University of Southern California Dissertations and Theses

Decision support system in radiation therapy treatment planning

(USC Thesis Other)

Decision support system in radiation therapy treatment planning

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content Copyright 2024 Wejdan Ali Alshehri
Decision Support System in Radiation Therapy Treatment Planning
by
Wejdan Ali Alshehri
A Thesis Presented to the
FACULTY OF THE USC VITERBI SCHOOL OF ENGINEERING
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
MASTER OF SCIENCE
(BIOMEDICAL ENGINEERING: MEDICAL IMAGING & TELEMEDICINE)
August 2024

Copyright 2024 Wejdan Ali Alshehri ii
Acknowledgments
I would like to express my deepest appreciation to my principal investigator, Dr. Brent Liu, for his
invaluable guidance, mentorship, and support throughout this journey. His insights and expertise
have been instrumental in shaping my research.
I am immensely grateful for the unwavering support and encouragement from my family and
friends. Their love and understanding have been invaluable, and I truly appreciate having them in
my life. Additionally, my sincere thanks go to Trent Benedick for his contributions and assistance
throughout this process.I extend my gratitude to everyone at the Image Processing and Informatics
Lab. Their collaboration has enriched my experience and expanded my knowledge.
I sincerely thank Jazan University, Saudi Arabia, for their sponsorship of my educational journey.

Copyright 2024 Wejdan Ali Alshehri iii
TABLE OF CONTENTS
Acknowledgements....................................................................................................................................... ii
List of tables..................................................................................................................................................iv
List of figures................................................................................................................................................iv
Abstract......................................................................................................................................................... v
Chapter 1 Introduction.................................................................................................................................. 1
1.1Medical Imaging Informatics for Decision Support in Healthcare............................................. 2
1.2 Imaging Informatics in Radiation Therapy................................................................................ 3
1.3 Research Aims and Scope .........................................................................................................4
Chapter 2 Literature Review......................................................................................................................... 5
Chapter 3 Methodology................................................................................................................................. 8
3.1Decision Support System Architecture....................................................................................... 8
3.2 Data Collection..........................................................................................................................11
3.2.1 Data Acquisition .......................................................................................................11
3.2.2 Database Construction ..............................................................................................12
3.3Similarity Matching Algorithm..................................................................................................14
3.3.1 Gower's Dissimilarity Equation.................................................................................17
3.3.2 EMD and Implementation .........................................................................................18
Chapter4 Discussion.....................................................................................................................................22
4.1 Ongoing Work..............................................................................................................22
4.2 Testing and Validation.................................................................................................22
4.3 Future Work.................................................................................................................24
Chapter 5 Conclusion...................................................................................................................................25
References:.................................................................................................................................................. 26

Copyright 2024 Wejdan Ali Alshehri iv
List of Figures
Figure 1. Decision Support System Components...................................................................................8
Figure 2. Entity Relationship Diagram Illustrating the Data Structure fin MongoDB.......................13
Figure 3. Planning target volume PTV & Organ at risk OAR contours...............................................14

Copyright 2024 Wejdan Ali Alshehri v
Abstract
The radiation therapy treatment planning involves complex decision-making processes that rely
heavily on the expertise of clinicians. This research introduces an innovative decision support
system (DSS) designed to enhance personalized treatment recommendations in radiation therapy
by leveraging advanced informatics methodologies. This research introduces an innovative
decision support system for personalized radiation therapy treatment recommendations using
advanced informatics methodologies. To compare patient profiles based on anatomical
characteristics, it employs the integration of Gower's similarity measure and Earth Mover's
Distance (EMD) analysis, enabling a nuanced assessment of dissimilarities.
The DSS computes EMD values for spatial target signatures (STS) and overlap volume histograms
(OVH) across multiple regions of interest (ROIs). The effectiveness of the EMD metric is in
quantifying dissimilarities between radiation therapy plans is validated through a testing
methodology. This involves randomly shifting the contours of organs at risk (OARs) and target
volumes while maintaining their original shape and size, allowing us to evaluate the EMD metric's
sensitivity to anatomical location variability, and its accuracy in identifying similar treatment
plans.
The implementation of this DSS involves the acquisition of diverse patient datasets from multiple
institutions, ensuring the generalizability of the research outcomes. The data is efficiently
organized and managed using MongoDB, a NoSQL database solution that allows for the storage
of complex and varied data types.The successful integration of these advanced informatics
methodologies paves the way for more personalized, precise, and effective radiation therapy
treatment planning.

Copyright 2024 Wejdan Ali Alshehri 1
Chapter 1: Introduction
Treatment planning in radiation therapy(RT) involves a degree of subjectivity, often relying
heavily on the expertise and judgment of clinicians. This research introduces an DSS designed to
support RT treatment planners using quantified insights gleaned from the analysis of past treatment
plans. The system offers decision support by providing estimates for treatment parameters aimed
at minimizing radiation exposure to healthy tissue surrounding the tumor.
The core functionality of the DSS lies in its ability to analyze and learn from historical treatment
planning data. Through a database of previously treated patients, the system identifies patterns and
correlations between patient characteristics. By applying advanced similarity matching algorithms,
such as Gower's distance, the system compares a new patient's profile with historical cases to
identify the most similar patients through computing Earth Mover's Distance (EMD) across STS
and OVH within a Gower's similarity framework to get the similarity match.
The integration of Gower's similarity measure and EMD analysis enables the system to capture the
multidimensional nature of patient data, considering factors such as anatomical structures,
dosimetric parameters, and spatial relationships. By quantifying the dissimilarity between patients,
the system can identify clusters of patients with similar characteristics and recommend treatment
plans.
Furthermore, this research emphasizes the importance of Machine learning ML as new patient data
and treatment outcomes become available, the decision support system can adapt and refine its
recommendations, leveraging the power of ML techniques. The system's architecture is designed

Copyright 2024 Wejdan Ali Alshehri 2
to facilitate the incorporation of advanced algorithms and models, enabling it to evolve and provide
accurate and personalized treatment suggestions over time.
This thesis details the pattern recognition features, metrics, and techniques utilized for this
purpose, along with an evaluation of their efficacy. In the following chapters, we will delve into
the detailed methodology, system architecture, and evaluation of the decision support system. We
will explore the application of Gower's similarity measure and EMD analysis in the context of
radiation therapy, and discuss the development and implementation of the system, and validation
methods. Through this research, we seek to make a significant contribution to the field of medical
informatics and advance the state of personalized radiation therapy treatment planning.
1.1 Medical Imaging Informatics for Decision Support in Healthcare
Advanced medical imaging techniques like CT and MRI have led to a significant increase in
electronic health data production. This has created a need for effective solutions to store, manage
and distribute digital healthcare data. The field of medical imaging informatics emerged to address
this demand, giving rise to industry standards, information systems and protocols that form the
foundation of today's healthcare IT industry (Deshpande, 2016).
Radiology, being heavily dependent on digital health data, provided the initial foothold for medical
imaging informatics. Early research led to the development of standards like DICOM (Digital
Imaging and Communications in Medicine) and systems like PACS (Picture Archiving and
Communications Systems). DICOM standardized storage and communication of medical images,
while PACS streamlined clinical workflow in radiology[4].

Copyright 2024 Wejdan Ali Alshehri 3
With the widespread adoption of these systems, a new avenue of imaging informatics research
involves leveraging the vast amounts of data routinely stored in these systems. One approach is to
search for patterns and develop mathematical models that can assist clinicians in their decisionmaking process, ultimately aiming to improve patient outcomes.. However, challenges remain
regarding model explainability, scarcity of annotated data for model training, and effective
translation of these automated techniques into routine clinical practice.
1.2 Imaging Informatics in Radiation Therapy
Radiation therapy is one among the medical specialties poised to benefit significantly from the
integration of imaging informatics. This is attributed to its growing dependence on medical
imaging for creating accurate three-dimensional representations of patients' internal anatomy.
These representations are significant for determining tumor characteristics such as size, location,
spread, and spatial distribution relative to surrounding anatomical structures. Such information is
essential for optimizing treatment parameters to deliver a uniform radiation dose to the tumor while
minimizing exposure to healthy surrounding tissues. While a high target dose is necessary for
tumor eradication, excessive radiation exposure to normal tissue can lead to harmful side effects
and compromise patients' post-treatment quality of life. Therefore, precise knowledge of each
patient's individual anatomy provided by medical imaging is fundamental to formulating effective
treatment strategies.
Radiation therapy data encompasses more than just images, including radiation dose distributions,
region of interest contours, histograms, and beam intensities. Although standards exist for these
data objects, most vendors still employ proprietary formats for storage and communication,
hindering the development of universal imaging informatics applications capable of interfacing
with products from multiple vendors. Consequently, many research groups opt to develop

Copyright 2024 Wejdan Ali Alshehri 4
applications tailored to the treatment systems used at their institutions, posing challenges for
integrating data across different countries, institutions, and treatment planning systems within the
same institution. Despite these infrastructure limitations, there is untapped potential in utilizing
pre-existing treatment planning data to power decision support systems that assist clinicians in
developing treatment plans. decision support systems in radiation therapy aim to assist clinicians
in planning treatments and, ideally, improving treatment outcomes.
1.3 Research Aims and Scope
This research aims to enhance the decision support system for radiation therapy treatment planning
by integrating advanced informatics methodologies and algorithms, with a primary focus on
improving the personalization of treatment recommendations through the comprehensive analysis
of patient data, specifically anatomical similarities, dosimetric variations, and outcome
predictions. The objectives encompass the integration of Gower's similarity measure and Earth
Mover's Distance (EMD) calculation to effectively capture dissimilarities between patient dose
profiles and identify anatomically similar cases. The novel integration of Earth Mover's Distance
(EMD) with STS and OVH within a Gower's similarity framework, as proposed in this thesis, aims
to address this limitation and enable more nuanced identification of similar patients based on
dosimetric and anatomic characteristics. This research further aims to investigate the composition
and characteristics of the EMD values stored within the system, optimizing their representation
and storage for efficient retrieval and analysis, and integrating the EMD analysis results into
propose validation and testing methodologies to ensure the similarity matching algorithm
reliability in identifying anatomical location variability.

Copyright 2024 Wejdan Ali Alshehri 5
Chapter 2: Literature Review
Personalized treatment planning is a critical aspect of radiation therapy, as it aims to optimize
treatment outcomes while minimizing adverse effects on healthy tissues. Decision support systems
have emerged as valuable tools to assist clinicians in developing individualized treatment plans
[5]. These systems leverage the analysis of past treatment plans and patient data to provide
recommendations for treatment parameters that minimize radiation exposure to healthy tissue
surrounding the tumor.
To generate effective personalized treatment recommendations, it is essential to consider the
anatomical and dosimetric similarities between patients. Previous studies have explored various
methods to quantify these similarities, such as using machine learning techniques like clustering
and regression [5]. Spatial Target Signatures (STS) and Overlap Volume Histograms (OVH) have
been employed to characterize spatial relationships and organ-at-risk exposure, enabling the
analysis of dosimetric variations and anatomical similarities.
Alfonso et al. [1] introduced a decision-support system called the Dose Distribution Index (DDI)
to assist in comparing and evaluating radiotherapy treatment plans. The DDI takes into account
key dosimetric factors such as dose coverage, conformity, and homogeneity for the planning target
volume (PTV), as well as sparing of organs at risk (OARs) and the remaining volume at risk
(RVR). The system utilizes dose-volume histogram (DVH) data provided by treatment planning
systems to calculate separate indices for the PTV, OARs, and RVR, which are then combined into
a single DDI score. The authors' retrospective evaluation of the DDI for fifty brain cancer patients
demonstrated its potential as a straightforward, non-subjective tool for quantitative comparison of

Copyright 2024 Wejdan Ali Alshehri 6
treatment plans, aiding in the decision-making process and potentially improving patient
outcomes[1].
Momin et al,2021 provided a comprehensive review of knowledge-based planning (KBP) methods
for dose prediction in radiation therapy. The authors categorized KBP approaches into traditional
methods using features like overlap volume histogram and distance-to-target histogram, and more
recent deep learning (DL) based methods that automatically learn features from raw data. They
noted that a common limitation of traditional KBP methods is the considerable effort required to
design handcrafted features that capture anatomical information. In contrast, DL methods can
automatically learn relevant features. The novel framework proposed in this thesis seeks to bridge
this gap by integrating Earth Mover's Distance (EMD) with shape-to-shape (STS) distance and
OVH within a Gower's similarity measure. This approach aims to leverage the strengths of both
traditional and modern techniques, enabling more sophisticated identification of similar patients
based on a comprehensive set of dosimetric and anatomic characteristics.
Recent research has highlighted the importance of organ-specific analysis in pediatric radiation
therapy planning. Huijskens et al. investigated interfractional position variations of abdominal
organs in children undergoing image-guided radiotherapy. Their findings revealed weak
correlations between position variations of contralateral organs and moderate correlations between
diaphragm and upper abdominal organ movements. While differences in systematic and random
errors between organs were small, the study emphasized the need for anisotropic margins and
careful consideration of organ-specific motion patterns when defining planning target volumes and
organs at risk. This work underscores the value of comprehensive organ motion analysis in
improving the accuracy and personalization of radiotherapy treatment plans for pediatric patients.

Copyright 2024 Wejdan Ali Alshehri 7
This study's findings on organ-specific motion and spatial relationships align with our research
focus on enhancing personalized radiation therapy treatment planning through advanced similarity
matching algorithms that consider individual patient anatomy and organ variability.
In summary, the DDI proposed by Alfonso et al. [1] and the KBP methods reviewed by Momin et
al. [7] provide valuable insights into the development of decision-support systems for radiotherapy
treatment planning. The approach proposed in this thesis builds upon these concepts by integrating
advanced similarity measures, such as EMD, STS, and OVH, within Gower's similarity
framework. This integration aims to address the limitations of traditional KBP methods and enable
more nuanced identification of similar patients based on dosimetric and anatomic characteristics.
Moreover, the importance of organ-specific analysis, as highlighted by Huijskens et al. [6], is
considered in the proposed approach, facilitating the personalization of treatment plans based on
individual patient anatomy. By leveraging these advancements, the proposed decision support
system has the potential to significantly enhance anatomical similarity matching to get
personalized and effective radiation therapy treatment planning.

Copyright 2024 Wejdan Ali Alshehri 8
Chapter 3: Methodology
This chapter outlines System Architecture and the methods and procedures used in developing and
implementing the decision support system for radiation therapy treatment planning. The aim is to
enhance treatment planning through the use of advanced imaging informatics techniques,
specifically focusing on the integration of the Earth Mover's Distance (EMD) within similarity
matching algorithms is utilized to measure the dissimilarities in STS and OVH, which represent
the spatial and volumetric discrepancies between planned and delivered radiation doses,
respectively. This assessment helps in optimizing and personalizing radiation therapy treatment
plans based on precise anatomical data and radiation dose delivery accuracy."
3.1 Decision Support System Architecture
The DSS for radiation therapy treatment planning consists of several interconnected components
designed to optimize the planning process. The Graphical User Interface (GUI), which serves as
the primary interface between the user and the system. It provides a visual and interactive means
for users to interact with the software, input data, and receive outputs.
Figure1: Decision Support System Components

Copyright 2024 Wejdan Ali Alshehri 9
The GUI allows clinicians to upload patient-specific CT images and contours, providing
a user-friendly interface to input necessary data for treatment planning. Once the patient data is
uploaded through the GUI, the DICOM Image/Metadata Parser queries the data. Specifically, this
component extracts relevant metadata from the DICOM files, ensuring that the necessary
information is captured for further processing. The parser runs on the DSS 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 (Deshpande et al., 2016).
The extracted and anonymized metadata is then stored in a database, which serves as a central for
all the relevant information. The database is organized to correspond to different DSS components,
facilitating efficient data retrieval and processing. This structured approach enables the system to
effectively manage and utilize the collected data.
Following storage of the metadata in the database, the Feature Extraction component is activated,
as illustrated in Figure 1 (component 3). This component processes the raw metadata to extract
key features that are essential for treatment planning. Two important features extracted by this
component are the Spatial Target Signature (STS) and the Overlap Volume Histogram (OVH).
These features provide valuable information about the spatial relationships and volumetric overlap
between the target volume and surrounding organs. The extracted features are then utilized by the
Decision Support Manager Module, which forms the core of the decision support system. This
module integrates with the database to retrieve the necessary data and leverages the outputs of the
Feature Extraction component. It plays a crucial role in analyzing the patient data and generating
personalized treatment recommendations.
To determine the similarity between patients, the system employs a Similarity Calculation
component that utilizes Earth Mover's Distance (EMD) as a metric. The EMD calculation

Copyright 2024 Wejdan Ali Alshehri 10
quantifies the dissimilarity between patients based on their extracted features, enabling the system
to identify anatomically similar cases.
The Data Mining component further enhances the decision support capabilities by performing
similarity matching. It compares the data of a new patient against historical data stored in the
database, identifying prior successful cases that closely resemble the current patient's
characteristics. This similarity matching process allows the system to draw insights from past
experiences and adapt treatment recommendations accordingly.
The clinical workflow for radiation therapy treatment planning seamlessly integrates the decision
support system. When a patient opts for radiation therapy, their information is entered into the
oncology system, and a CT exam is scheduled. The acquired CT study is then stored in the
database, and the images are retrieved and imported into the Treatment Planning System (TPS).
At this stage, the front-end Radiation Therapy Decision Support System, powered by the GUI,
assists clinicians in decision-making. It integrates with the TPS to provide personalized treatment
recommendations based on the analysis of historical patient data. The GUI presents relevant
information, such as Intensity Modulated Radiation Therapy (IMRT) dose constraints and dosevolume histograms, enabling clinicians to make informed decisions.
The back-end servers handle the core processing and analysis tasks of the decision support system.
These servers execute the algorithms that generate personalized treatment recommendations based
on patient data and prior successful cases. They interact with the database system to retrieve and
store relevant data and calculate essential metrics, such as the OVH and STS.
In summary, the decision support system architecture combines various components, including the
GUI, DICOM Image/Metadata Parser, database, Feature Extraction, Decision Support Manager
Module, Similarity Calculation, and Data Mining. These components work together to process

Copyright 2024 Wejdan Ali Alshehri 11
patient data, extract relevant features, identify similar cases, and generate personalized treatment
recommendations. The system seamlessly integrates into the clinical workflow, assisting clinicians
in making informed decisions and optimizing radiation therapy treatment planning. In the
following sections, we will delve deeper into the data collection process, exploring the acquisition
of diverse patient datasets, decision support metrics STS and OVH, and the construction of the
MongoDB database
3.2 Data Collection
3.2.1 Data Acquisition
In collaboration with Roswell Park Cancer Institute, the research collected more than 250
anonymized cases in the thesis that are sourced from three different datasets to ensure a diverse
and comprehensive pool of data for analysis, this approach supports the development and
validation of the decision support system across varied clinical environments, enhancing the
generalizability of the research outcomes. The datasets, encompassing CT scans saved in DICOM
format of patients with head and neck cancers.
Table 1: Dataset sources and number of cases
Instituition Cases
University of California, Los Angeles (UCLA) 137
State University of New York (SUNY) at Buffalo 70
Technical University of Munich (TUM) 50

Copyright 2024 Wejdan Ali Alshehri 12
All data were handled with strict adherence to privacy and ethical standards, including encryption
during transfer and restricted access to authorized personnel, ensuring compliance with HIPAA
and other data protection regulations.
In this system, each patient's data is intricately linked to multiple radiation therapy (RT) treatments,
encapsulating various studies, each potentially comprising multiple CT scans and DICOM RT
series. These are meticulously aligned within the same patient coordinate system, using patient
position and orientation attributes, to create a unified three-dimensional model. This model
supports the computation of key decision support metrics like the Overlap Volume Histogram
(OVH) and Spatial Target Signature (STS), crucial for the system's functionality. These features,
derived from detailed analysis and computational methods rather than directly from DICOM
objects, facilitate the generation of personalized treatment plans by highlighting similarities among
patient treatments, as will be explored further in subsequent chapters.
3.2.2 Database Construction
The system architecture utilizes MongoDB to handle all data. MongoDB is a powerful, flexible,
and scalable NoSQL database solution that uses a document-oriented approach to store and
manage data. As a NoSQL database, MongoDB does not require a fixed schema, allowing it to
handle diverse data types and adapt to changing data structures without modifying existing data.
This makes it an excellent choice for managing large volumes of structured, semi-structured, and
unstructured data, which is particularly relevant in the context of medical imaging informatics.
The data in MongoDB is organized into collections, which are analogous to tables in traditional
relational databases. However, unlike tables, collections in MongoDB do not enforce a rigid
structure, enabling us to store complex and varied data types within the same collection. This

Copyright 2024 Wejdan Ali Alshehri 13
flexibility is beneficial during the implementation of similarity-matching algorithm, where patient
data may contain different types of information.
The MongoDB database efficiently organizes patient data, treatment plans, imaging data, and
calculated metrics. The Patient collection serves as the central hub, storing essential patient
information and connecting to the Studies collection via the PatientName field. The Studies
collection further links to the Series collection, which contains detailed information about each
study series. The ROI and RTContour collections store derived regions of interest and contours,
respectively. The STS and OVH collections contain calculated metrics used as inputs for the EMD
calculations, which measure the dissimilarity between patient cases. The EMD collection stores
Figure 2: Entity Relationship Diagram Illustrating the Data Structure fin MongoDB

Copyright 2024 Wejdan Ali Alshehri 14
the calculated EMD values, enabling efficient retrieval and comparison of patient similarities. This
structured database design allows for the seamless integration of patient data, imaging files, and
calculated metrics, facilitating the development
We will examine closely the similarity matching algorithm in the following section, exploring the
significance of Gower's distance equation and the Earth Mover's Distance in capturing the
complexities and variabilities inherent in radiation therapy patient profiles. Leveraging these
powerful techniques, the algorithm enables the identification of similar patients, establishing the
basis for personalized treatment planning and improved clinical outcomes.
3.3 Similarity matching algorithm
Gower's distance equation provides a unified framework for comparing diverse variables by
normalizing their contributions to the overall dissimilarity measure.
The similarity matching algorithm, powered by Gower's distance and EMD calculations, plays a
pivotal role in identifying anatomically similar patients and guiding personalized treatment
planning. By interfacing with the database system, the algorithm retrieves relevant patient data,
including ROIs, OVHs, and STSs, which serve as inputs for the similarity assessment process.
Regions of Interest ROIs are anatomical structures
contoured on medical images, such as the planning
target volume (PTV) and organs at risk (OARs). These
structures are the basis for deriving OVHs and STSs.
Overlap Volume Histograms OVHs quantify the
volumetric overlap and distance between an OAR and
the PTV. They represent the percentage of OAR volume
Figure 3:
Planning Target Volume (PTV)
Right Parotid
(a) Planning target volume (PTV, in red) is situated very close
to the right parotid (solid green) and much farther from the left
parotid(green outline).(b) The PTV is situated midway between
the two parotids

Copyright 2024 Wejdan Ali Alshehri 15
within a certain distance from the PTV surface. OVHs capture the spatial relationship between the
PTV and OARs, which influences the achievable dose distribution.
The main objective of radiation therapy is to deliver the highest possible radiation dose to the
tumor while minimizing the exposure to surrounding healthy tissues. However, this poses a
challenge due to the proximity of tumors to critical organs. Figure 3 illustrates this difficulty, with
part (a) depicting a scenario where the right parotid gland (shown in solid green) is at a high risk
of radiation exposure because of its close vicinity to the planning target volume (PTV) (shown in
red). In contrast, part (b) presents a case where the right parotid gland has a lower risk of
overexposure, as it is situated at a greater distance from the PTV, making it easier to spare during
treatment.
Moreover, STSs characterize the three-dimensional spatial distribution of the PTV relative to an
OAR. They divide the space around the OAR into virtual 3D grid cells and quantify the percentage
of PTV volume in each cell. STSs provide a comprehensive representation of the PTV shape, size,
and location with respect to the OARs.
OVHs and STSs are derived from ROI contours and serve as similarity features in the DSS. They
comprehensively capture the geometric relationships between the PTV and OARs, which influence
the achievable dose distribution. The DSS in radiation therapy computes OVHs and STSs for a
new patient and compares them with a database of retrospective cases to identify anatomically
similar patients.
The process begins by computing the OVHs and STSs for a new patient. These computations are
carried out by extracting and analyzing the spatial data from the patient's CT images and ROI
contours.

Copyright 2024 Wejdan Ali Alshehri 16
Initially, the OVHs are calculated by measuring the volumetric overlap between the PTV and each
OAR, quantifying how much of an organ is likely to receive radiation. Secondly, STSs are
generated by determining the spatial distribution of the PTV in relation to the surrounding OARs,
providing a detailed map of relative positions and distances.
These newly computed OVHs and STSs are then matched against a database of retrospective
patient cases that also contain these similarity features. The DSS employs similarity matching
algorithms, potentially incorporating Earth Mover's Distance (EMD), to compare the OVHs and
STSs of the new patient with those in the database. This comparison helps identify cases with
anatomical similartiy to the new patient, thereby facilitating the derivation of personalized
treatment plans that are informed by previously successful outcomes.
Gower's distance equation is a crucial component of the similarity-matching algorithm due to its
ability to handle variables of different dimensions and data types. In the context of radiation
therapy treatment planning, patient profiles often consist of a mix of variables.
The importance of Gower's distance lies in its flexibility and adaptability to the complex nature of
patient data. By treating each variable independently and scaling their dissimilarities accordingly,
Gower's equation ensures that the similarity assessment accounts for the heterogeneous
characteristics of patient profiles. This approach enables the algorithm to capture subtle differences
and similarities between patients, facilitating accurate matching and treatment recommendations.
In conjunction with Gower's distance, the similarity-matching algorithm employs Earth Mover's
Distance (EMD) as a robust metric for quantifying the dissimilarity between patient profiles. EMD
measures the minimum cost of transforming one distribution into another, taking into account both
the magnitude of the differences and the spatial relationships between the compared distributions.

Copyright 2024 Wejdan Ali Alshehri 17
The strength of EMD lies in its ability to handle variable-length representations and allow for
partial matching. In the context of radiation therapy, patient profiles may have varying numbers of
ROIs or dosimetric features. EMD accommodates these variations by considering the optimal
transport of mass between the distributions, ensuring that the similarity assessment remains
meaningful even when the compared profiles have different lengths.
The combination of Gower's distance equation and EMD forms a powerful similarity-matching
algorithm that effectively addresses the complexities and variabilities inherent in radiation therapy
patient profiles. By leveraging Gower's ability to handle diverse data types and EMD's robustness
to partial matching and quantization, the algorithm enables the identification of similar patients.
3.3.1 Gower's Dissimilarity Equation
Gower's distance calculates the dissimilarity between two cases by considering each variable
separately and normalizing their contributions to the overall dissimilarity. For numeric variables,
the absolute difference is normalized by the range of the variable. For categorical variables, a
simple matching coefficient is used, where a value of 0 is assigned for a match and 1 for a
mismatch. Binary variables are treated as a special case of categorical variables.
By combining these individual variable dissimilarities, Gower's distance provides a
comprehensive measure of the overall dissimilarity between two patient cases. The normalization
step ensures that each variable contributes equally to the final distance, regardless of its scale or
type. This is particularly important when dealing with variables that have different units or
magnitudes.

Copyright 2024 Wejdan Ali Alshehri 18
Gower's distance is useful for clustering or classification tasks involving mixed-type data, as it can
handle different variable types simultaneously and provide a unified dissimilarity measure. in
belew sections, we will utilize Gowers distance.
1. DGower (Xi,Xj) =∑ ��
where:
DGower (Xi, Xj): represents the Gower's distance between objects i and j objects
dijc: is the dissimilarity measure between the i-th and j-th objects for the c-th variable
v: is the total number of variables that refer to OARs in this research
If the c-th variable is nominal or alternative, the dissimilarity between two categories xic and xjc
is treated as zero for matches of categories, and as one otherwise.
The significance of Gower’s equation is it can handle different types of data and bring them to a
common scale, by treating each variable independently and normalizing their contributions to the
overall dissimilarity, Gower's distance provides a unified scale for comparing objects with diverse
attributes. This is particularly important when dealing with patient profiles that may include a
combination of numerical measurements.
3.3.2 EMD and Implementation
The Earth Mover's Distance EMD is a metric that quantifies the dissimilarity between two
distributions or signatures. EMD is used to calculate the dissimilarity between individual metrics,
such as OVH and STS.
v
c=1

Copyright 2024 Wejdan Ali Alshehri 19
EMD treats these individual metrics (OVH and STS) as distributions or signatures and calculates
the minimum cost of transforming one distribution into another. The cost is determined by the
amount of work required to move the mass from one distribution to another, considering both the
distance between the elements and the amount of mass being moved.
In the case of OVH and STS, EMD calculates the dissimilarity by considering the differences in
the volumetric overlap and spatial distribution between two patient cases. It considers not only the
magnitude of the differences but also the spatial relationships between the elements of the
distributions.
EMD itself serves as a metric of similarity by combining the dissimilarities calculated for
individual metrics (OVH and STS). By aggregating these individual dissimilarities, EMD provides
an overall measure of similarity between patient cases, taking into account multiple aspects of their
anatomical and dosimetric characteristics.
Gower's similarity measure was integrated with EMD calculations to effectively capture
dissimilarities between patient profiles with different OARs types, enabling the identification of
anatomically similar patients.
(2) �� = 1 − 7�!"#$%&% ()*+,-
!"#!"! $%&'()
"+.!"#!"! $%&'()
+
�!"#$(/0 ()*+,-
!"#$*+ $%&'()
"+.!"#$*+ $%&'()
+ �!"#$%&% ()*+,1
!"#!"! $%&'(,
"+.!"#!"! $%&'(,
+
�!"#$(/0 ()*+,1
!"#$*+ $%&'(,
"+.!"#$*+ $%&'(,
+ ⋯ ;

Copyright 2024 Wejdan Ali Alshehri 20
The similarity index is calculated as the Weighted Earth Mover's Distance (WEMD) for STS and
OVH calculated and normalizes these by their respective maximum EMD values across a dataset.
This normalization is significant as it scales the dissimilarity to a range where a comparison
between different patient cases becomes meaningful.
(3) The weighting factor in the equation �!"# = -
234567 89 :;< = 1
arises because for each OAR, there are two distinct types of EMD values considered; EMD for
organ volume histograms (OVH) and EMD for spatial transformation signatures (STS).
Subtracting this value from 1 converts the dissimilarity measure into a similarity score; where a
score of 1 indicates perfect similarity and a score of 0 indicates maximum dissimilarity. This
aggregation into an overall similarity score enables assessment of patient cases, thereby facilitating
personalized treatment planning that closely aligns with previously successful cases.
A Python implementation for Gower's equation works when comparing a new patient case to the
existing cases in the database using Gower's distance; we precompute the maximum EMD values
for each region of interest (ROI) for each OAR type in the new case. This approach ensures that
the EMD calculation is performed only once for the new patient case at the beginning, rather than
repeatedly for each subsequent Gower's distance application.
To determine the maximum EMD value for a specific ROI type (e.g., STS for Organ Type 1), we
retrieve all the EMD values for all ROIs in that OAR type1. By comparing these values, we identify
the maximum EMD value for each ROI in OAR type1.

Copyright 2024 Wejdan Ali Alshehri 21
This process is repeated for both STS and OVH features for all organ types present in the current
case. The maximum EMD values for each feature type and organ are then stored in a local object.
The EMD calculation itself is performed using the scipy.stats.wasserstein_distance_nd function
from the SciPy library in Python. This function computes the Wasserstein-1 distance between two
N-dimensional discrete distributions, which aligns with our requirement to compare STS and OVH
for different patient cases.
By precomputing and storing the maximum EMD values for each ROI type, we can efficiently
normalize the EMD values during the subsequent Gower's distance calculations when comparing
the current case with other cases in the database. The implementation of this optimized EMD
calculation approach significantly reduces the computational overhead associated with repeated
calculations, thereby improving the overall performance of the decision support system.

Copyright 2024 Wejdan Ali Alshehri 22
Chapter 4: Discussion
4.1 Ongoing Work
The integration of Gower's similarity measure and Earth Mover's Distance (EMD) analysis into
the decision support system for radiation therapy treatment planning has shown promising results
in enhancing the personalization of treatment recommendations. By effectively capturing the
dissimilarities between patient dose profiles, the system can identify anatomically and
dosimetrically similar patients, thereby facilitating the generation of similarity-matching treatment
plans. However, it is essential to acknowledge that the implementation process is still ongoing.
The current focus is on refining the codebase and optimizing the EMD calculation using the
scipy.stats.wasserstein_distance_nd function from the SciPy library in Python.
4.2 Validation and Testing
To validate the effectiveness of the Earth Mover's Distance (EMD) in quantifying the dissimilarity
between radiation therapy plans, we have developed a systematic testing methodology that
involves randomly shifting the contours of organs at risk (OARs) while maintaining the original
shape and size of the target volume. This approach allows us to assess the sensitivity and accuracy
of the EMD metric in identifying similar plans based on anatomical location variability.
The validation process focuses on evaluating the similarity matching algorithm's ability to
accurately identify and group similar cases based on anatomical variations. We have divided the
validation into four groups, each representing a different level of anatomical variation determined
by the magnitude of the contour shift applied to the OARs, groups in this example are:

Copyright 2024 Wejdan Ali Alshehri 23
1. Group 1: 10-pixel shift
2. Group 2: 20-pixel shift
3. Group 3: 30-pixel shift
4. Group 4: 60-pixel shift
For each group, we will perform a randomization process, generating 100 different location
variations of the contour shifts for each group. This randomization ensures that the algorithm is
tested against a diverse set of anatomical configurations within each group, covering a wide range
of possible scenarios and assessing the algorithm's performance under different conditions.
After generating the randomized variations for all four groups, we will have a total of 400 test
cases (100 cases per group). These test cases represent a comprehensive set of anatomical
variations with different levels of contour shifts.
The next step is to apply the similarity matching algorithm to the entire set of 400 test cases. The
algorithm will compare each test case with every other test case and determine their similarity
based on the implemented metrics and criteria.
Ideally, the similarity matching algorithm should be able to accurately group the test cases that
belong to the same group. Test cases with a 10-pixel shift should be matched with other test cases
that have a 10-pixel shift, as they represent the most similar anatomical variations. Similarly, test
cases with a 20-pixel shift should be grouped together, and so on.

Copyright 2024 Wejdan Ali Alshehri 24
The expected outcome is that the algorithm will successfully identify and cluster the test cases
based on their similarity in terms of the magnitude of the contour shift. Test cases within the same
group should have a higher similarity score compared to test cases from different groups.
By evaluating the algorithm's performance in this manner, we can assess its ability to handle
different levels of anatomical variations and its sensitivity to the magnitude of the contour shifts.
If the algorithm can accurately group the test cases based on their similarity, it indicates that it is
effective in capturing and quantifying the anatomical differences between cases.
Additionally, by mixing all the test cases together and applying the similarity matching algorithm,
we are evaluating the algorithm's robustness and ability to handle a heterogeneous dataset. The
algorithm should be able to distinguish between cases with different levels of contour shifts and
group them accordingly, even when they are presented in a mixed fashion.
4.3 Future Work Utilizing Generative AI for Evaluation
While working on this project we are looking beyond the current scope implementing-similarity
matching algorithms, a promising future direction is the application of generative AI techniques
in medical imaging. Generative AI, such as Generative Adversarial Networks(GANs), Would
enable the creation of a comprehensive and diverse knowledge base for the decision support
system, it adds importance to conducting evaluations and tests on a large number of different cases.
Allows us to generate additional test cases that cover a broader spectrum of shape, size, and
location variability. This expanded dataset can be used to further evaluate and refine the similarity
matching algorithm, which results in a greater accuracy in calibrating the success of the methods
used to enhance the diagnostic assistance system for radiation therapy treatment planning.

Copyright 2024 Wejdan Ali Alshehri 25
Chapter 5: Conclusion
This research aimed to enhance the decision support system for radiation therapy treatment
planning by integrating Gower's similarity measure using Earth Mover's Distance (EMD) analysis.
The primary objective was to improve the personalization of treatment recommendations by
effectively capturing the dissimilarities between patient dose profiles and identifying anatomically
similar distance cases.
Through the implementation of Gower's similarity measure and the incorporation of EMD
calculations, the system has demonstrated promising results in quantifying the dissimilarity
between patients' dose distributions, spatial target signature (STS), and Overlap Volume
Histograms (OVH). This approach enables a more comprehensive assessment of patient
similarities, facilitating the generation of tailored treatment plans that closely align with previously
successful cases across different organ types.
In conclusion, the integration of Gower's similarity measure and EMD analysis into the decision
support system represents a advancement in the field of medical imaging informatics, particularly
in radiation therapy. By focusing on assessing the sensitivity and accuracy of the EMD metric in
the context of anatomical location variability, this research contributes to the advancement of
personalized radiation therapy treatment planning. The proposed validation approach, involving
the systematic shifting of organ contours and the generation of simulated dataset, provides a robust
framework for evaluating the effectiveness of the EMD in quantifying plan similarity and
dissimilarity among mixture of different anatomical location. The results of this study will not only
validate the use of the EMD but also lay the foundation for further research and refinement of
similarity assessment techniques in the field of radiation therapy decision support systems.

Copyright 2024 Wejdan Ali Alshehri 26
References:
[1] Alfonso, L. J. C., Herrero, M. A., & Núñez, L. (2015). A dose-volume histogram based
decision-support system for dosimetric comparison of radiotherapy treatment plans. Radiation
Oncology (London, England), 10(1), 263–263.
[2] Akay, Ö., & Yüksel, G. (2018). Clustering the mixed panel dataset using Gower's distance and
k-prototypes algorithms. Communications in Statistics-Simulation and Computation, 47(10),
3031-3041.
[3] Bang, C., Bernard, G., Le, W. T., Lalonde, A., Kadoury, S., & Bahig, H. (2023). Artificial
intelligence to predict outcomes of head and neck radiotherapy. Clinical and Translational
Radiation Oncology, 39, 100590.
[4] Deshpande, R. R., DeMarco, J., Sayre, J. W., & Liu, B. J. (2016). Knowledge-driven decision
support for assessing dose distributions in radiation therapy of head and neck cancer. International
journal of computer assisted radiology and surgery, 11, 2071-2083.
[5] Foster, I., Spezi, E., & Wheeler, P. (2023). Evaluating the use of machine learning to predict
expert-driven pareto-navigated calibrations for personalised automated radiotherapy planning.
Applied Sciences, 13(7), 4548.

Copyright 2024 Wejdan Ali Alshehri 27
[6] Huijskens, S. C., van Dijk, I. W., Visser, J., Balgobind, B. V., te Lindert, D., Rasch, C. R., ...
& Bel, A. (2018). Abdominal organ position variation in children during image-guided
radiotherapy. Radiation oncology, 13, 1-9.
[7] Momin, S., Fu, Y., Lei, Y., Roper, J., Bradley, J. D., Curran, W. J., ... & Yang, X. (2021).
Knowledge‐based radiation treatment planning: a data‐driven method survey. Journal of applied
clinical medical physics, 22(8), 16-44.
[8] Panayides, A. S., Amini, A., Filipovic, N. D., Sharma, A., Tsaftaris, S. A., Young, A., ... &
Pattichis, C. S. (2020). AI in medical imaging informatics: current challenges and future directions.
IEEE journal of biomedical and health informatics, 24(7), 1837-1857.
[9] Rubner, Y., Tomasi, C., & Guibas, L. J. (2000). The earth mover's distance as a metric for
image retrieval. International journal of computer vision, 40, 99-121.
[10] Šulc, Z., Matějka, M., Procházka, J., & Řezanková, H. (2017). Evaluation of the Gower
coefficient modifications in hierarchical clustering. Metodoloski zvezki, 14(1), 37-48.
[11] Winz, I. (2004). A decision support system for radiation therapy treatment planning.
Department of Engineering Science, University of Auckland, MsC.

Abstract (if available)

Abstract The radiation therapy treatment planning involves complex decision-making processes that rely heavily on the expertise of clinicians. This research introduces an innovative decision support system (DSS) designed to enhance personalized treatment recommendations in radiation therapy by leveraging advanced informatics methodologies. This research introduces an innovative decision support system for personalized radiation therapy treatment recommendations using advanced informatics methodologies. To compare patient profiles based on anatomical characteristics, it employs the integration of Gower's similarity measure and Earth Mover's Distance (EMD) analysis, enabling a nuanced assessment of dissimilarities.

The DSS computes EMD values for spatial target signatures (STS) and overlap volume histograms (OVH) across multiple regions of interest (ROIs). The effectiveness of the EMD metric is in quantifying dissimilarities between radiation therapy plans is validated through a testing methodology. This involves randomly shifting the contours of organs at risk (OARs) and target volumes while maintaining their original shape and size, allowing us to evaluate the EMD metric's sensitivity to anatomical location variability, and its accuracy in identifying similar treatment plans.

The implementation of this DSS involves the acquisition of diverse patient datasets from multiple institutions, ensuring the generalizability of the research outcomes. The data is efficiently organized and managed using MongoDB, a NoSQL database solution that allows for the storage of complex and varied data types.The successful integration of these advanced informatics methodologies paves the way for more personalized, precise, and effective radiation therapy treatment planning.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Knowledge‐driven decision support for assessing radiation therapy dose constraints

PDF

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

PDF

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

PDF

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

Asset Metadata

Creator Alshehri, Wejdan Ali (author)

Core Title Decision support system in radiation therapy treatment planning

School Viterbi School of Engineering

Degree Master of Science

Degree Program Biomedical Engineering (Biomedical Imaging and Telemedicine)

Degree Conferral Date 2024-08

Publication Date 08/09/2024

Defense Date 05/08/2024

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

Tag algorithm development,anatomical similarity,biomedical engineering.,decision support system (DSS),DICOM,Earth Mover's Distance (EMD),Gower's similarity measure,machine learning (ML),medical imaging,medical imaging informatics,OAI-PMH Harvest,organ at risk (OAR),overlap volume histogram (OVH),personalized treatment,radiation therapy (RT),spatial target signature (STS),testing and validation,treatment planning

Format theses (aat)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Liu, Brent (committee chair), Varghese, Bino (committee member), Zhou, Qifa (committee member)

Creator Email walshehr@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-oUC113998TEY

Unique identifier UC113998TEY

Identifier etd-AlshehriWe-13367.pdf (filename)

Legacy Identifier etd-AlshehriWe-13367

Document Type Thesis

Format theses (aat)

Rights Alshehri, Wejdan Ali

Internet Media Type application/pdf

Type texts

Source 20240813-usctheses-batch-1196 (batch), 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 author, as the original true and official version of the work, but does not grant the reader permission to use the work if the desired use is covered by copyright. It is the author, as rights holder, who must provide use permission if such use is covered by copyright.

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

Repository Email cisadmin@lib.usc.edu