has the potential to
improve expected patient outcomes indirectly.
2. Treatment plan evaluation
Once the Treatment Planning System (TPS) receives the initial treatment parameters
and IMRT dose constraints, it optimizes a treatment plan to accommodate the specified
constraints, and generates the resulting dose distribution matrix. Clinicians must now
evaluate the safety and effectiveness of the dose distribution by visually assessing the
dose volume histograms (DVHs), dose washes and isodose curves of the optimized plan
for the current patient (R2 in figure 4-6). Once again, the decision support system can
assist clinicians to evaluate the current DVH by comparing it against retrospective DVHs
of similar cases, since this provides a yardstick for comparison (D2 in figure 4-6). This
empirical evidence could help clinicians decide whether further improvements are
possible, thus enabling better patient outcomes; or it may save clinicians considerable
time and effort in failed attempts at re-planning in case no further improvements are
possible.
3. Dose constraint re-adjustments for re-planning
After evaluation (R2 in figure 4-6), if a plan is found to be unsatisfactory, clinicians must
re-adjust the IMRT dose constraints based on their judgment and expertise and prepare
the case for re-planning (R3 in figure 4-6). The decision support system can help
clinicians to identify the next best IMRT dose constraints based on treatment plans of
54
similar retrospective patients (D3 in figure 4-6). This could help avoid future re-planning
(R3 to R2 in figure 4-6) loops, and increase clinician confidence.
4.6 Use Case Scenarios
This section demonstrates the value added by this decision support system to the
treatment planning workflow, using examples drawn from the retrospective test
dataset.
1. Selecting Dose Constraints
Before the TPS can begin plan optimization for IMRT, clinicians must specify dose
constraints as inputs to the planning algorithm. This is usually done by following general
guidelines and by drawing on prior experience with similar cases. These guidelines are
not tailored to individual patient anatomy, and clinician expertise is often subjective.
The decision support system aims to make the planning workflow more objective,
evidence-based and personalized. Consider the case shown in figure 4-7, where
clinicians require decision support for a new patient. The system retrieves the most
similar database cases by setting a similarity score threshold. The dose volume
histograms (DVHs) of all relevant structures are then presented to the clinician through
the interactive plan assessment tool shown marked (A) figure 4-7. The grey curves are
the left parotid DVHs of anatomically similar database cases, while the red curves
represent a 95% DVH confidence interval for the new patient. The clinician can now use
this range to estimate the potential upper and lower bounds for a new patient’s DVH.
Additionally, the concentration of grey curves may be used to infer the regions of
highest probability of occurrence.
55
Picking the D50 constraint: The D50 constraint is “the dose delivered to 50% of the left
parotid”. The 50% volume mark on the y-axis is marked ‘B’ in figure 4-7. D50 is one of
the several dose objectives that clinicians may specify to the plan optimization
algorithm. Clinicians may infer that the dose value corresponding to the middle of the
95% confidence interval (marked point ‘C’) would be a safe choice for the D50
constraint. Point C may be traced back to the x-axis to determine the selected D50 value
of 24Gy, marked point ‘D’. A mean dose of 25Gy is accepted the tolerance limit for the
parotid glands according to the QUANTEC guidelines. Since the D50 parameter of 24Gy
falls under this limit, it conforms to the guidelines and is thus an acceptable choice. The
user may click the appropriate point (dose=24Gy, volume=50%, marked point ‘B’)
onscreen using the cross-hair cursor to select and save this dose constraint.
Selecting multiple dose constraints: In a similar manner, the user may select and save
several dose constraints for the left parotid. Following that, dose constraints may also
be selected for the right parotid, brainstem and any other OARs that are deemed to be
at risk of over-exposure. The saved constraints may be retrieved at a later time, or
entered immediately into the treatment planning system for plan optimization.
56

Figure 4-7 Interactive Plan Assessment Tool used for Selecting IMRT Dose Constraints
Integrated Image Viewer: The left pane of the plan assessment tool in figure 4-7 displays
an interactive CT image stack (part ‘E’ of figure 4-7) and a 3-dimensional model (part ‘F’
of figure 4-7) of the current patient that may be referred to while picking dose
constraints. The user may scroll through the CT stack, and manipulate the 3-dimensional
model to view it from various angles.
2. Treatment Plan Evaluation
Once the TPS has optimized a treatment plan, clinicians examine the resulting dose
distribution to evaluate whether the specified dose constraints were satisfied. At times,
optimizing for specified dose constraints may lead to unexpected negative
consequences, such as over-exposure of other ROIs or lack of dose uniformity resulting
in undesired hot spots or cold spots. If this happens, clinicians must revise their initial
dose constraint estimates to resolve such conflicts, leading to multiple re-planning
57
iterations (R3 to R2 in figure 4-6) until the plan is deemed acceptable. Without the
decision support system, dose constraint revisions are based solely on individual
judgment. In contrast, utilizing the system provides empirical reference standards for
evidence-based assessment. With respect to the example shown in figure 4-7, the TPS
optimized a plan based on the left parotid D50 constraint of 24Gy. The optimized plan
resulted in a dose distribution that generated the left parotid DVH highlighted in bold
blue in figure 4-8, marked as ‘A’.

Figure 4-8 Interactive Plan Assessment Tool for Treatment Plan Evaluation
Plan Evaluation with decision support: Since the Dose Volume Histogram (DVH) is a
popular tool used by clinicians in treatment plan evaluation, it has been utilized for the
same purpose by the decision support system as well. Figure 4-8 shows the most
recently optimized DVH as a thick blue line, superimposed on the DVHs of anatomically
similar database cases. Although this DVH conforms to the specified D50 parameter of
58
24Gy, it exceeds the 95% confidence interval (red curves) under the 40% volume mark,
as shown by the green ellipse marked ‘B’. Since the confidence interval indicates the
region of highest probability of occurrence for the new DVH, figure 4-8 indicates that it
might be possible to lower the dose in the bottom half of the curve to fall within the
95% Confidence Interval range, as indicated by the green ellipse marked ‘C’. This would
have been difficult to identify without the decision support system, and could have led
to acceptance of an inferior plan. Instead, the clinician can now attempt to add further
dose constraints in order to improve the plan further.
Dose constraint re-adjustments: Since the currently optimized plan was not approved,
the clinician attempts to incorporate an additional constraint in order to draw the lower
half of the curve into the 95% confidence interval band marked in red. The chosen dose
constraint is marked with the cross-hair cursor within the green ellipse labeled ‘C’ in
figure 4-8: D30 < 35Gy. This means that the dose delivered to 30% of the volume should
not exceed 35Gy.
Result of dose constraint revisions: Re-planning in order to optimize the newest dose
constraint results in a DVH shown in bold blue in figure 4-9, where the newest D30
constraint has been achieved as indicated by the region marked ‘A’. The interactive tool
shows that this is empirically one of the best possible DVHs for the left parotid, leading
to the lowest dose profile for the left parotid glands. With this reassurance, the clinician
accepts the plan and starts preparing for treatment delivery. In this case, the decision
support system increased clinician confidence by providing empirical standards for
59
comparison and facilitated an early termination of the re-planning loop, thus saving
considerable time and effort.

Figure 4-9 Interactive Plan Assessment Tool for Re-adjusting IMRT Dose Constraints
This chapter has described the decision support module in depth, including the features
used for calculating anatomical similarity, the methodology used for retrieving relevant
retrospective dose distributions, the concept and GUI design of the plan assessment
tool, as well as a demonstration of the impact of the decision support module on the
clinical workflow. The next chapter reports the methodology and results of evaluating
the decision support module.

60
Chapter 5 Evaluation
Evaluation of the decision support system was conducted in two parts.
(i) Similarity Matching: Validating the ability of the similarity matching algorithm to
detect differences in anatomical PTV characteristics such as location, volume and
shape.
(ii) Dose Profile Matching: Evaluating the system’s ability to extract relevant dose profiles
for the query case by testing the hypothesis that dose similarity may be deduced
through anatomical similarity.
5.1 Data Collection

The dataset used for evaluating the decision support system consists of 137 retrospective head
and neck cancer treatment plans. This is also the dataset that will eventually constitute the
knowledge base when the system is deployed for pilot studies or for clinical implementation. For
a new prospective patient, a search query will utilize all 137 retrospective database cases for
decision support. This data was collected from the Department of Radiation Oncology at the
University of California Los Angeles (UCLA). An initial dataset of 200 cases was filtered to remove
instances with missing data, incorrect labeling and inconsistent contouring. All DICOM files were
anonymized at the clinical site using DICOMCleaner, a freely available open source tool, in order
to remove any identifying information related to the patient, physician and institution including
name, ID, accession number, address, birth date and sex. Table 5-1 describes the data objects
contained within a single retrospective ‘case’ associated with a patient.

61
Table 5-1 Data Objects contained within dataset
Data Object Number of instances per case File Format
CT Image Series 85 to 195 DICOM
Structure Set 1 DICOM
Dose 1 DICOM
Treatment Plan 1 DICOM

All collected plans were Intensity Modulated Radiation Therapy (IMRT) cases, planned with the
Inverse Treatment Planning (ITP) approach. Table 5-2 describes some relevant characteristics of
the dataset. Several plans received boost treatments, wherein multiple targets (PTVs) are
defined, each with a different prescribed radiation dose. Sequential boost IMRT delivers
treatment in stages, each stage corresponding to a different PTV. The decision support system
considers treatment plans belonging to different stages in sequential boost treatments as
separate cases. The Simultaneous Integrated Boost (SIB) methodology treats all the PTVs in one
stage, such that each PTV receives different dose exposures simultaneously, resulting in more
than one defined PTV per case, as shown in Table 5-2. There are a total of 52 SIB cases in the
dataset, and as a result, the decision support module was equipped to deal with multiple PTVs.
Table 5-2 Characteristics of the dataset used for development and testing
Number of Targets PTV* Dose OAR** Overdose Treatment Modality
No. of
PTVs*
No. of
Cases
Mean 86.7Gy OAR % Cases Modality % Cases
1 72
Standard
Deviation
16.1Gy
Parotid
Glands
60
Tomotherapy
(Accuray)
43
2 13 Maximum 106.2Gy Mandible 70
Novalis
(Varian)
22
3 52 Minimum 12.2Gy Brainstem 35
TrueBeam
(Varian)
35
* PTVs: Planning Target Volumes **OAR: Organ At Risk

62
The mean dose delivered to each PTV depends on the prescription specified by the radiation
oncologist, and ranges from 12Gy to 106Gy. The prescription dose has a significant impact on the
ultimate dose delivered to OARs. Also, since one of the primary goals of this system is to help
clinicians minimize OAR dose, it is important to assess the number of cases receiving a dose
greater than the recommended tolerance limit. It can be observed that the rate of overdose is
significantly high in case of the mandible and the parotid glands, but much lower in case of the
brainstem. Details regarding tolerance limits and the impact of this trend will be discussed later
in this chapter. The table also shows that various treatment modalities are well represented in
this dataset.
The meta-data from all the DICOM files was extracted by the DICOM parser and stored in the
database for efficient query-retrieval.
5.2 Validation of the Anatomical Similarity Matching Algorithm

The methodology used to validate the system’s ability to detect differences in anatomical PTV
characteristics is analogous to creating virtual ‘digital phantoms’. A randomly selected test case
‘O’ forms the baseline for comparison. The segmented PTV object of a randomly selected test
image is subjected to a number of morphological or image warping transformations to yield a set
of transformed PTV objects. The magnitude of the transformation coefficients represents the
anatomical dissimilarity between the test PTV and its transformed PTV. Thus, “known”
dissimilarities have been artificially introduced to the original PTV. This known dissimilarity is
then compared to the dissimilarity scores calculated by the decision support system. It is
important to note that only the PTV object is manipulated, while the original OAR objects remain
undisturbed and intact to preserve the frame of reference for calculating similarity.
63
Three types of transformations were utilized to simulate different characteristics, as shown in
figure 5-1. Transformations were introduced with respect to the right parotid gland, and the
decision support system also considered the right parotid as the reference OAR for calculating
similarity.
1. Translation is implemented along the x, y and z axes simultaneously to simulate
differences in PTV location and distance to the OAR.
2. Scaling, implemented either via morphological erosion or dilation (picked at random),
simulates differences in total PTV volume, as well as volumetric distribution due to
addition or deletion of voxels.
3. Shearing is implemented for the x-y, y-z and x-z planes along the z, x and y axes
respectively in order to simulate differences in PTV shape, volumetric distribution,
location and distance. It is randomly applied either along the x, y or z axis at for a single
instance.
64

Figure 5-1 The three transformations applied in order to introduce artificial dissimilarity to the test case. ROIs:
Parotid glands in dark blue, brainstem in green and the PTV in red.
Ten test cases were selected at random. Each transformation was randomly applied 50 times to
each test case. Overall, this resulted in a total of 500 instances for evaluation. 500 transformation
coefficients representing the “known” or “introduced” dissimilarity were correlated with 500
dissimilarity scores calculated by the decision support system. The resulting correlation
coefficients were used to validate the ability of the decision support system to accurately rank
anatomical differences, or similarities. The results are presented in table 5-3. For each type of
transformation, the mean correlation coefficient was greater than 0.9, which indicates a strong
correlation. Smaller transformations represented small anatomical differences between the
original and transformed image, which led to lower dissimilarity scores. Larger transformations
represented greater differences in anatomy, which led to higher dissimilarity scores. This shows
that the decision support algorithm is good at detecting and quantifying differences in PTV shape,
size, distance and location with respect to a particular OAR.
65
Table 5-3 Correlation between artificially introduced dissimilarity represented by the translation coefficient, and the
dissimilarity calculated by the decision support algorithm
Transformation Translation Coefficient Average Correlation between the
transformation coefficient and
dissimilarity score for
Translation Radial distance 0.94
Scaling Size of Structuring element 0.9
Shearing Shearing Factor 0.91

5.3 Evaluation of the decision support module

The decision support system assists clinicians in achieving good dose end-points for new
prospective patients by showing them a range of the most probable dose distributions based on
the patient’s anatomy and empirical evidence gained from a database of retrospective treatment
plans. Therefore, the decision support system must identify retrospective dose distributions that
are most likely to resemble the new patient’s dose distribution. Since every dose distribution is
governed by the spatial relationship between a PTV and its surrounding critical structures, the
new patient’s dose profile is more likely to match the dose profiles of anatomically similar
patients than those of anatomically dissimilar patients. Thus, the system’s ability to extract
relevant dose distributions can be demonstrated by validating the claim that dose similarity may
be deduced through anatomical similarity.
Figure 5-2 shows an example comparing two cases – (a) and (b). The anatomical similarity score
is 95.4%, which is high. This can be confirmed through observation – there is a strong
resemblance in shape and volume of the PTV (yellow) as well the location of the PTV with respect
to the two parotid glands (dark blue), the mandible (red) and the brainstem (light blue).
66

The Dose Volume Histograms
(DVHs) of the right parotid,
mandible and brainstem are also
similar, while the left parotid DVHs
are so similar that they overlap to a
great extent. The Dose Volume
Histogram, as explained in greater
detail in Chapter 2, describes the distribution of dose across the OAR volume, and hence similar
DVHs indicate similar dose distributions. This example illustrates the trend that the decision
support system relies on. The evaluation stage of my research
investigates whether the entire dataset expresses this trend with
adequate statistical significance.
In order to carry out this evaluation, new prospective patients were
simulated with retrospective test data. The dose distribution of the
simulated prospective cases is known and may thus be regarded as a
gold standard for comparison. The dose profiles of various critical
structures are represented by their Dose Volume Histograms (DVHs),
and similarity between pairs of DVHs is calculated using the Earth
Mover’s Distance [24]. The entire set of 137 retrospective treatment
plans was utilized to conduct a leave-one-out cross validation for
Figure 5-2 Relationship between anatomical similarity and DVH
similarity
Figure 5-3 Leave-one-out Cross
Validation
67
determining the success rate of retrieving cases with matching dose distributions. Each test case
was used to simulate a new prospective query case, resulting in 137 iterations as shown in figure
5-3. In iteration #1, the first database case simulates the query patient and is compared to the
rest of the database (excluding case #1 itself); in iteration #2, the first database case is returned
to the pool, and the second database case simulates a new query patient and is compared to the
rest of the database (excluding case #2 itself); and so on.
5.3.1 Evaluation Per Iteration

For each simulated prospective patient P
q
(query case), i.e. for each iteration, I tested whether
the Dose Volume Histogram (DVH) of P
q
resembled the DVHs of similar database cases more than
the DVHs of dissimilar database cases. This concept is illustrated in figure 5-4.
Figure 5-4 Dose Volume Histogram comparison of anatomically similar and dissimilar cases
68
Part (a) shows the left parotid DVHs of all the database cases in grey, superimposed on the DVH
of P
q
, shown in light blue. The database cases are then split into two groups – similar and
dissimilar, based on a dynamic threshold. Part (b) of figure 5-4 shows the DVHs of similar
database cases, which are clustered close to P
q
, indicating similarity of dose profiles. Part (c)
shows the DVHs of dissimilar database cases, which are scattered further away from P
q
, indicating
dissimilarity of dose profiles. The mean spread of database DVHs for the ‘similar’ group shown in
figure 5-4(b) is compared to the mean spread of database DVHs of the dissimilar group, shown
in figure 5-4(c).
In order to accomplish this, the query DVH P
q
is first compared to each database DVH from the
‘similar’ group, as well as the ‘dissimilar’ group, by calculating the Earth Mover’s Distance (EMD)
between each query-database pair. Parts (d1), (d2), (d3) and (d4) of figure 5-4 show the query
DVH P
q
being compared to four database DVHs from the ‘similar’ group. Parts (e1), (e2), (e3) and
(e4) of figure 5-4 show the query DVH P
q
being compared to four database DVHs from the
‘dissimilar’ group. The Earth Mover’s Distance was chosen because it is widely used to quantify
the distance between two discrete probability density functions or histograms, as explained in
Chapter 4. The mean spread of each group is given by the mean of the Earth Mover’s Distance of
each database DVH from P
q
, as outlined in equations 5-1 and 5-2.
!"#$ &!'
()*)+,-
=
1
$
&!' '01
2
,'01
)
4
)56
[5-1]
7ℎ"9" $ = $:;<"9 => ?@;@A#9 B#?"?, '01
)
∶ '01 => Dℎ" @
EF
B#?" @$ Dℎ" ′?@;@A#9
H
I9=:J
!"#$ &!'
K)(()*)+,-
=
1
;
&!' '01
2
,'01
)
*
)56
[5-2]
7ℎ"9" ; = $:;<"9 => L@??@;@A#9 B#?"?, '01
)
∶ '01 => Dℎ" @
EF
B#?" @$ Dℎ" ′L@??@;@A#9
H
I9=:J
69
In order to prove that the system successfully identified retrospective dose profiles that are most
likely to resemble the dose profile of the new patient, I tested the claim made in equation 5-3.
!"#$ &!'
()*)+,-
< !"#$ &!'
K)(()*)+,-
[5-3]
This tests whether the DVHs in figure 5-4(b) are clustered closer to the query DVH P
q
(light blue)
than the DVHs in figure 5-2(c). The Wilcoxon Rank Sum test was conducted to test equation 5-3.
If the hypothesis in equation 5-3 is satisfied with statistical significance, then the system is
successful in identifying relevant retrospective dose profiles for the current test case P
q
. Thus,
the result either positively verifies the system’s ability to extract cases with similar dose
distributions (indicating conclusive success), or fails to provide a conclusive verification.
5.3.2 Evaluation for the entire dataset

The Wilcoxon Rank Sum test is performed on each of the 137 test cases in a leave-one-out
methodology as described earlier, and the number of iterations that produce a statistically
significant verification of success are counted. This count is a measure of the success rate of the
system’s algorithm. This evaluation procedure was conducted separately for four different critical
structures – the left parotid, the right parotid, the mandible and the brainstem. The results of the
evaluation are outlined in table 5-4.
Table 5-4 Success rate of the decision support algorithm in identifying relevant database cases for
comparison

Organ At Risk Number of Iterations
with conclusive
verification
Total
Number of
Iterations
Conclusive
Success Rate
(%)
Significance
Left Parotid 95 118 81.35 p < 0.001
Right Parotid 89 121 73.55 p < 0.001
Mandible 97 125 77.6 p < 0.001
Brainstem 74 127 58.27 p = 0.06

70
The success rates for the left parotid gland, right parotid gland and the mandible are high –
81.35%, 73.55% and 77.6% respectively. This means that there is an 81.35%, 73.55% and 77.6%
chance of the algorithm providing relevant retrospective dose profiles for the left parotid, right
parotid and mandible respectively. The algorithm’s similarity score for these critical structures is
an effective indicator of dose distribution similarity, with a statistical significance of p<0.001.
It must be noted that the dose profiles of the retrospective cases formed the gold standard. Since
there is no guarantee of the current gold standard representing the most optimal dose profiles,
it is difficult to attain a 100% success rate due to inevitable outliers. Moreover, the limited size
of the test dataset resulted in inadequate ‘similar’ cases for certain iterations. Beyond these two
factors, the discussion section proposes improvements that may be incorporated to increase
algorithm success rates.
The original plan for evaluating the decision support algorithm included the parotid glands,
mandible and brainstem because these were the most frequently occurring and the most
consistently contoured structures in all treatment plans. However, after observing the relatively
lower success rate for the brainstem, I related it back to the low overdose rate for the brainstem
mentioned earlier in this chapter. Further, I discovered that the maximum brainstem dose is 46.9
Gy on an average across all the database cases. This is in compliance with the recommended
tolerance limit of 54Gy. On an individual basis, 75% of cases comply with this recommended limit,
as compared to other OARs, which have much lower compliance rates ranging from 41% to 30%,
as shown in Table 5-5.

71
Table 5-5 Tolerance dose compliance rates of OARs

Organ At
Risk
Tolerance
Parameter
Tolerance
Limit
% Cases that
comply
Brainstem Maximum Dose 54 Gy 75
Right
Parotid
Mean Dose 25 Gy 39
Left Parotid Mean Dose 25 Gy 41
Mandible Maximum Dose 70 Gy 30

After consulting with our collaborating senior radiation physicist, Dr. John DeMarco, we
concluded that since the brainstem did not get overdosed frequently, clinicians did not actively
attempt to reduce the brainstem dose in 75% of cases. As a result, the anatomical PTV-brainstem
relationship did not actively impact the dose distribution in 75% of cases. Further investigation
revealed that the average volumetric overlap of the brainstem with the PTV is 0.04%, with a
standard deviation of 0.32%. This is much lower compared to the other critical structures, as
shown in Table 5-6, which have a higher mean overlap as well as greater variation in overlap.
Table 5-6 Trends of volumetric overlap with the PTV for the various Organs At Risk

Organ At Risk Volumetric Overlap (%)
Mean Standard Deviation
Brainstem 0.04 0.32
Right Parotid 0.81 2.1
Left Parotid 1.09 3.08
Mandible 3.72 7.6

PTV overlap often leads to OAR overdose. But the data in table 5-6 shows that this is not a strong
concern for the brainstem, reinforcing the idea that the ultimate dose distribution is not very
heavily influenced by the PTV-brainstem relationship. This conclusion was also verified by the
collaborating clinical expert. In clinical practice, most of the current cases would not have
involved any concern for brainstem overdose, and it is unlikely that clinicians would have utilized
72
the decision support system for the brainstem due to the relatively low risk of overdose in those
cases.
In conclusion, I validated the accuracy of the similarity matching algorithm and verified the ability
of the decision support module to retrieve relevant retrospective dose distributions for helping
clinicians select good dose constraints. The next chapter explores ways and strategies for
enriching the knowledge base in order to improve the performance of the decision support
system even further.

73
Chapter 6 Multi-Site Collaborative Data Aggregation

The performance and efficacy of knowledge-driven decision support systems depends to a great
extent on the size and diversity of the knowledge base. Variation in the knowledge base is
particularly important to ensure that various combinations of Planning Target Volume anatomy
are represented so that every query case produces an adequate number of database matches for
comparison. Data diversity may also be enhanced by representing the best practices from
multiple institutions in the knowledge base. This is important because each institution usually
follows a set of standard treatment practices and protocols. This may be the result of prescribed
departmental guidelines or established practices passed down from experts to trainees.
Combining data from multiple institutions results in clinicians benefiting from a range of
techniques from different institutions, thus drawing from the strengths of each.
6. 1 Challenges in Implementing Collaborative Data Sharing

Implementation of multi-institutional data sharing presents a logistic challenge due to the Health
Insurance Portability and Accountability Act (HIPAA). The HIPAA Privacy and Security rules
regulate sharing of Protected Health Information (PHI) by “covered entities” such as hospitals,
health insurers, etc. for research purposes. Protected Health Information (PHI) is defined in 45
CFR 160.103 (where CFR means the Code of Federal Regulations) as individually identifiable
health information that is either transmitted or maintained by electronic or any other form or
medium. The HIPAA Privacy Rule covers protected health information in any medium while the
HIPAA Security Rule covers electronic protected health information. However, existing data or
documents may be collected without any conflict with the HIPAA Privacy Rule if they have been
74
completely de-identified according to the requirements laid out in Section 164.514 (a) to (c) of
the Code of Federal Regulations (CFR). In order to comply with regulations, the de-identification
procedure must be conducted by the contributing institution within its internal network. This
involves using a third-party application for anonymization, followed by a thorough check of every
DICOM field to ensure the absence of identity-related elements. Private DICOM tags further
complicate this procedure, since the contents of these tags are not known in advance. This
requirement could discourage participation due to the manual effort involved. I have developed
a simple HIPAA-compliant data sharing protocol in order to circumvent some of these obstacles.
6.2 RT Data Sharing Research Initiatives

There have been some recent efforts at building solutions to support large-scale collection of
radiation therapy data for research purposes. Two examples that stand out are the ‘Transfer of
Images and Data’ (TRIAD) platform [17] and the ‘National Radiation Oncology Registry’ (NROR)
[18]. TRIAD is a medical data sharing platform sponsored by the American College of Radiology
(ACR) that enables exchange of DICOM and non-DICOM data from various participating sites for
the purposes of clinical trials, accreditations, etc. The Radiation Therapy Oncology Group (RTOG)
has chosen to use TRIAD in collecting data for its clinical trials. However, since this platform is
intended to facilitate the flow of data involved in sharing and collecting files, it does not address
the need to catalogue this data into databases for easy query and retrieval. This is one of the key
components of the system presented here. Additionally, it requires installation of de-
identification software, which further entails platform dependence and regular installation of
software updates.
75
The NROR aims to build a large-scale national registry of treatment delivery and health outcomes
data to fuel a broad range of research and quality assurance activities. It functions on the basis
of a Business Associate agreement between NROR and participating sites, and as such, it does
not require a de-identification module since it is considered a “covered entity” by the HIPAA
Privacy Act. As a result, only tried and tested decision support algorithms can be integrated with
it. However, research and validation of new decision support algorithms cannot be conducted
with NROR data. This decision support system, in contrast, provides a small but robust platform
for developing and testing new decision support techniques. Moreover, the modular nature of
this system allows future integration with systems such as NROR. For instance, the decision
support component or the client-side metadata extraction tool may be integrated into the NROR,
or data from other systems may be integrated into the current decision support system.
6.3 HIPAA-Compliant Data Sharing Protocol

The HIPAA-Compliant Data Sharing Protocol developed as part of this project circumvents the
need for tedious anonymization by employing a client-side DICOM parser that extracts a limited
number of de-identified DICOM attributes through a web browser. The cornerstone of this
protocol is a list of relevant DICOM attributes, that is identified in advance for every DICOM file
type. These attributes are carefully examined to ensure that they do not contain any patient
identifiers. Only the contents of these pre-approved attributes are extracted from the DICOM
files. This parsing and extraction is carried out on the client’s machine, and so the DICOM files
never leave the local file system. Since this extracted content is guaranteed to be anonymous, it
can safely be transferred to the system server. This data sharing protocol incorporates the
following steps, as indicated by the corresponding numbers in Figure 6-1.
76

1. The protocol is initiated by a data-sharing request from the client. This is an indication
from a user wishing to upload data to the server.
2. The server sends a JavaScript / HTML5 package to the client. This contains a DICOM parser
written in JavaScript, which has the capability to parse DICOM files on the client’s machine
through the user’s browser, and extract relevant metadata from these DICOM files.
3. The client’s local web browser executes this JavaScript package as soon as it is received.
4. The user is prompted to select a folder containing the DICOM files to be uploaded.
5. For every DICOM file that is identified, values from the pre-approved list of attributes are
extracted, and reformatted for transmission to the server.
6. A pseudo ID is requested from the system for every patient ID that is encountered in the
set of DICOM files marked for upload.
Figure 6-1 A HIPAA-Complaint Data Sharing Protocol
77
7. A mapping between these pseudo IDs and the DICOM Patient IDs is created and stored in
an encrypted format on the client’s file system, along with the encryption key.
8. The reformatted data is sent to the centralized server for archival, post-processing and
knowledge extraction.
This protocol resolves a lot of difficulties associated with data collection, which other groups such
as the RTOG are currently facing. Two key features from this protocol distinguish it from other
data collection systems. First is the client-side parser that allows users to access this service by
utilizing their local web browsers. This is a zero footprint approach, which circumvents the need
to download and install additional components, and provides platform independence. Programs
may be revised and updated regularly on the server without requiring the client to install newer
software versions, as is the case with desktop applications. Secondly, extraction of a limited set
of pre-approved data attributes from DICOM files guarantees de-identification, thus eliminating
the need for contributors to perform anonymization on their end. If the user so desires, an
encrypted mapping may be saved between the original patient ID and the system-generated
pseudo-ID to enable upload of follow-up data in the future.
6.4 Integration of Multi-Institutional Data into the Decision Support System

The parser, database and decision support algorithms were initially developed and tested with a
dataset of 50 head and neck treatment planning cases collected from the UCLA Department of
Radiation Oncology. This dataset was further supplemented with 87 cases from UCLA. These cases
were exported from three different treatment planning systems belonging to two different
vendors – RapidArc from Varian, TrueBeam from Varian and the TomoTherapy system from
78
Accuray. This demonstrated the system’s ability to aggregate data from different treatment
planning systems.
While the decision support system was undergoing evaluation, I also collected 52 head and neck
cancer treatment planning cases from the Techniche Universitat Munchen (TUM), Germany.
These cases were exported from Varian’s Clinac system and Accuray’s TomoTherapy system. The
TUM dataset was successfully parsed and integrated into the database, further demonstrating the
system’s robustness and compatibility with different data sources. Since different institutions and
clinicians follow differing naming conventions for ROIs, the ROI names had to be manually
standardized for the decision support system.
Treatment planning data was thus collected from two different institutions – 137 cases from UCLA
and 52 cases from TUM, in order to test the system’s ability to support multi-institutional data.

79
Chapter 7 Discussion & Conclusion
The design, prototyping and evaluation of a knowledge-driven radiation therapy decision support
system has been completed. Original contributions made during the course of this project include
the following:
(i) Development and testing of features that quantify the spatial relationship between
the Planning Target Volumes and Organs At Risk
(ii) A similarity matching algorithm that utilizes those features
(iii) Results of evaluating the system’s decision support module
(iv) A robust DICOM parser and database
(v) Design and prototype for an interactive Graphical User Interface
(vi) A HIPAA-compliant data sharing protocol for multi-institutional collaboration
This work was evaluated in multiple stages with a total of 192 treatment planning cases of head
and neck cancer patients, collected in collaboration with UCLA and TUM - two institutions from
different countries. This attests to the robustness of the system, and shows promise for future
compatibility with data from multiple sources.
The contributions listed above may be accessed by the scientific and clinical community either
by replicating the system specifications outlined in journal articles, by downloading some of the
open-source components that are being developed, or by collaborating with the IPILab research
group. Individuals and groups seeking to replicate and expand this work may do so independently
in separate institutions, or choose to form multi-institutional collaborations. Another avenue for
launching this system is integration with or adoption by a larger consortium such as the RTOG or
80
regional and national cancer registries such as NROR. Yet another potential approach to utilizing
some of the decision support methods outlined in this work is by direct integration into existing
commercial treatment planning systems. I intend to continue my work in collaboration with the
IPILab and other interested institutions to refine my existing methodology and develop a finished
product, ready for deployment in the clinical environment.
7.1 Targeted User Community

In the course of treating a new patient with radiation therapy, treatment planning experts often
draw upon memories of anatomically similar patients from personal experience. The system’s
ability to find anatomically similar database cases is analogous to providing clinicians with virtual
clinical experience in treatment planning. This is particularly useful to residents, trainees and new
clinicians with a limited background in treatment planning. The system can also benefit more
experienced radiation oncologists and dosimetrists by providing quantified evidence-backed data
to reinforce their decisions, and by allowing them to draw from the techniques and standards
employed by other clinicians.
The treatment planning personnel who could benefit from this system include dosimetrists and
radiation oncologists. Radiation oncologists are responsible for prescribing the treatment
protocol and target dose, delineating the RT target, providing initial treatment parameters as
well as supervising and signing off on the final treatment plans. Dosimetrists, on the other hand,
are responsible for fulfilling the radiation oncologist’s requirements, fine tuning treatment plan
parameters and performing the treatment plan iterations that result in an acceptable plan. This
decision support system can benefit both oncologists and dosimetrists in their respective roles.
Advanced queries on the data collected in the knowledge base may also benefit radiation
81
physicists in performing regular quality control checks, evaluating departmental safety
compliance, comparing treatment machines and protocols, etc.
7.2 Expanding the Scope of Research
Some of the concepts and modules developed as a part of this research may translate to
applications in other areas. For instance, the multi-institutional data sharing protocol developed
for this project may be utilized for collaborative research enterprises beyond radiation therapy
since it provides HIPAA compliance. One potential use case is collection and cataloguing of
radiological images for longitudinal research studies, such as identification of disease biomarkers.
The radiation therapy decision support knowledge base itself may be expanded, maintained and
modified for a number of possible research endeavors. Well-organized dose-volume data from
treatment plans is very valuable for knowledge discovery in oncology research and epidemiology;
development of more sophisticated treatment planning methods; quality control; tracking
departmental and institutional trends; etc.
7.3 Challenges Encountered

The knowledge base integrates treatment planning data generated by different clinicians from
different institutions, through a variety of treatment planning systems. The biggest challenge
encountered while doing this is the lack of a standardized ontology for dealing with Regions of
Interest such as target volumes and critical structures. Until a common naming convention is
standardized and adopted universally, there is a need for tools that can facilitate semi-automatic
standardization of ROI names.
Another commonly faced issue is that of ROIs that are artificially generated to manipulate the
TPS to produce better optimization results. These artificial ROIs or “optimization” structures are
82
often subsets of real ROIs, most often intersecting regions of critical structures and the PTV as
well as areas that are prone to hot spots or cold spots. Sometimes, the non-intersecting parts of
a critical structure may constitute a new “optimization” ROI, such that only the non intersecting
parts of the organ are used for optimization. Segmentation tools that help resolve such conflicts
may need to be incorporated into the system to counter this problem.
7.4 Suggested System Improvements for Workflow Enhancement
During the development and testing phase, I identified several potential improvements for
inclusion in future implementations of the system.
(i) Currently, data must be manually exported in DICOM format from the treatment
planning systems before being uploaded to the decision support system. This step can
be made either automatic, or at least semi-automatic, by developing a module that
interfaces directly with various treatment planning systems using the Application
Programming Interfaces (APIs) provided by their vendors. This is an elaborate
undertaking, mainly because of the need to integrate with multiple treatment
planning systems from different vendors. Additionally, these APIs are often not well-
documented, and developers may need to liaise with TPS vendors to complete this
task successfully.
(ii) Similarly, it would help to include IMRT dose constraints as DICOM attributes for
decision support. As of now, dose constraint selection and plan evaluation is
conducted on the basis of the DVH alone, and dose constraints for the new patient
are extrapolated from the DVHs of similar retrospective patients. API calls to transfer
dose constraints selected by clinicians from the decision support system directly to
83
the TPS would also facilitate the decision support workflow and complete the
feedback loop within the clinical workflow.
(iii) The GUI plan assessment tool currently shows DVHs of anatomically similar cases, as
well as the tentative DVH of the new query patient during evaluation. In order to assist
the evaluation process, the GUI could include a visual depiction of preferred dose-
volume guidelines superimposed on the dose volume histograms. These guidelines
could be well-known clinical tolerance limits from the literature, such as QUANTEC, or
they could be institution-specific standards or recommendations. The guidelines
could then be further modified by specific users based on their personal preferences.
Automatic dose checks could create alerts if the selected dose constraints or tentative
query DVHs violate these guidelines.
(iv) Finally, the STS and OVH bin intervals have currently been set so as to achieve feasible
run-times in MATLAB. However, if the programs were to be optimized and
implemented as executables, histogram bin intervals could be feasibly narrowed, thus
increasing resolution and potentially improving system performance.
7.5 Scope and Direction of Future Research
The decision support system has been successfully prototyped, the decision support module has
been evaluated and certain system components have been partially optimized. However, in order
to integrate the system into the clinical environment, it is essential to fully optimize the system
code and develop a clinically usable product that is fast and user friendly. Once the system is
ready for clinical deployment, the immediate next step is to perform a clinical study that
measures the impact of the system on the treatment planning workflow. Retrospective cases
84
that show a possibility of increased dose sparing will be re-planned based on system input, and
resulting changes to the OAR dose will be measured, such that a reduction in OAR dose indicates
a successful outcome. A clinical study such as this will rely on further collaborations, since it
requires participation from radiation oncologists and dosimetrists. After the current decision
support system has been successfully tested for head and neck cancer, further research should
be conducted to expand its application to other disease sites such as prostate cancer, breast
cancer, etc.
The final step in the decision support pipeline is the incorporation of follow-up data to link the
system’s recommendations to tangible patient outcomes. This outcomes data may include
radiation toxicity assessments, survival data or even post-treatment follow-up imaging studies.
This would help to quantify the direct impact of radiation on the patient’s quality of life as
opposed to assuming an implicit linear relationship between the dose delivered and the toxicity
incurred. However, outcomes measures are usually subjective and difficult to quantify, often
involving questionnaires with ratings and scale systems. Also, multiple factors can affect
outcomes – patient history, compounded effects of multiple treatment types such as surgery and
chemotherapy, etc. This leads to a host of factors that must be accounted for, and expands the
scope of research from radiation therapy to oncology in general.
7.6 Final Thoughts

Although research in healthcare has historically been the domain of chemists and biologists, the
most recent advancements have been fueled by physicists and engineers in the form of imaging
technologies, implants, robotic surgery and radiation therapy. My research attempts to help in
revolutionizing the landscape of medicine by exploiting the possibilities offered by information
85
technology. Knowledge discovery and personalized medicine are very favorably poised to ride
the wave of big data analytics and high performance computing, and become permanent fixtures
in the world of healthcare delivery. It is my hope that the techniques developed and evaluated
herein will serve as a stepping stone in this metamorphosis.

86
Bibliography

[1] Digital Imaging and Communications in Medicine (DICOM) Part 3. Information object
definitions. PS3.3 (2011)
[2] Huang HK (2010) PACS and Imaging Informatics: Basic Principles and Application. Wiley-
Blackwell Ed. 2
[3] Law MYY and Liu BJ (2009) Informatics in Radiology: DICOM-RT and Its Utilization in
Radiation Therapy. Radiographics 29(3):655-667
[4] Law MY, Huang HK (2003) Concept of a PACS and imaging informatics-based server for
radiation therapy. Comput Med Imaging Graph 27:1–9
[5] Law MY (2005) A model of DICOM-based electronic patient record in radiation therapy.
Comput Med Imaging Graph 29:125–136.
[6] Le AH, Liu BJ, Schulte R, Huang HK (2011) Intelligent ePR system for evidence-based research
in radiotherapy: proton therapy for prostate cancer. Int J CARS 6:769–784.
[7] Le AH (2010) Mining an ePR System using a Treatment Plan Navigator for Radiation Toxicity
to Evaluate Proton Therapy Treatment Protocol for Prostate Cancer. PhD Dissertation
[8] Radiation Therapy Oncology Group Core Lab TRIAD.
http://www.rtog.org/CoreLab/TRIAD.aspx
[9] Efstathiou JA, Nassif DS, McNutt TR, et al (2013) Practice-based evidence to evidence-based
practice: building the National Radiation Oncology Registry. J Oncol Pract 9(3): e90-5.
[10] Palta JR, Efstathiou JA, Bekelman JE, Mutic S, et al (2012) Developing a national radiation
oncology registry: From acorns to oaks. Pract Radiat Oncol 2(1):10-7. doi:
10.1016/j.prro.2011.06.002
[11] Argiris A, Karamouzis M, Raben D, Ferris R (2008) Head and neck cancer. Lancet
371(9625):1695-709
[12] Jensen SB, Pedersen AM, Vissink A, et al (2010) A systematic review of salivary gland
hypofunction and xerostomia induced by cancer therapies: prevalence, severity and impact on
quality of life. Support Care Cancer 18 (8): 1039-60
[13] Deasy JO, Moiseenko V, Marks L, Clifford Chao KS, Nam J, Eisbruch A (2010) Radiotherapy
Dose-Volume Effects on Salivary Gland Function. Int. J. Radiation Oncology Biol. Phys.
76(3):S58–S63
[14] Jereczek-Fossa BA and Orecchia R (2002) Radiotherapy-induced mandibular bone
87
complications. Cancer Treat Rev. 28(1):65-74
[15] Bhandare M and Mendenhall WM (2012) A Literature Review of Late Complications of
Radiation Therapy for Head and Neck Cancers: Incidence and Dose Response. J Nucl Med Radiat
Ther S2:009. doi: 10.4172/2155-9619.S2-009.

[16] Emami B, Lyman J, Brown A, et al (1991) Tolerance of Normal Tissue to Therapeutic
Irradiation. Int J Radiat Oncol Biol Phys. 21(1):109-22.
[17] Bentzen SM, Constine LS, Deasy JO, Eiscruch A (2010) Quantitative Analyses of Normal
Tissue Effects in the Clinic (QUANTEC): An introduction to the scientific issues. Int. J. Radiation
Oncology Biol. Phys. 76(3): S3–S9
[18] Marks LB, Yorke ED, Jackson A, et al (2010) Use of Normal Tissue Complication Probability
Models in the Clinic. Int. J. Radiation Oncology Biol. Phys. 76(3):S10–S19
[19] HI Amols, CC Ling, SA Liebel (2003) Overview of the IMRT Process. A Practical Guide to
Intensity Modulated Radiation Therapy Chapter 2
[20] Beadle BM, Liao KP, Elting LS, Buchholz TA, et al (2014) Improved survival using intensity-
modulated radiation therapy in head and neck cancers: A SEER-Medicare analysis. Cancer
120(11): 1754
[21] Jones D (1994) ICRU Report 50 - Prescribing, Recording and Reporting Photon Beam
Therapy. Med Phys 21:833-834. doi:http://dx.doi.org/10.1118/1.597396
[22] Digital Imaging and Communications in Medicine (DICOM) Supplement 11 – Radiotherapy
Objects
[23] Moore KL, Brame RS, Low DA, Mutic S (2011) Experience based quality control of intensity-
modulated radiotherapy planning. Int J Radiat Oncol Biol Phys 81(2):545-51
[24] Yuan L, Ge Y, Yu WR, Yin FF, Kirkpatrick JP, Wu QJ (2012) Quantitative Analysis of the
factors which affect the interpatient organ-at-risk dose sparing variation in IMRT plans. Med
Phys 39(11):6868-78
[25] Appenzoller LM, Michalski JM, Thorstad WL, Mutic S, Moore KL (2012) Predicting dose-
volume histograms for organs-at-risk in IMRT planning. Med Phys 39(12):7446-61
[26] Schreibmann E and Fox T (2014) Prior knowledge treatment planning for volumetric arc
therapy using feature-based database mining. J Appl Clin Med Phys 15(2):4596
[27] Kazhdan M, Simari P, McNutt T, et al (2009) A Shape Relationship Descriptor Radiation
Therapy Treatment Planning. Med Image Comput Comput Assist Interv 12(2):100-8
88
[28] Wu B, Ricchetti F, Sanguineti G, Kazhdan M, Simari P, et al (2010) Data Driven Approach to
Generating Achievable Dose-Volume Histogram Objectives in Intensity Modulated Radiotherapy
Planning. Int J Radiation Oncology Biol Phys 79(4): 1241-1247

[29] Simari P, Wu B, Jacques R, King A, McNutt T (2010) A Statistical Approach for Achievable
Dose Querying in IMRT Planning. MICCAI 6363: 521-528.

[30] Wu B, Ricchetti F, Sanguineti G, et al (2009) Patient geometry-driven information retrieval
for IMRT treatment plan quality control. Med Phys 36(12):5497-505
[31] Rubner Y, Tomasi C, Guibas LJ (2000) The Earth Mover’s Distance as a Metric for Image
Retrieval. International J Comput Vis 40(2):99-121

89
Author Publications and Presentations

Peer-Reviewed Manuscripts

• Deshpande R., DeMarco J, Liu BJ, “Knowledge Driven Decision Support for Assessing Dose
Distributions in Radiation Therapy of Head and Neck Cancer”, Int J Comput Assit Radiol
Surg (Status: In Revision)
• Huang H. K., Deshpande R., Documet J., Le A. H., Lee J., Ma K., Liu B. J., “Medical imaging
informatics simulators: a tutorial”, Int J Comput Assist Radiol Surg. 2013 Sep 14. [Epub
ahead of print]
• Moin P., Deshpande R., Sayre J., Messer E., Gupte S., Romsdahl H., Hasegawa A., Liu B. J.,
“An observer study for a computer-aided reading protocol (CARP) in the screening
environment for digital mammography”, Academic Radiology 2011 Nov; 18(11):1420-9

Conference Proceedings

• Deshpande R., DeMarco J., Liu B. J., “Design and Evaluation of an Imaging Informatics
System for analytics-based decision support in radiation therapy”, Proceedings of SPIE Vol.
9418 94180Y (2015)
• Deshpande R., Thupimdang W., DeMarco J., Liu B. J., “A Collaborative Framework for
Contributing DICOM RT PHI (Protected Health Information) to augment data mining in
clinical decision support”, Proceedings of SPIE Vol. 9039, 90390K (2014)
• Meier J., Deshpande R., Liu B. J., Neumuth T., “A concept of a generalized electronic
patient record for personalized medicine”, Proceedings of SPIE Vol. 9039, 903907 (2014)
• Deshpande R., Zhou, A., Zhang, J., et al., "Role of an imaging informatics-based DICOM-RT
cancer registry in evaluating treatment parameters of IMRT for prostate cancer",
Proceedings of SPIE Vol. 8674, 86740O (2013)
• Fernandez J., Deshpande R., Hovanessian-Larsen L., et al., "Comparison of 2D versus 3D
mammography with screening cases: an observer study", Proceedings of SPIE Vol. 8318,
831818 (2012)
• Deshpande R., DeMarco J., Low D., et al., "Data mining DICOM RT objects for quality
control in radiation oncology", Proceedings of SPIE Vol. 8319, 83190Q (2012)
• Deshpande R., Requejo P., Sutisna E., et al., "A web-based electronic patient record (ePR)
system for data integration in movement analysis research on wheel-chair users to
minimize shoulder pain", Proceedings of SPIE Vol. 8319, 83190W (2012)
• Wang X., Fernandez J., Deshpande R., et al., "A computer-aided detection (CAD) system
with a 3D algorithm for small acute intracranial hemorrhage", Proceedings of SPIE Vol.
8319, 831913 (2012)

90
• Deshpande R., Li H., Requejo P., et al., "Utilization of DICOM multi-frame objects for
integrating kinetic and kinematic data with raw videos in movement analysis of wheel-
chair users to minimize shoulder pain", Proceedings of SPIE Vol. 8319, 83190S (2012)
• Fernandez J., Deshpande R., Wang X., et al., "Evaluation of a stand-alone computer-aided
detection system for acute intra-cranial hemorrhage in emergency environments",
Proceedings of SPIE Vol. 7967, 79670H (2011)
• Deshpande R., Fernandez J., Lee J. K., et al., "Multi-site evaluation of a computer aided
detection (CAD) algorithm for small acute intra-cranial hemorrhage and development of a
stand-alone CAD system ready for deployment in a clinical environment", Proceedings of
SPIE Vol. 7628, 76280H (2010)
• Le A., Deshpande R., Liu B. J., "Decision support tools for proton therapy ePR: intelligent
treatment planning navigator and radiation toxicity tool for evaluating of prostate cancer
treatment", Proceedings of SPIE Vol. 7628, 76280F (2010)

Conference Presentations

• Deshpande R., DeMarco J., Liu B., “Moving Towards Big Data Analytics in Radiation
Therapy - Dynamic Decision Support through Data Mining”, RSNA 2014 Educational Exhibit
(Computer Demonstration)
• Deshpande R., Le A., DeMarco J., Liu B., “Cloud computation of anatomical features from
imaging studies to discover radiation toxicity trends using a DICOM-based decision support
system”, RSNA 2013 Educational Exhibit (Computer Demonstration)
• Liu B., Deshpande R., DeMarco J., Documet J., “An open source, rich-client web
application for visualizing DICOM RT data”, RSNA 2013 (Computer Demonstration)
• Deshpande R., DeMarco J., Low D., “Data Mining Utilizing DICOM RT objects for
Knowledge Discovery in Radiation Therapy for Prostate Cancer”, RSNA 2011 (Paper Poster)
• Fernandez J., Documet J., Liu M., Deshpande R., et al., “Demonstration of a Stand-alone
Computer-aided Detection System for Acute Intra-cranial Hemorrhage in the Battle Field
of TBI (Traumatic Brain Injuries) in Emergency Environments”, RSNA 2010 Educational
Exhibit (Computer Demonstration)
• Le A., Ma K., Suh J., Park Y., Deshpande R., “A 3D Volumetric CAD of Multiple Sclerosis
Using Multi-core CPUs and General Purpose Graphics Processing Units (GPGPU)”, RSNA
2012 (Computer Demonstration)
• Moin P., Deshpande R., “Mind Over Matter: Using a neural headset to interact with a
PACS workstation”, RSNA 2009 (Computer Demonstration)

Abstract (if available)

Abstract Treatment planning in radiation therapy (RT) is a partly subjective process, relying considerably on the associated clinicians’ personal experience and judgment. This dissertation presents an informatics system that aims to assist RT treatment planners through quantified knowledge gained by analyzing retrospective treatment plans. Decision support is provided in the form of treatment parameter estimates that are likely to help reduce radiation dose to healthy tissue around the cancer tumor. In order to compute good recommendations, the system learns from the dose profiles of retrospective patients that are anatomically similar to a new patient. The pattern recognition features, metrics and techniques employed for this purpose have been described in this manuscript, along with the results of evaluating their effectiveness. Informatics applications such as this are often incomplete without a scheme for assimilating the developed programs into the clinical environment. As a result, this work also focuses on clinical workflow analysis to help identify weaknesses in the operational workflow, and corresponding informatics tools and techniques that can provide a resolution through effective data management and integration. The clinical workflow, data standardization through the adoption of DICOM, the system architecture and specifications of major components are discussed. Since the knowledge base forms the crux of the system, the manuscript highlights mechanisms to facilitate data collection and promote data diversity by enabling multi-institutional collaborations. Four major innovations presented herein include prototypes of: the Spatial Target Signature (STS), which is a novel feature for assessing anatomical similarity

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Mining an ePR system using a treatment plan navigator for radiation toxicity to evaluate proton therapy treatment protocol for prostate cancer

PDF

Decision support system in radiation therapy treatment planning

PDF

A medical imaging informatics based human performance analytics system

PDF

Development of an integrated biomechanics informatics system (IBIS) with knowledge discovery and decision support tools based on imaging informatics methodology

PDF

An electronic patient record (ePR) system for image-assisted minimally invasive spinal surgery

PDF

Imaging informatics-based electronic patient record and analysis system for multiple sclerosis research, treatment, and disease tracking

PDF

Identifying injury risk, improving performance, and facilitating learning using an integrated biomechanics informatics system (IBIS)

PDF

Molecular imaging data grid (MIDG) for multi-site small animal imaging research based on OGSA and IHE XDS-i

PDF

Lower extremity control and dynamics during the golf swing

PDF

Fresnel beamforming for low-cost, portable ultrasound systems

PDF

Upper extremity control and dynamics during manual wheelchair propulsion at different speeds and wheelchair configurations

PDF

Refined screening guidelines for radiation-related late effects in childhood cancer survivors: a dose-volumetric guided approach

PDF

Understanding reactive balance control strategies in non-disabled and post-stroke gait

PDF

A wireless implantable MEMS micropump system for site-specific anti-cancer drug delivery

PDF

Sense and sensibility: statistical techniques for human energy expenditure estimation using kinematic sensors

PDF

Long-term effects on vertebral body height growth of dose sculpting intensity modulated radiation therapy for children with neuroblastoma

PDF

Investigation of preclinical testing methods for total ankle replacements

PDF

Parametric and non‐parametric modeling of autonomous physiologic systems: applications and multi‐scale modeling of sepsis

PDF

Control and dynamics of turning tasks with different rotation and translation requirements

PDF

Towards a high resolution retinal implant

Asset Metadata

Creator Deshpande, Ruchi R. (author)

Core Title Knowledge‐driven decision support for assessing radiation therapy dose constraints

School Viterbi School of Engineering

Degree Doctor of Philosophy

Degree Program Biomedical Engineering

Publication Date 02/23/2016

Defense Date 01/19/2016

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag biomedical imaging,clinical decision support,DICOM RT,medical imaging informatics,multi-institutional data sharing,OAI-PMH Harvest,radiation therapy

Format application/pdf (imt)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Liu, Brent J. (committee chair), Chung, Kirk (committee member), D'Argenio, David (committee member), McNitt-Gray, Jill (committee member)

Creator Email rrdeshpa@usc.edu,ruchi.deshpande@gmail.com

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c40-213812

Unique identifier UC11278916

Identifier etd-DeshpandeR-4148.pdf (filename),usctheses-c40-213812 (legacy record id)

Legacy Identifier etd-DeshpandeR-4148.pdf

Dmrecord 213812

Document Type Dissertation

Format application/pdf (imt)

Rights Deshpande, Ruchi R.

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA