Добавил:
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
Скачиваний:
0
Добавлен:
31.08.2026
Размер:
29 Мб
Скачать
Artificial Intelligence in Adaptive Radiation Therapy
Figure 13.5. Grad-CAM analysis of dose distributions for patients with LRexamples of dose distribution slices with Grad-CAM maps, including the time of local recurrence in years. (Reproduced with permission from [
151]. Copyright 2024 Elsevier.)
importance of each neuron in the model behavior. Typically, it is applied to the last convolutional layer, as it has the most signicant impact on the models decision. Its main advantage, and the reason why it is very popular in outcome models using imaging data, is the ability to visualize the models focus region spatially. See, for example, Grad-CAM applied in the deep learning model, predicting local failures in NSCLC treated with stereotactic body radiation therapy (SBRT) published by Dudas et al [151].
13.2.5 Sample use cases
13.2.5.1 Lung cancertumor local control prediction
A deep learning outcome model predicting local control in NSCLC patients treated with SBRT was adopted in the study by Dudas et al [151](figure 13.6). The model utilized multi-modality data. It was built from three separate neural networkstwo CNNs and one VAE. The CNNs were applied to extract features from planning CT images and planning dose distribution, while the VAE was used to extract under­lying features from clinical/demographic patient details. The outputs from all three NN were then concatenated and fed into a discrete-time survival NN, as shown in gure 13.6. The survival neural network predicted the conditional probability of tumor local control.
The dataset was split according to TRIPOD criteria type 2b [160]. The rst part, 80% of the data, were used for stratied ve-fold cross-validation. The remaining 20% were kept as an independent dataset for nal testing.
The explainability technique Grad-CAM was applied to identify the parts of the data which are the most signicant with respect to the local control prediction.
13-16
Artificial Intelligence in Adaptive Radiation Therapy
Figure 13.6. Diagram of the DL-based outcome model for prediction of local control in NSCLC patients treated with SBRT. (Reproduced with permission from [
151]. Copyright 2024 Elsevier.)
Subsequently, the results were used to guide the design of planning dose criteria to mitigate local failures.
13.2.5.2 Breast canceracute skin toxicity prediction
The study by Saednia et al [165] presented conventional ML models, based primarily on a random forest approach, to predict acute skin toxicity (radiation dermatitis)
13-17
Artificial Intelligence in Adaptive Radiation Therapy
after whole-breast RT. The model was trained on quantitative biomarkers, extracted as surface temperature and texture features from thermal images, acquired before and during the treatment (after the fth, tenth, and fteenth treatment fractions).
The dataset consisted of 90 patients, split into training (83%) and independent testing (17%) subsets. The model was cross-validated using the leave-one-out methodology. The highest prediction accuracy (testing accuracy = 0.87) was associated with thermal biomarkers obtained after the fth treatment fraction.
13.2.5.3 Head and neck cancerloco-regional failure, distant metastasis, and
overall survival prediction
About 300 multi-institutional patients from The Cancer Imaging Archive were involved in the study by Diamant et al [166] to develop DL-based models predicting loco-regional failures (LRF), distant metastasis (DM), and overall survival (OS). CNNs, followed by MLPs, were applied to extract features from pre-treatment CT images and model the endpoints.
The training/validation dataset included 194 patients, while the independent testing dataset included 106 patients. The AUCs of predicting DM, LRF, and OS were about 0.88, 0.65, and 0.70, respectively. The study also demonstrated the use of explainability techniques (grad-CAM), showing a signicant difference in the decision explanation of patients who did or did not develop distant metastasis.
13.2.5.4 Prostate cancerdisease progression prediction
Conventional ML methods were applied in the study by Nayan et al [167] to predict a grade progression in very-low and low-risk prostate cancer patients. Various clinical and demographic features, such as age, family history of prostate cancer, tumor stage, PSA level, prostate size, biopsy characteristics, and use of a 5-alpha reductase inhibitor were included in the modeling.
The study investigated several ML approaches, including logistic regression (the traditional and ML versions), SVM, RF, and ANN. Using a stratied random split, the dataset was split in a ratio of 80% (training/validation) and 20% (independent testing). The results showed a superior performance of ML approaches predicting progression on active surveillance for prostate cancer.
13.2.5.5 Liver canceroverall survival prediction
Three types of patient data (clinical, radiomics, and contrast-enhanced CT images) were used in the study by Wei et al [168] for training a multi-modality DL-based outcome model to predict the OS in hepatocellular carcinoma (HCC) patients treated with SBRT.
Three separate models were trained: (i) VAE latent representation of clinical details fed into DeepSurv NN [169]; (ii) VAE latent representation of radiomic features fed into DeepSurv NN; and (iii) CNN-extracted features from CECT fed into DeepSurv NN. Subsequently, all three models were combined in a joint survival model for clinical, radiomic, and CECT features, as shown in gure 13.7. The joint model reported the best performance (c-index 0.650; 95% CI 0.635–0.683).
13-18
Artificial Intelligence in Adaptive Radiation Therapy
Figure 13.7. VAE-based clinical survival NN, VAE-based radiomics survival NN, and CNN-based CECT survival NN were jointly employed in the OS prediction of HCC patients treated with SBRT. (Reproduced with permission from [
168]. Copyright 2021 Elsevier.)
The integrated gradient method provided the interpretation and concluded the increased importance of normal tissue status for the OS.

13.3 AI response-based adaptive radiotherapy

Adaptive radiotherapy (ART) has been advancing since the late 1990s [170]. Typically, it is associated with anatomy-based modications, requiring new PTV/ OARs delineations, or biology-based modications using simple population-based response models. However, with the emergence of AI, multi-modality imaging and biotechnology, more advanced and individual-based approaches to response mod­eling were adopted. Consequently, signicant improvements in the personalization of ART can be achieved. Patient-specic response models can accurately predict treatment outcomes while considering the actual state of a patient on treatment. Therefore, it allows for personalized treatment adjustments during the treatment course. This framework is often called knowledge-based response-adapted radiation therapy (KBR-ART) [171]. Its diagram, in the context of current and previous frameworks for ART, is in gure 13.8.
13.3.1 Requirements and challenges
All essential requirements and challenges related to response-based adaptive radio­therapy are encompassed in the acronym KBR-ART. These are knowledge, response, and adaptation. The adaptation process is formulated based on
13-19
Artificial Intelligence in Adaptive Radiation Therapy
Figure 13.8. Comparison of previous, current and KBR frameworks for ART. (A) Previous framework treatment planning provided on imaging data, no adaptation during the course, outcomes estimated using population-based response models. (B) Current frameworktreatment planning provided on imaging data, anatomy-based adaptation, outcomes estimated using population-based response models. (C) KBR-ART frameworkindividualized treatment planning utilizing imaging and multiomics data for informed response­based planning, personalized response-based adaptation, outcomes estimated using personalized response models. (Reproduced from [
171]. CC BY 4.0.)
predictions from the response model, which, in turn, relies on the availability of comprehensive patient datasets. As is evident, each KBR-ART component is interdependent and cannot work without the rest.
The rst requirement is knowledge (i.e. patient data). As introduced in section
13.1, various data resources prove valuable for outcome modeling. Particularly data
that capture characteristics of the patient, the patients disease, and planned treatment. That includes clinical and demographic details (see section 13.1.1), imaging biomarkers (see section 13.1.2), treatment planning data (see section
13.1.3), and multiomic biomarkers (see section 13.1.4).
The second requirement is a suitable and reliable response model. Modeling the treatment response is a key part of modern oncology research, as it enables treatment personalization. Radiotherapy outcomes are typically expressed in terms of TCP and NTCP. Both can be modeled either analytically or via data-driven models. While analytical methods have clear interpretation, ML/DL approaches are often poorly interpretable with limited or no radiobiological understanding. However, ML/DL-based outcome models usually outperform analytical models, since they are capable of nding important underlying data representations. Common outcome modeling approaches were reviewed in section 13.2.
13-20
*
*
*
Artificial Intelligence in Adaptive Radiation Therapy
Finally, the third requirement is an adaptation of decision-making, illustrated in gure 13.8 by the adaptation function π. The question is, how can we use outcome models in a strategically optimal manner to adapt a patients treatment plan? The resolution of this challenge leverages modern AI techniques. Architectures such as reinforcement learning (RL) have a crucial impact in this area, as they are specically suitable for adaptive systems and sequential decision-making [172
177]. Their primary advantages lies in their ability to explore the effects of
decision-making several steps ahead, for example, in dose regimen adaptations between radiotherapy fractions. Other ML methods applicable in KBR-ART involve Bayesian networks [178] or quantum deep RL [174].
The core component of RL is the Markov decision process (MDP), a framework applied in sequential decision-making models. Each step of the MDP can be described as four-tuple (s the system (s (a
A), R is the reward function for a given action, and γ is the importance rate,
t
S), atis the action undertaken in step t (i.e. from state stto s
t
, at, Rt, γt), where stdenotes the current state of
t
t + 1
which sets how far in the future the reward is considered. The system assigning actions to states (i.e. s
× at) is called the policy. The policy in KBR-ART is called the
t
adaptation function (π)[171, 172]. The ultimate goal of RL, and other MDP methods, is to nd a policy that maximizes cumulative reward in the modeled problem. The objective of MDP (the Q function), being maximized, is mathemati­cally expressed by the following equation:
The optimal policy
p
() ( ()) ( )
Qsa Rs s, E , . 13.6
p
is determined as
Q
t
gp=
t
=
tts,a
0
p
,
p
p
. However, equation (13.6)is
Qmax
impractical for implementation due to the innite sum. A more effective formulation is called the Bellman equation:
)
p
∼∼
Qsa E Rsa Qsa, , max , , 13.7
such that
() () ( ) ( )
+
i
1
=
QQlim
i
→∞
i
p
, and
()
sPsa
,
{}
′∼sPsa,()
state s to state s. As proved earlier, the convergent point
g=+
′∈ ′∈
aA;sS
i
denotes the transition probability from
p
is optimal [179]. The Q
Q
function is often solved utilizing deep learning techniquesdeep Q-networks (DQNs).
In KBR-ART, MDP simulates the clinical workow. Each step (i.e. delivery of each fraction) is described by a patient state s
, reward Rt, and external action atto
t
maximize the outcomes. The patient state is commonly characterized by TCP/NTCP models, utilizing various input data (see section 13.2.1). The reward is usually a predened combination of treatment outcomes, prioritizing maximum TCP and minimum NTCP. The external action may involve, for example, dose escalation, localized radiation boosting, adapting the number of fractions, adapting the target volume or even boosting the treatment with other modalities, such as chemotherapy
13-21
Artificial Intelligence in Adaptive Radiation Therapy
or immunotherapy. In practical deployment, the only option to change the current patient state and maximize treatment benets is by actions of the adaptation function. The optimal decision-making for each state is trained in the RL model via numerous trial/error (s
tRt
)
(s
, R
t + 1
a
t
) interactions.
t + 1
13.3.2 Prediction versus treatment optimization
Precision medicine can encompass many tasks in a real clinical environment, such as anatomy-based or outcome-optimized treatment adaptations. Personalized treatm ent adaptations may be performed before, during, and after the treatment.
The traditional approach involves adaptations as a result of on-treatment changes in the patients anatomy and geometry (anatomy-based adaptations). Typically, a new treatment plan must be created on new planning CT images when a signicant change in the geometry is detected. The adaptation can be ofine (between fractions), online (immediately prior to a fraction), or even real-time (during an ongoing fraction) [180]. All three approaches may utilize various ML-empowered tools, such as auto-image registration, auto-contouring, or auto-planning, but no outcome model is necessary in such a framework. However, outcome modeling can be applied to predict ART eligibility based on numerous imaging and multiomic biomarkers [181, 182]. Other applications have demonstrated the promising role of longitudinal CT radiomics in evaluating the point of the tumors most signicant change to trigger the anatomy-based ART [183]. In a study by Forouzannezhad et al [184], a promising value of longitudinal CT, FDG-PET, and SPECT radiomics, to predict survival in NSCLC patients for risk-adaptive cancer therapy, was presented.
One of the possible strategies involves treatment optimization prior to the initial fraction. Suitable response models, predicting TCP and NTCP, can guide the dose schedule adaptation to optimize outcomes [185].
Another potential, and perhaps more common , approach involves treatment adaptation during t he treatment course, based on actual (mid-treatment) patient data and information. Traditionally, baseline imaging biomarkers (e.g. PET or MRI) would be evaluated before the treatment, followed by mid-treatment evaluation. The adaptation in dose scheduling or target volume denition would then be determined based on the comparison of these two datasets [186, 187].
As discussed in the previous section, the state-of-the-art response-based ART is KBR-ART. This framework is built on advanced AI models, comprising two essential components: (i) the outcome model for predicting clinical outcomes and (ii) the adaptation model optimizing the treatment, typically dose scheduling, based on the outcome model prediction. Both models must work in an effective cooperation, the RL model, where the reward function utilizes the outcome model. Most commonly, the adaptation is not proposed between each fraction, but at a certain point of the treatment, when it is possible to evaluate current outcomes and tailor the adaptation [176].
13-22
Artificial Intelligence in Adaptive Radiation Therapy
13.3.3 Sample use cases
13.3.3.1 Response-based dose prescription
In the study by Lou et al [185], a deep learning outcome model was applied to individualize the dose prescription in lung cancer patients. The outcome model had an autoencoder-like architecture to extract features from CT images, which were fed into another neural network to predict the probability of tumor local control. Prediction results were combined with other patients clinical data, in a regression model, to provide a recommendation for personalized dose prescription, that reduced tumor local failure probability below 5%.
Dose recommendations, provided by the model, had a wider range and were more continuous than the actual delivered doses. It also suggested dose de-escalation in a number of patients. Therefore, it provides a more exible approach to treatment personalization than regular clinical protocol. This framework allows personalizing dose prescription according to the risk prediction, which potentially changes the radiotherapy planning paradigm from binary decisions to response-informed adjustments.
13.3.3.2 Response-based dose schedule adaptation
A framework for dose fractionation adaptation was presented in a study by Tseng et al [177]. There were three complementary neural networks: GAN, DNN and DQN. The GAN was used to generate more training data samples as the original number of patients involved in this retrospective study was insufcient. Input characteristics involved clinical, genetic, and imaging features. The DNN was applied to reconstruct a radiotherapy articial environment (RAE) to simulate transition probabilities, for a given action, between different states (i.e. dose schedule modications). The last neural network (DQN) was responsible for the evaluation of possible adaptation strategies and recommendation of the optimal one, considering the treatment outcomes modeled by the RAE. This framework was designed to provide dose schedule adaptations at about 2/3 period of the treatment, based on dosimetric variables, clinical details, multiomics features, and PET radio­mics, acquired before and after about 1/3 and 2/3 of scheduled fractions. The presented results were benchmarked against real clinical dose adaptations, per­formed on patients enrolled in a dose-escalation clinical protocol.
A similar approach was adopted in the study by Niraula et al [175], which introduced the Adaptive Radiotherapy Clinical Decision Support (ARCliDS) software. They used GAN to generate more training samples, an articial RT environment (ARTE) to estimate treatment outcomes (TCP/NTCP), and an optimal decision maker (ODM), built on DQN, to recommend an optimal treatment adaptation. The ARTE combined a linear-quadratic-linear response model with graph neural networks (GNNs), developed by Niraula et al [174]. The input characteristics for TCP/NTCP modeling were dosimetric, clinical, multiomics, and PET-radiomics features. The system was trained and validated in two different treatment types for two different diseases (NSCLC and HCC). A diagram of the
13-23
Artificial Intelligence in Adaptive Radiation Therapy
Figure 13.9. (a) Diagram of response-adaptive radiotherapy using the ARCliDS system. A treatment plan is created in phase 0, the patients treatment response is evaluated in phase 1, treatment adaptation is planned at the end of phase 1, and it is executed in phase 2. (b) ARCliDS consists of two componentsARTE and ODM. In learning mode, ARTE is trained in supervised learning, and its outputs are used in the reinforcement learning of ODM. In operation mode, both components run simultaneously, providing outcome estimates and optimal dose adaptation. (Reproduced with permission from [
175]. Copyright 2023 Springer Nature.)
framework and description of the ARCliDS operation/learning modes are shown in gure 13.9.

13.4 Challenges and recommendations

Any AI deployment is generally associated with multiple challenges. However, radiotherapy AI response modeling and response-based adaptations have several
13-24
Artificial Intelligence in Adaptive Radiation Therapy
specic requirements, in addition to the general ones, of which all developers and researchers in this eld have to be aware.
A trustworthy and reliable response model in radiation oncology should always be based on a comprehensive understanding of the biological mechanisms behind disease specics, progression, and control. Only then can the developer decide whether the available data capture sufcient information for the desired prediction. It is the lack of understanding of the disease biology and, hence, insufcient data representation, which often limits the accuracy and trustworthiness of developed response models. For the same reason, it is sometimes also difcult to select the most indicative features for a given task. Input features are mostly selected based on various statistical methods, not considering their real biological/clinical importance for the modeled endpoint.
Another critical challenge for AI response modeling is the models architecture. There is no universal architecture that ts all possible scenarios, clinical endpoints, and the variability of input datasets. Therefore, the architecture is heavily inuenced by the developers skills and knowledge. The optimal architecture depends, for example, on the problem denition (classication/regression), data type (structured/ unstructured), number of training samples, class imbalance ratio, and so on.
One of the most common issues in radiotherapy outcome modeling regarding dataset quality is the imbalance of modeled endpoints. There are multiple methods to overcome this problem, for example, over-sampling, under-sampling, class weights, and the synthetic minority over-sampling technique (SMOTE) and its variants [188192]. However, the most effective appears to be generative models, which can be utilized to generate new synthetic data points of any class. The applicability of this concept to radiotherapy outcome modeling was recently showcased in a study by Dudas et al [159].
The ultimate goal of all response and treatment adaptation models is their translation into clinical practice. This can be accomplished only for models which are properly validated and provide a robust explainability. The clinically preferred validation process involves a prospective validation using a randomized controlled trial or clinical trial. The randomized controlled trial is commonly regarded as the optimal choice for prospective validation since it is expected to offer an unbiased tool for causal inferences. An alternative to prospective validation is retrospective validation, which evaluates models in past patients, who were not preselected for the study. There are several scenarios for retrospective validation, and some of them are described in the TRIPOD report [160]. The optimal approach for retrospective validation involves a separate validation dataset, ideally from different institutions. However, retrospective validation can never completely reduce all biases. For example, it may suffer from biases such as clinicians/patientstreatment preferences or changes in care quality over time. Therefore, this is a drive towards prospective validation via randomized clinical trials [193].
Existing challenges can be diminished if developers adhere to a suitable and standardized checklist [1] (see the example in gure 13.10.
13-25