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
12.2.3 On-treatment QA
A patient undergoing treatment is monitored by physics, typically as weekly chart review. Weekly chart checks involve monitoring of treatment records and docu­mentation which are recorded in the oncology information system (OIS) [91]. The literature is scant on the use of AI to assist in that scope of weekly chart review, apart from the proposed software solutions [92, 93]. Although it is the physicians responsibility, the review of daily imaging and registration accuracy and learning the dose implication of positioning error or anatomical changes is a task that can benet from AI. Simple models were developed to detect gross alignment error in vertebral bodies [94, 95]. The same group implemented a CNN-based model that can detect not only vertebral body misalignment [96] but also soft tissue changes [97].
In vivo dosimetry is another form of on-treatment monitoring. EPID-based in vivo dosimetry, which captures exit radiation through a patient during treatment (transit dose), has drawn much interest due to its convenience and high spatial resolution. The EPID images can be compared to the uence from the treatment plan in a similar manner to pre-treatment PSQA. Wolfs et al built a hidden Markov model [98] and CNN-based [99] prediction system that can detect anatomical or positional change by analysing transit EPID images. Alternatively, EPID can be backprojected into 3D dose in patient geometry, which enables direct comparison of target or OAR dose against planned dose [100]. DL has been applied to enhance the accuracy of the reconstructed dose for an MR-linac to correct for the effect of the magnetic eld [101]. The same group modeled the generic deviation in a reconstructed in vivo dose using DL in order to increase the sensitivity of in vivo dosimetry in patient-related sources of deviations [102](figure 12.8).

12.3 Treatment delivery systems and instruments

12.3.1 Machine commissioning
Zhao et al [103] introduced a machine-learning-based approach to model linac beam data, streamlining the processes of linac commissioning and QA. The model, trained with 43 Varian TrueBeam beam data, sets encompassing PDDs and proles across various energies and eld sizes from different institutions. Figure 12.9 shows the workow for model building.
Using a 10 × 10 cm developed for predicting beam specic PDDs and proles for different eld sizes. The predictions for PDDs exhibited a mean absolute percent relative error (%RE) ranging from 0.19% to 0.35% across various beam energies, with a maximum mean absolute %RE of 0.93%. In prole prediction, the mean absolute %RE was within the range of 0.66%–0.93%, and the maximum absolute %RE was 3.76%. Figure 12.10 shows the comparison between the ground truth and the predicted PDDs and proles of 4 × 4 cm potential in simplifying the linac commissioning procedure, offering time and resource efciency while enhancing the accuracy of the commissioning process. It is particularly promising for its ability to address uncertainties, with the largest
2
eld as the input, a multivariate regression model was
2
and 30 × 30 cm2fields. Notably, this method showed
12-10
Artificial Intelligence in Adaptive Radiation Therapy
Figure 12.8. Example of anatomical changes that can be detected by 3D-U-Net based in vivo dose reconstruction prediction. Column (a): overlay between simulation and cone-beam CT demonstrating anatomical changes. Column (b): dose agreement in a 3D gamma map between an in vivo reconstructed dose and TPS. Column (c): U-Net reconstructed gamma map, highlighting patient-specic dose deviations. Column (d): U-Net reconstructed gamma map factoring out patient-specic changes, showing only generic (TPS or detector related) dose deviation. (Reproduced from [ Copyright 2023 American Association of Physicists in Medicine.)
102] with permission from John Wiley & Sons.
observed in the build-up region for PDD predictions and at the eld penumbra for prole predictions.
Liu et al [104] from the same research group rened the data acquisition process for linac beam data through the application of implicit neural representation (NeRP) learning, as illustrated in gure 12.11. This aimed to enhance the accuracy of beam data collection verication and streamline the linac commissioning and QA procedure.
12-11
Artificial Intelligence in Adaptive Radiation Therapy
Figure 12.9. Workow for model training and prediction of PDD and proles. (Reproduced with permission from [
103]. Copyright 2020 Elsevier.)
The authors incorporated prior knowledge of beam data into a multilayer perceptron network, achieved by learning the NeRP from a vendor-provided goldenbeam dataset. This network underwent training to align with clinical beam data collected at a specic eld size. Subsequently, it demonstrated the capability to predict beam data accurately for other eld sizes. To assess prediction accuracy, the authors compared the network-predicted beam data with measure­ments obtained from water tanks across 14 clinical linacs. They found that the linac beam data predicted by the model exhibited strong agreement with water tank measurements (averagely > 95% passing rates at 1%/1 mm criteria and < 0.6% mean absolute errors). Figure 12.12 shows beam prole predictions against the ground truth data. Moreover, the model unveiled instances of measurement errors by identifying inconsistent beam predictions when trained with correct versus erroneous data samples. These discrepancies were characterized by a GPR < 90%. It is concluded that the model veries beam data collection accuracy and holds promise of simplifying commissioning and QA processes by minimizing the number of required measurements without compromising the quality of medical physics service.
12-12
Artificial Intelligence in Adaptive Radiation Therapy
Figure 12.10. Example of PDD (A)–(C) and prole (D)(F) prediction of a 4 × 4 cm2field (yellow dots) and a 30 × 30 cm (ground truth) are depicted as black lines. (Reproduced with permission from [
Figure 12.11. Workow of NeRP learning for linac beam modeling and prediction. (Reproduced from [104] with permission from John Wiley & Sons. Copyright 2023 American Association of Physicists in Medicine.)
2
eld (blue dots) with 10 × 10 cm2field as input (red line). The measured PDDs and proles
103]. Copyright 2020 Elsevier.)
Wagner et al [105] introduced a machine-learning approach to expedite the modeling process for M6 CyberKnife integrated in Moderato. A machine-learning algorithm was trained to nd electron beam parameters for other M6 devices. The algorithm simulated dose curves with varying spot size and energy, optimizing its performance through cross-validation, and validating its accuracy with measure­ments from other institutions equipped with M6 CyberKnife devices. The agreement
12-13
Artificial Intelligence in Adaptive Radiation Therapy
Figure 12.12. Example beam prole predictions for 6 MV and 6FFF beam at 30 × 30 cm2field size. Enhanced consistency in water tank measurements was achieved by tting the prior-embedded network to sparse beam data, indicated by green circles, rather than using direct goldenbeam data. (Reproduced from [ permission from John Wiley & Sons. Copyright 2023 American Association of Physicists in Medicine.)
104] with
in the Monte Carlo model was achieved for a monoenergetic electron beam of 6.75 MeV with a Gaussian spatial distribution of 2.4 mm full-width-at-half-maximum (FWHM). Clinical plan dose distributions from Moderato exhibited an agreement within 2% with the TPS, and lm measurements further corroborated the precision of the model. During cross-validation of the prediction algorithm, minimal mean absolute errors of 0.1 MeV and 0.3 mm were observed for beam energy and spot size, respectively. The prediction agreements were within 3% with measurements, except for one device where differences up to 6% were detected. This approach can expedite the modeling of new machines within Monte Carlo systems, offering efciency and reliability in the intricate process of medical physics modeling.
12.3.2 Machine QA
Numerous investigations have explored the diverse applications of machine learning in linac QA [106, 107]. These studies have delved into various avenues, encompass­ing models constructed from beam data commissioning, as detailed in the preceding section, to models derived from delivery log les [108110]. Additionally, machine­learning models have been built using proton elds [111, 112], addressing image artifacts [113], and implementing automated QA through electronic portal imaging
12-14
Artificial Intelligence in Adaptive Radiation Therapy
Figure 12.13. Results from SPC using symmetry data from open eld beam (left) and EDW data (right). (Reproduced with permission from [
117]. Copyright 2023 IOP Publishing Ltd.)
device (EPID) images [114]. This multifaceted exploration underscores the versa­tility of machine learning across various aspects of linac QA, highlighting its potential to enhance efciency and accuracy in QA processes.
Chan et al [115] utilized ve years of daily linac QA data for visual analysis and correlation studies among dosimetric parameters and the researchers gained insights into the intricacies of the data. Subsequently, they developed a convolutional neural network (time-series modeling) to predict beam symmetry using the same daily QA dataset [116]. The same research group further utilized the daily QA data, but to include both open elds and enhanced dynamic wedge (EDW) measurements [117]. They employed statical process control and autoregressive integrated moving average modeling to predict linac target failure. The EDW mechanism, character­ized by nonuniform magni cation factors within its wedge-directed beam proles, played a pivotal role in their analysis. This nonuniformity introduced sensitivity to changing beam properties induced by a degrading target. Figure 12.13 illustrates two occurrences of target failures that can be effectively predicted from the daily symmetry data. The comprehensive approach contributes insights into the predic­tion of linac performance.
12.3.3 Dosimetry tool QA
Chang et al [118] presented a deep-learning hierarchical neural network (HNN) method to calibrate the EBT3 lm with better calibration accuracy than the conventional R-NOD method. They used the Keras functional application program interface to build an HNN, with the inputs of net optical densities, pixel values, and inverse transmittances to reveal the delivered dose and train the neural network with deep learning. About the aging effect, the percentage error of the HNN method is within 4% and proved to be unaffected, while the averaged percentage error of the conventional R-NOD method is about 6.8% (gure 12.14). This new technique can be improved by updating the new calibration data into the HNN training system whenever physicists perform the recalibration. Based on collecting calibration data
12-15
Artificial Intelligence in Adaptive Radiation Therapy
Figure 12.14. Percentage differences between the calculated dose and the delivered dose for the verication test of Lot A and Lot B lms. (Reproduced from [
118]. CC BY 4.0.)
with the HNN method, physicists could require less calibration time and reduce lm usage.
Zhuang et al [119] described the development of an ANN approach for processing EBT3 lms from various batches without requiring specic calibration for each batch. Utilizing PyTorch, researchers constructed a feed-forward ANN model that transforms pixel values from scanned images across different batches into the corresponding absorbed dose. To facilitate this, lms exposed to x-ray doses from different batches were scanned in transmission mode using an Epson 11000XL scanner, serving both for model training and validation. The dose map generated by the TPS was employed as the target output for the ANN model. To assess the methods effectiveness and its adaptability, a cross-validation study was conducted. The ANN model, once trained, was used to convert scanned images into dose maps, demonstrating a high level of consistency with the dose maps calculated by TPS. For lms exposed using the sliding window technique, the mean square errors (MSEs) were below 16.0 cGy for the training batches and under 18 cGy for the testing batches. In the case of patient IMRT lms, the γ (3%, 3 mm) for comparison between the dose maps from the ANN and TPS exceeded 97.5% for the training set and 97.0% for the testing set. These results indicate the potential of this method to accurately convert pixel values to absorbed doses for EBT3 lms without needing batch-specic calibrations, suggesting its applicability to other scenarios.
Avanzo et al [120] developed a machine-learning model to forecast skin dose from targeted intraoperative (TARGIT) treatment, facilitating the timely implementation of strategies to mitigate the risk of excessive skin dose. The process involved feature selection of predictors of in vivo skin dose, followed by the training of various machine-learning models using in vivo dosimetry results. The evaluation was conducted through tenfold cross-validation, with rankings based on both roots mean square error (RMSE) and adjusted correlation coefcient of true versus
12-16
Artificial Intelligence in Adaptive Radiation Therapy
predicted values (adj-R2). The identied predictors strongly correlated with in vivo dosimetry, including factors such as the distance of skin from source, depth-dose in water at the depth of the applicator in the breast, utilization of a replacement source, and irradiation time. Among the models, support vector regression (SVR) emerged as the most effective, achieving an RMSE of 0.746 (95% condence intervals, 0.737,
0.756) and an adj-R2 of 0.481 (95% CI 0.468, 0.494) during the tenfold cross­validation. This SVR model, trained on in vivo dosimetry results, holds signicant practicality. The authors concluded that it can be employed to predict skin dose during the patient set-up for TARGIT, enabling the timely adoption of strategies to prevent excessive skin dose, thus enhancing the overall safety and efcacy of the procedure.

12.4 Summary

Quality assurance plays a crucial role in ensuring the accuracy of radiotherapy planning, delivery, and instrumentation within the clinical workow. This critical process places a substantial demand on clinical physicistsresources. The integration of AI into QA brings forth a twofold enhancement. First, AI serves to accentuate components with heightened susceptibility to failure, enabling physicists to allocate their attention more effectively towards those vulnerabilities. By pinpointing these areas of concern, AI augments the specicity of QA process, facilitating a more targeted and efcient utilization of resources. Second, AI possesses the capability to discover intricate system errors that may elude human detection due to their complexity. The analysis of various inuencing factors is a forte of AI, enabling it to discern anomalies that might otherwise remain hidden, reinforcing the overall reliability of the QA process. In summary, the incorporation of AI into QA efforts contributes to an elevated level of efciency and standardization within the QA program, advancing the precision, safety, and consistency of radiotherapy practices.

References

[1] ASTRO 201 Safety is no accident Guide American Society for Radiation Oncology https://
astro.org/Patient-Care-and-Research/Patient-Safety/Safety-is-no-Accident
[2] McIntosh C, Svistoun I and Purdie T G 2013 Groupwise conditional random forests for
automatic shape classication and contour quality assessment in radiotherapy planning
IEEE Trans. Med. Imaging
[3] Chen H C et al 2015 Automated contouring error detection based on supervised geometric
attribute distributio n models for radiation therapy: a general strategy Med. Phys.
1048–59
[4] Hui C B et al 2018 Quality assurance tool for organ at risk delineation in radiation therapy
using a parametric statistical approach Med. Phys.
[5] Zhang Y, Plautz T E, Hao Y, Kinchen C and Li X A 2019 Texture-based, automatic
contour validation for online adaptive replanning: a feasibility study on abdominal organs
Med. Phys.
[6] Cardenas C E et al 2018 Deep learning algorithm for auto-delineation of high-risk
oropharyngeal clinical target volumes with built-in Dice similarity coefcient parameter optimization function Int. J. Radiat. Oncol. Biol. Phys.
46 4010–20
32 1043–57
45 2089–96
101 468–78
12-17
42
Artificial Intelligence in Adaptive Radiation Therapy
[7] Rhee D J et al 2019 Automatic detection of contouring errors using convolutional neural
networks Med. Phys.
[8] Gooding M J, Boukerroui D, Vasquez Osorio E, Monshouwer R and Brunenberg E 2022
Multicenter comparison of measures for quantitative evaluation of contouring in radio­therapy Phys. Imaging Radiat. Oncol.
[9] Rhee D J et al 2022 Automatic contouring QA method using a deep learning-based
autocontouring system J. Appl. Clin. Med. Phys.
[10] Cha E et al 2021 Clinical implementation of deep learning contour autosegmentation for
prostate radiotherapy Radiother. Oncol.
[11] Gooding M J et al 2018 Comparative evaluation of autocontouring in clinical practice: a
practical method using the turing test Med. Phys.
[12] Yang J, Veeraraghavan H, van Elmpt W, Dekker A, Gooding M and Sharp G 2020 CT
images with expert manual contours of thoracic cancer for benchmarking auto-segmenta­tion accuracy Med. Phys.
[13] Allozi R et al 2010 Tools for consensus analysis of expertscontours for radiotherapy
structure denitions Radiother. Oncol.
[14] Sherer M V et al 2021 Metrics to evaluate the performance of auto-segmentation for
radiation treatment planning: a critical review Radiother. Oncol.
[15] Brock K K, Mutic S, McNutt T R, Li H and Kessler M L 2017 Use of image registration
and fusion algorithms and techniques in radiotherapy: report of the AAPM Radiation Therapy Committee Task Group No. 132 Med. Phys.
[16] Rong Y et al 2021 Rigid and deformable image registration for radiation therapy: a self-
study evaluation guide for NRG oncology clinical trial participation Pract. Radiat. Oncol.
11 282–98
[17] Paganelli C, Meschini G, Molinelli S, Riboldi M and Baroni G 2018 Patient-specic
validation of deformable image registration in radiation therapy: overview and caveats
Med. Phys.
[18] Nenoff L et al 2023 Review and recommendations on deformable image registration
uncertainties for radiotherapy applications Phys. Med. Biol.
[19] Hussein M, Akintonde A, McClelland J, Speight R and Clark C H 2021 Clinical
use, challenges, and barriers to implementation of deformable image registration in radiotherapythe need for guidance and QA tools Br. J. Radiol.
[20] Yuen J et al 2020 An international survey on the clinical use of rigid and deformable image
registration in radiotherapy J. Appl. Clin. Med. Phys.
[21] Yang D et al 2017 A method to detect landmark pairs accurately between intra-patient
volumetric medical images Med. Phys.
[22] Paganelli C, Peroni M, Baroni G and Riboldi M 2013 Quantication of organ motion
based on an adaptive image-based scale invariant feature method Med. Phys.
[23] Han D, Gao Y, Wu G, Yap P T and Shen D 2015 Robust anatomical landmark detection
with application to MR brain image registration Comput. Med. Imaging Graph.
[24] Grewal M, Wiersma J, Westerveld H, Bosman P A N and Alderliesten T 2023 Automatic
landmark correspondence detection in medical images with an application to deformable image registration J. Med. Imaging
[25] Bender E T and Tomé W A 2009 The utilization of consistency metrics for error analysis in
deformable image registration Phys. Med. Biol.
45 e908–e22
46 5086–97
24 152–8
23 e13647
159 1–7
45 5105–15
47 3250–5
97 572–8
160 185–91
44 e43–76
68 24TR01
94 20210001
21 10–24
44 5859–72
40 111701
46 277–90
10 014007
54 5561–77
12-18
Artificial Intelligence in Adaptive Radiation Therapy
[26] Varadhan R, Karangelis G, Krishnan K and Hui S 2013 A framework for deformable
image registration validation in radiotherapy clinical applications J. Appl. Clin. Med. Phys.
14 4066
[27] Saleh Z H et al 2014 The distance discordance metric: a novel approach to quantifying
spatial uncertainties in intra- and inter-patient deformable image registration Phys. Med.
Biol.
59 733–46
[28] Neylon J, Min Y, Low D A and Santhanam A 2017 A neural network approach for fast,
automated quantication of DIR performance Med. Phys.
[29] Eppenhof K A J and Pluim J P W 2018 Error estimation of deformable image
registration of pulmonary CT scans using convolutional neural networks J. Med.
Imaging
[30] Galib S M, Lee H K, Guy C L, Riblett M J and Hugo G D 2020 A fast and scalable
method for quality assurance of deformable image registration on lung CT scans using convolutional neural networks Med. Phys.
[31] Smolders A, Lomax A, Weber D C and Albertini F 2023 Deep learning based uncertainty
prediction of deformable image registration for contour propagation and dose accumu­lation in online adaptive radiotherapy Phys. Med. Biol.
[32] Ge Y and Wu Q J 2019 Knowledge-based planning for intensity-modulated radiation
therapy: a review of data-driven approaches Med. Phys.
[33] Tol J P, Dahele M, Delaney A R, Slotman B J and Verbakel W F 2015 Can knowledge-
based DVH predictions be used for automated, individualized quality assurance of radiotherapy treatment plans? Radiat. Oncol.
[34] Cao W et al 2022 Knowledge-based planning for the radiation therapy treatment plan
quality assurance for patients with head and neck cancer J. Appl. Clin. Med. Phys.
e13614
[35] Stanhope C et al 2015 Utilizing knowledge from prior plans in the evaluation of quality
assurance Phys. Med. Biol.
[36] Fan J, Wang J, Chen Z, Hu C, Zhang Z and Hu W 2019 Automatic treatment planning
based on three-dimensional dose distribution predicted from deep learning technique Med.
Phys.
[37] Nguyen D et al 2019 3D radiotherapy dose prediction on head and neck cancer patients
with a hierarchically densely connected U-Net deep learning architecture Phys. Med. Biol.
64 065020
[38] Gronberg M P et al 2023 Deep learning-based dose prediction for automated, individu-
alized quality assurance of head and neck radiation therapy plans Pract. Radiat. Oncol.
e282–e91
[39] Ford E et al 2020 Strategies for effective physics plan and chart review in radiation therapy:
report of AAPM Task Group 275 Med. Phys.
[40] Ford E C, Terezakis S, Souranis A, Harris K, Gay H and Mutic S 2012 Quality control
quantication (QCQ): a tool to measure the value of quality control checks in radiation oncology Int. J. Radiat. Oncol. Biol. Phys.
[41] Gopan O, Zeng J, Novak A, Nyot M and Ford E 2016 The effectiveness of pretreatment
physics plan review for detecting errors in radiation therapy Med. Phys.
[42] Clark B G, Brown R J, Ploquin J and Dunscombe P 2013 Patient safety improvements in
radiation treatment through 5 years of incident learning Pract. Radiat. Oncol.
5 024003
47 99–109
10 234
60 4873–91
46 370–81
47 e236–e72
84 e263–9
44 4126–38
68 245027
46 2760–75
23
13
43 5181
3 157–63
12-19