Добавил:
Sekretar
kiopkiopkiop18@yandex.ru
t.me/Prokururor I Вовсе не секретарь, но почту проверяю
Опубликованный материал нарушает ваши авторские права? Сообщите нам.
Вуз:
Предмет:
Файл:Ординатура / Хирургия / Библиотека им академика М.И. Перельмана / Книга_5525_Библиотеки_им_академика_М_И_Перельмана.pdf
X
- •Foreword
- •Acknowledgments
- •Editor biographies
- •Yi Wang
- •X. Sharon Qi
- •List of contributors
- •1.2.3 Feature engineering and representation
- •1.2.4 Linear separability
- •1.2.5 Classical models
- •1.1 A brief introduction to AI
- •1.2 Machine learning basics
- •1.2.1 Learning paradigms
- •1.3 Artificial neural networks
- •1.3.1 Feed-forward neural networks
- •1.3.2 Recurrent neural networks
- •1.3.3 Convolutional neural networks
- •1.3.4 Attention
- •1.3.5 Training neural networks
- •1.3.6 Applications and use cases of deep learning
- •1.4 Model training and evaluation
- •1.4.1 Hyperparameters
- •1.4.2 Data split
- •1.4.3 Evaluation metrics
- •1.5 Generative models
- •1.5.1 Generative adversarial networks
- •1.5.2 Diffusion models
- •1.5.3 Applications and use cases
- •1.6 Ethical consideration and bias
- •1.6.1 Transparency and explainability
- •1.6.2 Bias and fairness
- •1.6.3 Data privacy violation
- •1.6.4 Risk and misuse
- •1.7 Summary
- •References
- •2.1 Introduction
- •2.1.2 Staff roles in radiation therapy
- •2.2 Overview of AI in radiation therapy
- •2.2.1 Patient evaluation and dose prescription
- •2.2.2 Treatment simulation
- •2.2.3 Contouring
- •2.2.4 Treatment planning
- •2.2.5 Quality assurance
- •2.2.6 Treatment delivery
- •2.2.7 Response assessment and toxicity management
- •2.3 Summary
- •3.1 Introduction
- •3.1.1 Introduction of clinical decision making and AI
- •3.1.2 The role of AI in clinical decision making
- •3.2 AI algorithms for clinical decision making
- •3.2.2 Radiomics
- •3.2.3 Data integration by AI
- •3.2.4 Interpretability of AI models
- •3.3 Application of AI in clinical decision making
- •3.3.1 Diagnosis and disease phenotyping
- •3.3.2 Personalized treatment
- •3.3.3 Treatment outcome and prognosis prediction
- •3.4 Challenges and future directions of AI in clinical decision making
- •3.4.1 Challenges and concerns
- •3.4.2 Future directions
- •3.5 Summary
- •References
- •4.1 Introduction
- •4.2 Imaging for treatment planning
- •4.2.1 CT simulation
- •4.2.2 4D-CT
- •4.2.3 PET/CT
- •4.2.4 MRI
- •4.3 Imaging for treatment guidance
- •4.3.1 Portal imaging
- •4.3.2 CBCT
- •4.3.3 CT-on-rail and CT-linac
- •4.3.4 MR-linac
- •4.3.5 PET-linac
- •4.4 Imaging for motion management
- •4.4.1 ExacTrac
- •4.4.2 Varian triggered imaging
- •4.4.3 4D-CBCT
- •4.4.4 Cine MRI
- •4.4.5 4D-MRI
- •4.4.6 Surface imaging
- •4.5 Imaging for treatment assessment
- •4.5.1 Contrasted CT
- •4.5.2 PET/CT
- •4.5.3 Functional MRI
- •4.6 Summary
- •5.1 Introduction to big data in radiation oncology
- •5.1.1 Overview of big data
- •5.1.2 Sources of big data in radiation oncology
- •5.1.3 Big data and AI in radiation oncology
- •5.2 Big data lifecycle in radiation oncology
- •5.2.1 Data aggregation and storage
- •5.2.2 Data sharing and security
- •5.4 The application of big data in radiation oncology
- •5.4.1 Medical image segmentation
- •5.4.2 Automatic treatment planning
- •5.4.3 Treatment response prediction
- •5.4.4 Quality assurance and patient safety
- •5.4.5 Clinical decision support
- •5.2.3 Data visualization
- •5.2.4 Knowledge creation and implementation
- •5.2.5 Data archiving and deletion
- •5.3 Big data analytics with AI
- •5.3.1 Data processing and integration
- •5.3.2 AI modeling
- •5.5 Challenges and future perspectives
- •5.6 Summary
- •Reference
- •6.1 The road to ART
- •6.1.1 3D conformal radiotherapy (3DCRT)
- •6.1.2 Intensity modulated radiotherapy (IMRT)
- •6.1.3 Image-guided radiotherapy (IGRT)
- •6.1.4 Adaptive radiotherapy (ART)
- •6.2 ART workflow and implementation
- •6.2.2 Current practice
- •6.2.3 Clinical impact
- •6.3 Considerations for implementing online ART
- •6.3.1 Time as a limiting factor
- •6.3.2 Implications for fast and reliable re-planning
- •6.3.4 Clinical considerations
- •6.4 Summary
- •7.1 Components of ART workflow
- •7.1.1 Simulation
- •7.1.2 Pre-planning
- •7.1.3 Online imaging and daily re-planning
- •7.1.4 Quality assurance
- •7.2 AI-driven ART
- •7.2.1 Simulation
- •7.2.2 Pre-planning
- •7.2.3 AI for delivery
- •7.3 Outlook and future directions
- •7.3.1 Real-time ART with AI
- •7.3.2 Dose escalation and functional adaption with AI
- •7.4 Summary
- •References
- •8.1 Introduction
- •8.2 Synthetic CT: deep learning methods
- •8.2.1 Conventional methods
- •8.2.2 U-Net
- •8.2.3 Generative adversarial networks
- •8.2.4 Denoising diffusion probabilistic model
- •8.3 Synthetic CT from CBCT
- •8.3.1 Noise and artifact reduction
- •8.3.2 Online dose calculation
- •8.3.3 Online image segmentation
- •8.4 Synthetic CT from MRI
- •8.4.1 Synthetic image accuracy
- •8.4.2 Dose calculation in MR-only radiation therapy
- •8.4.3 PET attenuation correction
- •8.4.4 Image registration
- •8.5 Discussion and outlook
- •8.6 Summary
- •References
- •9.1 AI-based image registration and segmentation for ART
- •9.1.1 Adaptive radiation therapy
- •9.2 Artificial intelligence
- •9.2.1 What is machine learning?
- •9.2.2 What is deep learning?
- •9.3 Deep learning: the basic components
- •9.3.1 Convolutional neural networks: looking at the picture
- •9.3.2 Pooling layers: keeping what matters most
- •9.3.3 Fully connected (dense) layers: bringing it all together
- •9.3.4 Activations
- •9.3.5 Loss: driving the model
- •9.3.6 Auto-encoders: remove the noise
- •9.3.7 Supervised versus unsupervised learning
- •9.3.8 Pre-trained convolutional neural networks
- •9.4 Image registration: bringing two images together
- •9.4.1 Registration similarity metrics
- •9.4.2 Types of registrations
- •9.5 AI-based image registration
- •9.5.1 Supervised learning
- •9.5.2 Unsupervised learning
- •9.5.3 Registration in ART
- •9.5.4 Commonalities in architectures
- •9.6 Image segmentation
- •9.6.1 Introduction: coloring by the numbers
- •9.6.2 Segmentation networks
- •9.6.3 Best practices
- •9.7 Summary
- •10.1 Introduction
- •10.1.1 Overview of chapter content
- •10.2 The landscape of AI-assisted dose prediction
- •10.2.1 Traditional machine learning for dose prediction
- •10.2.2 Deep learning-based dose prediction
- •10.2.3 Challenges in AI-assisted dose prediction
- •10.3 Re-planning workflows powered by AI
- •10.3.1 Deep learning for re-planning pipelines
- •10.4 Future directions of AI-assisted dose prediction and re-planning
- •10.5 Summary
- •11.1 Introduction
- •11.2.1 Imaging-based motion monitoring
- •11.2.2 Delivery system actions
- •11.2.3 Challenges for real-time ART implementation
- •11.3 AI in real-time ART workflows
- •11.3.1 Improving intrafraction motion monitoring through AI
- •11.3.2 Mitigating system latency through AI
- •11.4 AI for ART delivery: future directions
- •11.4.1 Management of non-respiratory motion
- •11.4.2 Training AI models with small or unpaired datasets
- •11.4.4 Biology-guided ART delivery
- •11.5 Summary
- •References
- •12.1 Introduction
- •12.2 Patient QA
- •12.2.1 Pre-planning QA
- •12.2.2 Pre-treatment plan QA
- •12.2.3 On-treatment QA
- •12.3 Treatment delivery systems and instruments
- •12.3.1 Machine commissioning
- •12.3.2 Machine QA
- •12.3.3 Dosimetry tool QA
- •12.4 Summary
- •References
- •13.1 Data resources for response modeling in radiotherapy
- •13.1.1 Clinical data
- •13.1.2 Imaging (radiomics)
- •13.1.3 Treatment planning (dosiomics)
- •13.1.4 Multiomics
- •13.2 Radiotherapy treatment outcome modeling
- •13.2.1 TCP/NTCP in radiotherapy
- •13.2.2 Clinical outcomes versus PROs
- •13.2.3 Machine learning response prediction
- •13.2.4 Explainability of ML response models
- •13.2.5 Sample use cases
- •13.3 AI response-based adaptive radiotherapy
- •13.3.1 Requirements and challenges
- •13.3.2 Prediction versus treatment optimization
- •13.3.3 Sample use cases
- •13.4 Challenges and recommendations
- •13.5 Summary
- •Acknowledgments
- •References
- •14.1 Overview of challenges in AI-driven ART
- •14.2 Data challenges
- •14.2.1 Data availability
- •14.2.2 Data quality
- •14.2.3 Data privacy
- •14.3 Technical challenges
- •14.3.2 Model robustness and generalizability
- •14.3.3 Model explainability and interpretability
- •14.4 Challenges associated with online and real-time workflows
- •14.4.1 Image quality
- •14.4.2 Dose calculation
- •14.4.3 Real-time ART
- •14.5 Operational challenges
- •14.5.1 Clinical validation
- •14.5.3 Staff training
- •14.5.4 User experiences
- •14.5.5 Quality management program
- •14.5.6 Financial challenges
- •14.6 Ethical, regulatory, and legal challenges
- •14.6.1 Ethical issues
- •14.6.2 Regulatory and legal issues
- •14.7 Summary
- •References
- •15.1 Clinical considerations for CT-based offline ART
- •15.1.1 Patient and site selection
- •15.1.2 Re-simulation
- •15.1.3 Re-planning
- •15.1.4 Plan summation and evaluation
- •15.1.6 Limitations and future directions
- •15.2 Clinical considerations for CBCT/CT-based online ART
- •15.2.2 Patient and site selection
- •15.2.3 Simulation
- •15.2.4 Pre-planning review
- •15.2.5 Reference planning
- •15.2.9 Limitations and future directions
- •15.3 Summary
- •References
- •16.1 Introduction
- •16.2 Overview of MRI-guided ART systems
- •16.3 MRI-guided ART workflow
- •16.4 AI applications for MRI-guided ART
- •16.4.1 Synthetic CT generation
- •References
- •16.4.2 Auto-segmentation
- •16.4.3 Image registration
- •16.4.4 Others
- •16.4.5 Future AI development and implementation
- •16.5 Summary
- •17.1 Functional PET-guided ART
- •17.1.1 PET-based functional imaging overview
- •17.1.2 From anatomy to function: the power of PET in radiation therapy
- •17.1.5 Conclusions and future prospects
- •17.2 Functional MRI-guided ART
- •17.2.1 From anatomy to function: the power of functional MRI in radiation therapy
- •17.2.4 Conclusion and future prospects
- •17.3 Summary
- •References
- •18.1 Proton ART
- •18.1.1 Clinical context and necessity
- •18.1.2 Patient populations
- •18.1.4 Rationale for AI in proton ART
- •18.2 AI in proton ART
- •18.2.1 Imaging
- •18.2.2 Deformable and rigid registration
- •18.2.3 Contour propagation
- •18.2.4 Dose calculations
- •18.2.5 Plan optimization
- •18.2.6 Other developments
- •18.3 Implementation of adaptive proton therapy
- •18.4 Summary
- •References
- •19.1 Designing clinical trials with AI
- •19.1.1 The essential role of clinical trials
- •19.1.2 Trial protocols and methodologies
- •19.1.3 AI-driven clinical trial design and execution
- •19.1.4 Incorporation of digital twins (DTs) in clinical trials
- •19.2 Implementation of AI in ongoing clinical trials
- •19.2.1 Integration with existing clinical trial frameworks
- •19.2.2 Quality assurance, compliance, and standardization
- •19.3 Case studies of AI in adaptive radiotherapy trials
- •19.3.1 Overview of guidance for advanced radiotherapy in clinical trials
- •19.3.2 AI in the radiotherapy clinical trial quality assurance processes
- •19.4 Ethical and regulatory considerations
- •19.4.1 Patient consent and data privacy
- •19.4.2 Bias, fairness, and transparency
- •19.4.3 Regulatory guidelines and compliance
- •19.5 Future directions and challenges
- •19.5.1 Emerging technologies and techniques
- •19.5.2 Alternative strategies
- •19.6 Conclusion
- •19.7 Summary
- •References
- •20.1 Risk management
- •20.1.1 Prospective risk assessments
- •20.1.2 Root cause analysis

Artificial Intelligence in Adaptive Radiation Therapy
Figure 13.2. Diagrams of feature-based and featureless radiomics workflow.
13-6

Artificial Intelligence in Adaptive Radiation Therapy
TCP/NTCP models is limiting for radiotherapy personalization, as they do not
reflect patient-specific conditions and factors. Therefore, there has been a great effort
in the last few years to apply machine and deep learning techniques in 3D dosiomics.
Deeper features, whether extracted as hand-crafted features, similar to those in
radiomics (histogram-based, shape-based, texture-based, etc), or using a deep
learning approach, can better reflect the real impact of the person-specific dose
distribution on clinical outcomes. Many studies have been published on dosiomic
predictors of radiation toxicities [79–83] and tumor control [84, 85], and they proved
to be promising for clinical application. Moreover, special attention should be paid
to models combining dosiomic features with radiomics or even other -omics (see
section 13.1.4), which often significantly improves the model’s performance [86–88].
13.1.4 Multiomics
The growing field of biotechnology and bioinformatics offers numerous valuable
tools and data that can positively contribute towards the accuracy of outcome
models, their clinical translation, and precision medicine in general. A wide range of
molecular biomarkers can currently be analysed and included in response modeling.
Most commonly, they are obtained by analysis of the tumor genome (genomics), its
further transcriptions into RNA (transcriptomics), translations into proteins (proteomics), and final metabolites (metabolomics) [89]. Combining all these biomarkers
in response models is often referred to as a multiomic approach, due to the
combination of different -omic datasets. Furthermore, extending the multiomics
approach by radiomics led to the emerge of radiogenomics [90–92].
Genomics represents the identification of structural variations in the DNA of the
studied specimen. Since there are various classes of genomic variables, it is always
important to consider the aim of the analysis and the specimen type before selecting
specific genomic biomarkers for modeling. However, the most common classes in
radiotherapy are single nucleotide polymorphisms (SNP) and copy number variations (CNV) [93, 94]. Typically, they can be associated with radiation sensitivity, and
thus they can, for example, work as prognostic factors of radiation toxicities. There
are millions of SNPs and CNVs per 100 nucleotides. Consequently, it is crucial to
have large sample sizes to assess the prognostic power of specific SNPs. One of the
largest collaborations investigating associations between SNPs and radiation
toxicities is the radiogenomics consortium (RGC) [95–97].
Transcriptomics is a quantitative and qualitative sequential analysis of RNA
molecules, including their expression levels, functions, and degradations [98, 99].
RNA is a translation between DNA and protein. It carries genes to ribosomes,
where proteins are created. However, it also regulates the expression levels of genes
and might have structural functions, such as ribosomes. Therefore, transcriptomics
can have a major impact on cancer prognosis as it captures extensive and complex
information. As is evident from various studies, transcriptomics is a promising part
of multiomics in radiotherapy [100–102].
Proteomics involves various analytical techniques to identify the composition,
structure, function, and interactions of proteins. It provides another necessary input
13-7

Artificial Intelligence in Adaptive Radiation Therapy
to the holistic approach of multiomics, as it describes the overall organism reaction
to the treatment better than just genomics. Perhaps the most helpful protein
expressions in radiotherapy are cytokines, secreted by different immune cells
responsible for speci fic organism responses to radiotherapy, such as radiation
toxicities [103–105].
Metabolomics is the profiling of metabolites from collected specimens.
Metabolites are final products of gene transcriptions, i.e. metabolism, which
conducts essential cell functions, such as energy conversion and metabolic waste
management. There is a vast potential for metabolomics in oncology, for example,
in hepatocellular carcinoma [106, 107], glioma [108, 109], and non-small cell lung
cancer (NSCLC) [110, 111]. In radiotherapy, metabolomics shows promising results
in predicting and evaluating treatment outcomes of various diagnoses [112–116].
13.2 Radiotherapy treatment outcome modeling
Radiotherapy outcome modeling historically originated from population-based
models, which are often inaccurate due to the lack of individualization of the
patient’s response. Population-based models use large cohorts of patients to
determine relationships between treatment details and clinical endpoints.
Consequently, they come up with relationships that describe average trends in the
cohort. Typical examples of such models are tumor control probability (TCP) and
normal tissue complication probability (NTCP), which were limited to dose
response. Currently, with the expansion of machine learning techniques, outcome
modeling is becoming a patient-specific task. Therefore, new requirements, such as
model’s explainability, are usually associated with it [2].
13.2.1 TCP/NTCP in radiotherapy
The most common outcome models in radiotherapy are TCP of the targeted volume
and NTCP of surrounding OARs. The goal of radiotherapy is to maximize TCP
while minimizing NTCP. Unfortunately, delivering dose to the target volume is
always associated with a certain risk of radiation toxicities in normal tissues.
Therefore, the main task in radiotherapy is to optimize the trade-off, i.e. the
therapeutic ratio (TR) between TCP and NTCP (see figure 13.3)[117, 118].
There are two overarching methods for TCP/NTCP modeling: analytical and
data-driven [32]. A more traditional (analytical) approach represents populationbased dose response models, that are mostly utilizing the mechanistic radiobiological linear-quadratic (LQ) model [119, 120] and DVH data. The LQ model is a
current standard for various clinical calculations and assessments, for example,
when clinical outcomes of different treatment regimens need to be compared. It can
be expressed as a clonogenic cell survival fraction (SF):
=
eSF ,
()
13.1
()
2
ab−−
DDGD
where D denotes total dose, α and β are organ-specific parameters describing lethal
damage induced by double and single strand DNA breaks, respectively, and G(D)is
13-8

Artificial Intelligence in Adaptive Radiation Therapy
Figure 13.3. Illustration of typical TCP and NTCP curves. Probability of tumor (local) control rises with the
dose. However, this also elevates the dose in surrounding organs and thus the probability of radiation toxicity
(NTCP) increases as well. The therapeutic ratio refers to a window between TCP and NTCP, defined as the
tumor complication-free TR = TCP × (1-NTCP).
the Lea–Catcheside factor accounting for clonogen repopulation between fractions
[121, 122].
Analytical TCP models can have different definitions, but one of the most usual
follows Poisson distribution and can be described as
NSF
·
=−eTCP ,
13.2
()
where N is the initial number of tumor cells. A detailed review on the use of TCP
models in radiotherapy was published by Zaider and Hanin [123].
Normal tissue complication modeling is similar to TCP, although it is more
difficult due to the higher complexity of the problem. Usually, several different
OARs are irradiated during the treatment, and their response differs. Moreover,
NTCP depends on the spatial dose distribution, which makes the task even more
complicated as higher dimensional data are involved. Therefore, published analytical NTCP models often focused on reducing DVH dimensionality into a single or a
few metrics, making routine implementation and understanding easier. The most
used model involves the Lyman–Kutcher–Burman (LKB) method, which is defined
by the following equations:
uD V,
⎛
⎜⎟
DV e dxNTCP ,
⎝
⎛
⎜⎟
uD V
,
⎝
1
⎞
=
2
⎠
DV
⎞
=
mV
⎠
()
∫
−∞
p
−
TD
50
TD
·()
50
2
−
x
()
2
, 13.3
()
, 13.4
()
()
13-9

Artificial Intelligence in Adaptive Radiation Therapy
TD 1
()
50
()
50
=V
, 13.5
n
V
()
TD
where TD50(V) is the tolerance dose for partial volume V that results in 50%
complication probability after 5 years, m is a measure of the slope of the dose
response curve (i.e. the standard deviation of TD
(1)) and n is a tissue-specific
50
volume effect factor, determining the serial/parallel type of the structure [32, 118].
The most impactful publications of dose–volume responses in various OARs are
summarized in QUANTEC review articles [124, 125] and subsequent HyTec for
hypofractionation studies [76]. Provided data are nowadays commonly used in
clinics as treatment planning criteria, limiting the probability of radiation-induced
post-treatment complications.
Another approach to TCP/NTCP modeling involves AI techniques, which can
usually provide better accuracy than analytical models, as they can capture the nonlinear relationship between input data and clinical endpoints. They create a new
opportunity to personalize the outcome and potentially the treatment by including
patient-specific data from different resources (see section 13.1) and exploring their
underlying interactions [32]. Many different types of machine learning techniques
can be effectively utilized in radiotherapy outcome modeling. More details are
summarized in following sections.
13.2.2 Clinical outcomes versus PROs
Next to conventionally followed and physician-assessed clinical endpoints, such as
tumor control, survival, andr various radiation toxicities, more patient quality of life
(QoL) focused endpoints have recently been studied. This typically involves patient
reported outcome (PRO) evaluations, which is a systematically prepared questionnaire, usually filled in by patient before, during, and after the treatment in follow-up
appointments. PRO questions and their evaluation are disease-specific. Therefore,
different departments and clinics might use other questionnaires, depending on a
patient’s needs that are specific in given medical department. One of the most
common questionnaires in oncology is the Edmonton Symptom Assessment System
(ESAS) [126, 127]. It is routinely used in several oncology clinics for symptom
management and patient QoL monitoring [128–132]. However, another potential
application is in the prediction of clinical endpoints, for example, the overall
survival. Recent studies showed evidence of ESAS PROs as significant predictors
of survival in cancer patients [133–135]. This might help with various clinical
decisions, such as transitioning from curative to palliative treatment.
The prognostic power of PROs offers a variety of potential applications in
outcome modeling. Incorporating time-sequential PROs into outcome models could
improve their accuracy and impact. One of the possible applications is outcome
models that could be used to optimize treatment plan parameters not only with
respect to typical clinical endpoints, but also concerning the patient-specific QoL.
An example of effort made in this area is a project led by the H Lee Moffitt Cancer
Center and Research Institute (Data Science to Improve Treatment Planning for
Advanced Prostate Cancer Patients Treated with Radiotherapy), supported by the
13-10

Artificial Intelligence in Adaptive Radiation Therapy
Department of Defense (DoD) Congressional Directed Medical Research Program.
The project focuses on using PROs in radiotherapy outcome modeling and
predicting patients’ QoL after the treatment.
Patient’s PROs are sequential data acquired at different time points.
Consequently, it has to be dealt with as a data sequence, which represents a specific
task regarding suitable machine learning architecture. The inspiration can be taken
from natural language processing (NLP) models, which are designed exactly for data
sequences, i.e. words and sentences. Typical architectures in this field include
recurrent neural networks (RNN), such as long short-term memory (LSTM), and
gated recurrent unit (GRU), or recently more advanced and famous transformers.
13.2.3 Machine learning response prediction
As a subcategory of AI, machine learning (ML) technologies influence almost all
fields currently. There are numerous possible medical applications, and major
benefits are also evident in radiotherapy outcome modeling. Machine learning
enables the aggregation of a large amount of data from different resources (i.e.
multiomics), allowing for further personalization in outcome prediction, while
increasing its accuracy. This can be attributed to AI capabilities in capturing
underlying relationships between the provided data and recognizing inter-human
cancer-specific differences, which are usually impossible for humans to detect and
process. Therefore, in the last decade, machine learning response prediction has
dominated over traditional approaches, based on simple TCP/NTCP dose–volume
responses. It is further expected that ML-based outcome modeling will play a
leading role in the future exploration of cancer biology and personalization of
treatment planning/adaptation, where the patient-specific response should always be
considered a priority. It is especially envisioned as a key part of future clinical
decision-support systems (CDSS), which can assist in several clinical tasks [1, 2].
Initially, it can serve as a decision-support system for determining the optimal
treatment procedure. Further, it can guide treatment planning for the optimized
outcomes. In the next step, it can be applied during radiotherapy adaptation of an
ongoing treatment course. Finally, it can provide an early prediction of recurrence
or post-treatment complications, enabling physicians to determine and apply the
optimal follow-up treatment. Therefore, there are various data resources and
formats that can be employed in the aforementioned tasks. Generally, there are
two data formats—structured and unstructured. Structured data are most often
tabular, while unstructured data involve images, 3D dose distributions, or unstructured clinician-provided notes.
There are numerous algorithms that can be used for structured data. Typically,
they belong to conventional ML algorithms, although artificial neural networks
(ANN) can also be used. Typical examples of conventional ML methods used in
radiotherapy outcome modeling are logistic regression, support vector machine
(SVM), decision tree (DT), random forest (RF), Bayesian network (BN), and
various ensemble methods, such as extreme gradient boosting (XGBoost).
13-11

Artificial Intelligence in Adaptive Radiation Therapy
Logistic regression is a classification ML technique. Typically, it is used for
modeling discrete outcomes by fitting a sigmoid function to extracted features
(multivariable logistic regression) [136–138]. The logistic regression requires the
features to be uncorrelated and independent. Therefore, a careful feature selection
has to be performed before the regression. In dose–volume space, this practically
implies using a single summarizing dose–volume metric, as DVH metrics are highly
correlated [118].
The support vector machine is a regression/classification ML technique, which
uses a kernel function to define a ‘hyperplane’ separating data points according to
their classes, for example, tumor local control/recurrence or survival/death [139,
140]. Depending on the kernel definition, SVM algorithms can be categorized as
linear, polynomial, sigmoid, or radial basis.
Another type of conventional ML technique used as a classifier in outcome
modeling is the Bayesian network (BN). It is a probabilistic model based on
conditional dependencies among input variables. It is a type of directed acyclic
graph (DAG), thus it is easily interpretable. Thanks to the probabilistic character of
the BN, it can work with partially missing input data. However, some of the
drawbacks are the need for larger datasets compared to other methods or poorer
identification of causality in the dataset. A simpler alternative to BN is the naïve
Bayesian network (NBN), which assumes independence in input features. Therefore,
it is more effective and easier to train, even though it does not often correspond to
real scenarios. Nevertheless, it is very popular in many applications, including
radiotherapy outcome prediction, often outperforming other ML methods [141].
Decision tree is a non-parametric classification/regression ML technique that
separates data points based on recursive partitioning analysis. Due to its clear and
simple design, it is very convenient for interpretation. A more advanced alternative
(random forest) utilizes an ensemble approach, building multiple DTs and averaging
their results. It is currently one of the most popular conventional ML techniques due
to its accuracy and direct interpretability, and it is usable in time-to-event modeling
as well [138, 142–146].
The ensemble technique is a concept combining more ML models to improve the
prognostic accuracy. One of the most utilized and popular is the gradient boosting
machine (GBM) [147]. It is based on sequentially adding new models focusing on
previously misclassified samples. The idea in each step is to create a model that
correlates with the negative gradient of the ensemble loss function. There are several
implementations of GBM and one of the most utilized is XGBoost [148–150].
Unstructured (raw) data (e.g. dose distributions, images, -omics, etc) are usually
processed using deep learning (DL) methods (i.e. deep neural networks (DNN)),
which are built by stacking up multiple layers of specifically designed mathematical
operations (see the featureless approach in figure 13.2.). This technique is able to
identify underlying representations and relationships to learn specific tasks without
any prior feature selection. Typically employed architectures are the multilayer
perceptron (MLP), convolutional neural networks (CNNs), recurrent neural network (RNNs), autoencoders (AEs), and generative neural networks, such as
13-12

Artificial Intelligence in Adaptive Radiation Therapy
Figure 13.4. Diagrams of basic deep learning architectures. (1) Multilayer perceptron consisting of several
fully connected layers; (2) variational autoencoder; (3) convolutional neural network followed by MLP; and
(4) recurrent neural network with LSTM units.
generative adversarial networks (GANs) and denoising diffusion probabilistic
models (DDPMs).
A multilayer perceptron is a basic neural network consisting of several fully
connected layers containing multiple nodes (see figure 13.4). One node is a weighted
sum of the nodes from the previous layer, followed by an activation function (e.g.
ReLu, SeLu, sigmoid, etc). The aim of outcome modeling is usually predicting an
endpoint probability. Therefore, the last layer of MLP is typically followed by a
sigmoid activation function to simulate the probability. The MLP works only with
1D data, thus any higher dimensional data has to be flattened before the MLP can
be applied [62, 88].
13-13

Artificial Intelligence in Adaptive Radiation Therapy
The convolutional neural network is an architecture typically used for 2D and 3D
data, although it can also be applied to 1D data. Compared to MLP, the CNN
contains fewer parameters, which makes training easier. Consequently, it requires
less training data than MLP. The main blocks of the CNN are convolutional layers,
which represent a convolution between a specific kernel and the previous layer. The
CNN maintains local connectivity and spatial invariance, which is one of its main
advantages and also the reason why it is so useful in higher dimensional unstructured data. A typical CNN consists of several layers, each usually followed by a
down-sampling using a pooling layer. The output of a CNN is a set of most
significant features that can be further used as an input for a small MLP to provide
the outcome prediction (see figure 13.4)[62, 151].
An autoencoder is an unsupervised DL technique with many implementations.
One of the most popular, due to its continuous latent space described by a
probability distribution, is the variational autoencoder (VAE) (see figure 13.4).
Even though the VAE cannot be used directly for outcome prediction, since it is an
unsupervised technique, it is still a useful approach for feature extraction. Any
autoencoder consists of an encoder and a decoder. The encoder extracts features and
maps them to a latent space while the decoder learns to reconstruct the original input
from the latent space. The latent space representation is often used in outcome
modeling for feature extraction and dimensionality reduction. It can also be
combined with some supervised techniques, such as MLP, and utilized in outcome
modeling directly in an end-to-end approach [152].
Another type of data, common in outcome modeling, is sequential or longitudinal
data, such as a time-sequence of images or textual information. For this type of
unstructured data, RNNs (see figure 13.4) are commonly used. The main idea
behind an RNN is to capture longitudinal associations between the input data. One
of the most common and effective architectures is long short-term memory (LSTM)
or gated recurrent unit (GRU) [153, 154]. A more recent technique that outperforms
traditional RNN methods in many tasks, especially in NLP, is the transformer [155,
156]. It is an architecture utilizing a multi-head attention mechanism, making it
faster and more robust than LSTM or GRU.
Generative neural networks represent a category of machine learning frameworks
which aims to learn contextual information and underlying representations to
generate unique new samples. The most typical examples of such models are
GANs and, recently, DDPMs. Previously, most of the generative tasks were
assigned to VAEs, which could generate new samples by randomly sampling the
latent space. A more recent architecture involves GAN, which trains two neural
networks to compete against each other (the generator and discriminator). The
discriminator is trained to distinguish between real and synthetic images, while the
generator is trained to create synthetic samples to deceive the discriminator. Even
though GAN produces more realistic images, it is often associated with a mode
collapse [157]. The most recent and highly promising generative architecture,
outperforming GANs, is based on denoising diffusion models (DDPM) [157, 158].
Generative models in outcome modeling can be used for data augmentation or
13-14

Artificial Intelligence in Adaptive Radiation Therapy
addressing the class imbalance in the training dataset, as shown in the study by
Dudas et al [159].
Deep learning methods are promising alternatives for radiotherapy outcome
modeling. They usually exhibit excellent performance and offer a large variability in
the input data type. However, their explainability is not as simple and straightforward as in conventional machine learning techniques. Another disadvantage is the
amount of data, which has to be big enough, since DL techniques are usually
associated with more trainable parameters than conventional ML methods.
Consequently, DL models have to be trained cautiously to avoid overfitting. All
critical aspects regarding training, validation and testing of outcome models are
summarized in the TRIPOD (‘Transparent reporting of a multivariable prediction
model for individual prognosis or diagnosis’) reporting guideline [160]. In general,
predictive models should always be cross-validated and tested. The cross-validation
can be performed using one of the recommended methods, for example, k-fold or
leave-one-out. The testing can be done prospectively or retrospectively on a held-out
independent dataset [1].
13.2.4 Explainability of ML response models
The ultimate goal of ML outcome models, whether used alone or as a part of a
complex clinical system, is to provide decision-making support to clinicians.
Therefore, the key part of this human–machine interaction is the machine’s
explainability. Clinicians must understand the reasoning behind the algorithm’s
decision [161]. Moreover, explainability allows for better alignment with ethical
principles in applications with an impact on individuals. It can also help identify
errors, thus facilitating further improvements to the model. Some ML techniques,
such as decision trees or Bayesian networks, are inherently well-explainable due to
their simple architecture, although their performance is limited. In most cases,
models with higher performance usually have poorer explainability. For example,
deep learning methods often outperform conventional ML models, but their
explainability is complicated, since they work with implicitly learned features, which
are abstract to the human mind [1].
Three main explainability techniques are commonly used in outcome modeling.
These are local interpretable model-agnostic explanations (LIME), Shapley values,
and gradient-weighted class activation mapping (Grad-CAM).
LIME is an agnostic model, which approximates the model with a simple
interpretable model locally. Then, it perturbs selected input samples and evaluates
its impact on the local model approximation [161, 162]. It can be applied to any
outcome model which incorporates a classifier.
The Shapley values method, another agnostic model, arises from a cooperative
game theory, where marginal contributions of each input variable towards the
model output are evaluated. The most popular implementation, easily applicable to
any ML/DL architecture, is Shapley additive explanation (SHAP) [161, 163].
Grad-CAM [164] is a model-specific method, especially suitable for CNNs
(figure 13.5). It utilizes gradient information from a specific layer to illustrate the
13-15
Соседние файлы в папке Библиотека им академика М.И. Перельмана
