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
difference 2.3 ± 0.1%
difference < 1%
0.6%
Artificial Intelligence in Adaptive Radiation Therapy
0.27%
± 0.79%
difference
< 1% for both photon
N/A
and proton plans
difference
< 1% (proton plan)
Gamma passing rate: >
+
95% at (1%, 1 mm)
for photon plan,
(Continued)
> 90% at (2%, 2 mm)
for
proton plan
0.5% (proton plan)
MAE (HU): 47.2 ± 11.0 Mean DVH metrics
validation/18% testing
(brain)
MAE (HU): 55.7 ± 9.4
Brain: 24, leave-one-out cross
Brain: T1w
validation
Pelvis: T2w
50.8 ± 15.5 (pelvis)
validation
Pelvis: 20, leave-one-out cross
MAE (HU): 34.1 ± 7.5 PTV V95% difference <
14, 25/11
MAE (HU): 72.87 ± 18.16 Mean DVH metrics
validation
MAE (HU): 51.32 ± 16.91 Mean DVH metrics
validation
MAE (HU): (82, 147)
testing
Gupta et al [74] U-Net 3 T in-phase Dixon T1w Brain: 47 training/13 testing MAE (HU): 17.6 ± 3.4 Mean target dose
Largent et al [28] GAN 3 T T2w Pelvis: 39, training/testing: 25/14, 25/
Kazemifar et al [85] GAN 1.5 T post-gadolinium T1w Brain: 77, 70% training/12%
Lei et al [39] CycleGAN
Liu et al [89] U-Net 1.5 T T1w Brain: 30 training/10 testing MAE (HU): 75 ± 23 PTV V95% difference
Liu et al [40, 42] CycleGAN 3 T/1.5 T T1w Liver: 21, leave-one-out cross
8-15
Liu et al [42] CycleGAN 1.5 T T2w Pelvis: 17, leave-one-out cross
Neppl et al [22] U-Net 1.5 T T1w Brain: 57 training/28 validation/4
CycleGAN 1.5 T T1w Brain: 50 MAE (HU): 54.55 ± 6.81 PTV D95 difference <
[90]
Olberg et al [44] GAN 0.35 T T1w Breast: 48 training/12 testing MAE (HU): 16.1 ± 3.5 PTV D95 difference <1%Shafai-Erfani et al
Key findings in
Key findings in image
dosimetry
quality
Artificial Intelligence in Adaptive Radiation Therapy
difference in
< 1%
difference
< 1%
MAE (HU): 75.7 ± 14.6 N/A
target = 1.3%
MAE (HU): 48.5 ± 6 Maximum dose
Table 8.3. (Continued )
Site, and # of patients in training/
testing
Author, year Network MR parameters
Wang et al [91] U-Net 1.5 T T2w Head and neck: 23 training/10 testing MAE (HU): 131 ± 24 N/A
Florkow et al [82] U-Net 3 T T1w Dixon Pelvis: 27, 3-fold cross validation MAE (HU): (33, 40) N/A
Koike et al [92] GAN T1w + T2w + FLAIR Brain: 15 MAE (HU): 108.1 ± 24.0 DVH metrics difference
Head and neck: 30 training/15 testing MAE (HU): 69.98 ± 12.02 Mean average dose
Qi et al [81] GAN T1w + T2w + contrast-
enhanced T1w + contrast-
enhanced T1w Dixon water
testing from one scanner
validation
Pelvis: 11 training from two scanner/8
Head and neck: 32, 8-fold cross
contrast T1w + T2w
scanners
Numbers in parentheses indicate minimum and maximum values.
N/A: not available, i.e. not explicitly indicated in the publication.
AE: Autoencoder.
*
Tie et al [83] GAN 1.5 T pre-contrast T1w + post-
Brou Boni et al [93] GAN 1.5 T and 3 T T2w from three
+
8-16
Artificial Intelligence in Adaptive Radiation Therapy
Table 8.4. Summary of studies on MR-based synthetic CT for PET attenuation correction. (Adapted from [1]. CC BY 4.0.)
Site, and # of patients in
Author, year NetworkMRparameters
training/testing Key findings in PET quality
Gong et al [84] U-Net Dixon and
ZTE
Jang et al [18] U-Net 3 T UTE Brain: 30 pre-training/6
Leynes et al [24] U-Net 3 T Dixon and
ZTE
Liu et al [17] U-Net 1.5 T T1w Brain: 30 training/10
Spuhler et al [27] U-Net 1.5 T T1w Brain: 44 training/11
Torrado-Carvajal
et al [23]
Blanc-Durand et al
[79]
Ladefoged et al [80] U-Net UTE Brain: 79 (pediatric), 4-
Arabi et al [94] GAN 3 T T1w Brain: 40, 2-fold cross
U-Net Dixon-VIBE Pelvis: 28 pairs from 19
U-Net ZTE Brain: 23 training/47
Brain: 14, leave-two-out Absolute bias < 3% among 8
training/8 testing
Pelvis: 26, 10 training/16
testing
testing
validation/11 testing
patients, 4-fold cross validation
testing
fold cross validation
validation
VOIs
Bias (%): 0.8 ± 0.8–1.1 ±
1.3 among 23 VOIs
RMSE (%): 2.68 among 30
bone lesions, 4.07 among 60 soft-tissue lesions
Bias (%): 3.2 ± 1.3–0.4 ±
0.8
Global bias (%): 0.49 ± 1.7
for 11C-WAY-100 635–
1.52 ± 0.73 for 11C­DASB
Bias (%): 0.27 ± 2.59 for fat
0.03 ± 2.98 for soft tissue
0.95 ± 5.09 for bone
Bias (%): 1.8 ± 1.9–1.7 ±
2.6 among 70 VOIs
Bias (%):
Absolute bias < 4% among
0.2–0.5 in 95%
CI
63 VOIs
In the majority of the studies, the MAE of the synthetic CT within the patient’s body typically falls within the range of 40–70 HU. Some reported results even approach the uncertainties observed in standard CT simulation. Specically, several studies highlight MAEs for soft tissue that are less than 40 HU [21, 28, 30, 7275], demonstrating relatively accurate intensity mapping in this region. However, due to the indistinguishable contrast of bone or air on MR images, the MAE for these tissues tends to exceed 100 HU, indicating higher discrepancies. Misalignment between CT and MR images in patient datasets emerges as a common source of error. This misalignment, particularly on bone structures, not only contributes to intensity mapping errors during training but also results in an overestimation of error during evaluation. This is because the error from misalignment registers as synthetic error in the assessment metrics. Notably, two studies reported signicantly higher MAE for the rectum (70 HU) compared to other soft tissues [28, 76]. This discrepancy may be attributed to mismatches in CT and MR imaging, potentially arising from variable lling of the rectum. Considering that the number of bone
8-17
Artificial Intelligence in Adaptive Radiation Therapy
pixels is considerably fewer than those of soft tissue, the training process may tend to map pixels to the low HU region during the prediction stage. Potential solutions to address these challenges could include assigning higher loss weights on bone structures or incorporating bone-only images during the training process [21].
In multiple studies, learning-based methods consistently outperform conventional methods, showcasing superior accuracy in generating synthetic CTs [16, 29, 73, 76]. This highlights the advantage of adopting a data-driven approach over traditional model-based methods. For instance, synthetic CTs generated by atlas-based methods were observed to be more susceptible to noise and registration errors, resulting in signicantly greater MAE compared to learning-based methods. Despite the advantages of learning-based methods, there are limitations to consider. The performance of these methods can be unpredictable when applied to datasets that signicantly differ from the training sets. Such differences may stem from unusual or abnormal anatomy, or images with degraded quality due to severe artifacts and noise. In contrast, atlas-based methods generate a weighted average of templates derived from prior knowledge. This characteristic makes them less prone to failure in handling unexpected or unusual cases, contributing to their robustness in scenarios with signicant variations in image quality [76].
The diverse datasets, training approaches, and testing strategies employed across these studies make the direct comparison of results challenging, precluding the determination of a single best methodology for all applications. However, some studies have conducted comparisons with competing methods using the same datasets, shedding light on relative advantages and limitations. For example, in a study involving fteen brain cancer patients, a GAN-based method demonstrated better preservation of detail and closer similarity to real CT with less noise when compared to an autoencoder-based method [30]. The GAN-based synthetic CT exhibited higher accuracy at the bone–air interface and in determining ne structures, with approximately 10HU less error by MAE. Another study comparing U-Net and GAN with different loss functions on 39 patients with prostate cancer revealed quantitative results indicating that U-Net methods had signicantly higher MAE than their GAN counterparts. Interestingly, the perceptual loss in both U-Net and GAN did not contribute to reducing MAE or provide benets for dose calculation accuracy [28]. A comparison between CycleGAN and GAN-based methods on patients with brain and prostate cancer demonstrated a signicant improvement in MAE with CycleGAN. CycleGAN also exhibited better visual results in terms of ne structural detail and contrast. Notably, CycleGAN results were less sensitive to local mismatches in the training CT/MR pairs, resulting in less blurry bone boundaries compared to GAN results [39]. Similar comparison results were reported in a study comparing CycleGAN and GAN on liver stereotactic body radiation therapy (SBRT) cases. While dosimetry comparison showed minimal difference, attributed to the insensitivity of volumetric modulated arc therapy (VMAT) plans to HU inaccuracy, CycleGAN exhibited improved MAE and visual results over GAN [40].
Among the reviewed studies, various MR sequences have been employed for synthetic CT generation, with the choice often dictated by their availability.
8-18
Artificial Intelligence in Adaptive Radiation Therapy
The optimal sequence yielding the best performance has not been conclusively determined. T1-weighted and T2-weighted sequences, being two of the most common general diagnostic MR sequences, are widely used due to their availability. These sequences enable models to be trained on relatively large datasets containing co-registered CT and T1- or T2-weighted MR images. T2-weighted images may be preferable to T1-weighted ones due to their intrinsically superior geometric accuracy within regions of signicant anatomic variability, such as the nasal cavity, and reduced chemical shift artifacts at fat and tissue boundaries. However, both T1- and T2-weighted MR images lack contrast for air and bone, which can impede the extraction of features corresponding to these structures in learning-based methods.
The two-point Dixon sequence, capable of separating water and fat, has been utilized in commercial PET/MR applications for segmentation [77, 78]. However, its limitation lies in poor bone contrast, resulting in the misclassication of bone as fat. To enhance bone contrast and facilitate feature extraction in learning-based methods, ultrashort echo time (UTE) and/or zero echo time (ZTE) MR sequences have been employed recently to generate positive image contrast from bone [17]. While studies by Ladefoged et al and Blanc-Durand et al demonstrated the feasibility of UTE and ZTE MR sequences using U-Net in PET/MR attenuation correction, respectively [79, 80], a direct comparison with conventional MR sequences under the same deep learning network is lacking. Therefore, the advantage of these specialized sequences has not been conclusively validated. Moreover, compared with conventional T1- or T2-weighted MR images, UTE/ ZTE MR images may have limited diagnostic value for soft tissue and longer acquisition times. This may potentially reduce their clinical utility, particularly in poorly tolerated, long-duration exams such as whole-body PET/MR.
Several studies have explored the use of multiple MR images with varying contrasts as training inputs to enhance the overall predictive power and accuracy of synthetic CT generation. Qi et al proposed a four-channel input comprising T1, T2, contrast-enhanced T1, and contrast-enhanced T1 Dixon water images. The results from the four-channel input demonstrated lower mean absolute error (MAE) compared to results from fewer channels, highlighting the potential benets of incorporating diverse contrast information [81]. Florkow et al investigated single­and multi-channel inputs using magnitude MR images and Dixon-reconstructed water and fat images obtained from a single T1 multi-echo gradient-echo acquisition [82]. Their ndings indicated that multi-channel input can improve synthetic CT generation over single-channel input, with the Dixon sequence input outperforming other congurations. Tie et al employed T2 and pre- and post-contrast T1 MR images in a multi-channel, multi-path architecture, demonstrating additional improvement over multi-channel single-path and single-channel results [83]. Combining UTE or ZTE sequences with Dixon sequences, which provide contrast for bone against air and fat against soft tissue, respectively, has been considered an attractive combination [ 24, 84]. Leynes et al showed that synthetic CT using both ZTE and Dixon MR sequences has less error than using Dixon alone, showcasing the potential benets of combining these contrast sources [24]. While the resulting improvement in image quality has been validated, the necessity of performing
8-19
Artificial Intelligence in Adaptive Radiation Therapy
additional MR sequences for synthetic CT generation requires further study in specic applications to justify the associated costs and acquisition time.
In the reviewed studies, CT and MR images in the training datasets were acquired separately on different machines, necessitating image registration between the CT and MR images to create CT-MR pairs for training. The registration error is generally minimal at the level of the brain but may be more signicant within the pelvis, owing to variable bladder and rectum lling, and in the abdomen, due to variations introduced by respiratory motion and peristalsis. Methods such as U-Net and GAN-based approaches can be susceptible to registration errors, particularly when utilizing a pixel-to-pixel loss function. These errors can be exacerbated by physiological motion, making accurate registration challenging. To address this issue, Kazemifar et al proposed a potential solution using mutual information as the loss function in the GAN generator. This approach aims to bypass the registration step during training, potentially mitigating the impact of registration errors on the performance of the model [85]. CycleGAN-based methods, developed for unpaired image-to-image translation, exhibit greater robustness to registration errors. This is attributed to the role of the cycle consistency loss, which enforces structural consistency between the original and cycle-generated images. For instance, in the context of synthetic CT generation from MR images, the cycle consistency loss ensures that a cycle MRI generated from synthetic CT remains similar to the original MRI. This characteristic makes CycleGAN-based methods more resilient to registration errors, contributing to their effectiveness in scenarios where accurate image registration is challenging [19, 33, 35, 86].
8.4.2 Dose calculation in MR-only radiation therapy
In studies with applications in radiation therapy, many have evaluated the dosimetric accuracy of synthetic CT by calculating the radiation treatment dose from the original treatment plan and comparing it against ground truth CT simulation imaging. It has been observed that the dose difference is approximately 1%, which is relatively small compared to typical total dose delivery uncertainties over an entire treatment course (5%). For reference, in the bulk-density assignment method, Lee et al observed that the differences between the dose of CRT plans on bulk-density, when compared to CT, were less than 2% [3]. Similarly, Jonsson et al reported a comparable result, noting that the maximum difference in monitor units (MU) required to reach the prescribed dose was 1.6% [4]. The improvement in dosimetric accuracy provided by deep learning-based methods in radiation therapy, when compared to image accuracy, is relatively small and may lack clinical relevance [73, 76]. One potential reason for this phenomenon is that dose calculation on photon plans tends to be forgiving to image inaccuracy, particularly within homogeneous regions such as the brain. In VMAT, the contribution to dosimetric error from random image inaccuracy also tends to cancel out within an arc. However, the small dosimetric improvement observed may be of signicance in scenarios such as stereotactic radiosurgery (SRS) and stereotactic body radiation therapy (SBRT), where small volumes are treated to very high doses. In such cases,
8-20
Artificial Intelligence in Adaptive Radiation Therapy
signicant dosimetric errors may arise from otherwise negligible errors in CT synthesis, particularly in the region surrounding the target volume [95]. These ndings underscore the importance of considering the clinical context and the specic treatment scenario when assessing the impact of synthetic CT accuracy on dosimetry in radiation therapy applications.
Studies have also assessed the use of synthetic CT in the context of proton therapy for various cancers, including prostate, liver, and brain cancer [41, 42, 90]. Proton beams, unlike photon beams, exhibit a sharp dose gradient (Bragg peak) at the distal end of the beam, allowing for highly conformal dose delivery to the target by superimposing proton beams from several angles. Any inaccuracies in HU along the beam path on the planning CT can lead to a shift in the highly conformal high-dose area. This shift may result in the tumor being substantially under-dosed or the organs at risk being over-dosed [96]. In studies such as the one by Liu et al most of the dose differences resulting from the use of synthetic CT were observed at the distal end of the proton beam [42]. Liu et al reported that the largest and mean absolute range differences were 0.56 and 0.19 cm among their 21 liver cancer patients, and
0.75 and 0.23 cm among 17 prostate cancer patients, respectively [41, 42]. These ndings emphasize the critical importance of accurate synthetic CT generation in proton therapy, where precision in dose delivery is crucial due to the unique characteristics of proton beams.
In addition to dosimetric accuracy for treatment planning, the evaluation of synthetic CT imaging must also consider geometric delity for treatment set-up. However, studies specically focusing on synthetic CT positioning accuracy are limited. Fu et al conducted patient alignment testing by rigidly aligning synthetic CT and real CT to the CBCT acquired during the delivery of the rst fraction of a fractionated radiotherapy treatment course [21]. The average translation vector distance and absolute Euler angle difference between the two alignments were found to be less than 0.6 mm and 0.5°, respectively. Gupta et al performed a similar study and reported that the translation difference was less than 0.7 mm in one direction [74]. Although studies have addressed alignment with CBCT, the alignment between the digitally reconstructed radiograph (DRR) derived from the synthetic CT and on­board kilovolt (kV) imaging of the patient is also clinically important. However, no studies on DRR alignment accuracy were found in the reviewed literature. It is worth noting that the geometric accuracy of synthetic CT is inuenced not only by the synthetic methods employed but also by the geometric distortion on MR images caused by magnetic eld inhomogeneity, as well as subject-induced susceptibility and chemical shift. Therefore, methods to mitigate MR distortion are crucial for improving synthetic CT accuracy in patient positioning, contributing to the overall success of radiotherapy treatment set-up.
8.4.3 PET attenuation correction
In studies focused on PET attenuation correction, the evaluation has primarily centered around the bias introduced in PET quantication due to synthetic CT errors. While it is challenging to dene a specic error tolerance that signicantly
8-21
Artificial Intelligence in Adaptive Radiation Therapy
impacts clinical decision-making, a general consensus is that quantitative errors of 10% or less typically do not have a substantial impact on decisions in diagnostic imaging [15]. A thresholding on MR images of UTE sequences resulted into an average error of 5% in brain PET images [12]. Kops and Herzog demonstrated that the segmentation-based and the registration-based methods proposed by them have similar performance in PET image reconstruction [14]. A thresholding on MR images of UTE sequences resulted in an average error of 5% in brain PET images [12]. Kops and Herzog demonstrated that the segmentation-based and the registra­tion-based methods proposed by them have similar performance in PET image reconstruction [14]. Most of the proposed deep learning methods in the reviewed studies met this criterion based on the average relative bias reported. However, it is essential to note that due to variation among study subjects, the bias in some volumes-of-interest (VOIs) may exceed 10% for certain patients [24, 79]. This emphasizes the importance of considering both the mean and standard deviation of the bias when interpreting results, as proposed methods may exhibit poor local performance affecting specic patients. Reporting alternative results that list or plot all data points, or at least their range, could provide a more comprehensive understanding of the proposed methodsperformance.
Bone accuracy on synthetic CT is crucial for PET attenuation correction since bone has the highest capacity for attenuation due to its high density and atomic number. Unlike applications in radiation therapy, the bias and geometric accuracy of bone on synthetic CT are more frequently evaluated for PET attenuation correction. Several studies have demonstrated that improved accuracy of bone representation in CT synthesis leads to more globally accurate PET [23, 79, 84, 94]. In the reviewed studies, PET attenuation correction by conventional CT synthesis methods exhibited an average bias of about 5% among selected VOIs. In contrast, learning-based methods reduced the bias to around 2%, highlighting the signicant improvements achieved in PET accuracy with more accurate synthetic CT images generated by these methods [17, 18, 23, 24, 84].
8.4.4 Image registration
In addition to its applications in radiation treatment planning and PET attenuation correction, MR-based CT synthesis has demonstrated promise in facilitating inter­modality image registration. Direct registration between CT and MR images is challenging due to disparate image contrast, and this challenge is further amplied in deformable registration, where signicant geometric distortion is allowed. McKenzie et al proposed a CycleGAN-based method to synthesize CT images, utilizing the synthetic CT to replace MR imaging in MR-CT registration in the head and neck [97]. By doing so, they transformed an inter-modality registration problem into an intra-modality one. As summarized in table 8.5, their ndings revealed that, using the same deformable registration algorithm, the average landmark error decreased from 9.8 ± 3.1 mm in direct MR-CT registration to 6.0 ± 2.1 mm when using synthetic CT as a bridge. Similar positive results were reported in the inverse CT-MR registration task.
8-22
Artificial Intelligence in Adaptive Radiation Therapy
Table 8.5. Summary of study on MR-based synthetic CT for registration. (Adapted from [1]. CC BY 4.0.)
Site, and # of patients in
Author, year NetworkMRparameters
McKenzie et al [97] CycleGAN 0.35 T Head and neck: 25, 5-fold
training/testing
cross validation
Key findings in registration accuracy
Landmark error (mm):
6.0 ± 2.1 (MR-to-CT)
6.6 ± 2.0 (CT-to-MR)

8.5 Discussion and outlook

Recent years have seen a surge in the utilization of deep learning within the realm of medical imaging. Cutting-edge networks and techniques borrowed from computer vision have been adapted to cater to specic clinical tasks in radiology and radiation oncology. This chapter reviews the emerging and active eld of CT synthesis, with most of the studies covered being published within the last three years. With ongoing advancements in both articial intelligence and computing hardware, it is antici­pated that more advanced learning-based methods will further enhance the clinical workow with novel applications. While the reviewed literature showcases the success of deep learning-based image synthesis in various applications, there are still some open questions that need addressing in future studies.
The incorporation of novel network architectures, including transformers and diffusion models, holds great promise for advancing the eld of CT synthesis. The transformer architecture, renowned for its success in natural language processing and image recognition, may offer enhanced capabilities in capturing long-range dependencies and contextual information within medical images. The attention mechanism in transformers enables the model to focus on relevant image regions, potentially improving the synthesis accuracy, particularly in complex anatomical structures. Similarly, the diffusion model, such as the DDPM, has emerged as a powerful tool for deep generative tasks. Its unique two-stage approach involving noise addition and subsequent denoising offers stability during training, making it less susceptible to issues such as mode collapse and hyperparameter sensitivity. As research in this area progresses, the application of diffusion models could contribute to more robust and accurate CT synthesis, addressing challenges faced by current deep learning-based methods.
The selection between 2D and 3D models for CT synthesis is a pivotal decision that hinges on the specic demands and constraints of the application. 2D models exhibit advantages in computational efciency and training data availability, making them suitable for scenarios with limited resources and large datasets. However, challenges arise in their ability to capture 3D context and potential slice discontinuities. On the other hand, 3D models inherently provide spatial context and more homogeneous synthesis but demand greater computational resources and extensive training data. Fu et al compared the performance of 2D and 3D models using the same U-Net implementation, nding that 3D-generated synthetic CT
8-23
Artificial Intelligence in Adaptive Radiation Therapy
exhibited smaller MAE and more accurate bone regions [21]. Hybrid approaches, combining 3D patches or multiple adjacent slices, offer a compromise [98]. The ongoing development of techniques that optimize both 2D and 3D models may provide a balanced solution, ensuring that the choice aligns with the unique requirements of medical imaging tasks, such as CT synthesis, where understanding volumetric relationships is critical for accurate clinical applications.
The reviewed studies underscore the superiority of learning-based methods over conventional approaches in terms of performance and clinical utility. Learning­based methods consistently surpass conventional ones by producing synthetic images that closely resemble real images and exhibit superior quantitative metrics. While the training process for learning-based methods demands hours to days, the application of a trained model to new patients enables the rapid generation of synthetic images within seconds to minutes. In contrast, conventional methods display a broad spectrum of run times due to diverse methodologies, with iterative approaches such as compressed sensing (CS) proving less favorable due to substantial time and computational resource requirements.
While learning-based methods have demonstrated clear advantages, it is crucial to acknowledge the potential unpredictability of their performance when dealing with input images during production that signicantly differ from the training images. Many reviewed studies tend to exclude unusual cases, but the clinical reality may present scenarios that deviate from the norm. Instances such as hip prostheses, causing severe artifacts on both CT and MR images, could impact the application of learning-based methods, and understanding such effects is essential. Unusual cases, ranging from medical implants introducing artifacts to challenges posed by obesity and anatomic deformities, may arise in various imaging modalities, warranting further investigation to ensure the robustness and reliability of learning-based models in diverse clinical scenarios.
Before integrating learning-based models into the clinical workow, addressing several challenges is paramount. To accommodate potentially unpredictable synthetic images arising from non-compliance with imaging protocols in the training data or unexpected anatomic variations, the implementation of additional quality assurance (QA) steps becomes essential in clinical practice. QA procedures would be designed to routinely assess or verify the consistency of model performance, either through periodic checks or after upgrades, involving re-training the network with additional patient datasets. This approach ensures the reliability of synthetic image quality across a range of cases in diverse clinical scenarios.

8.6 Summary

In recent years, the increasing integration of deep learning into medical imaging has been notable. Borrowing from computer vision, advanced techniques of AI are now being tailored for clinical use in radiology and radiation oncology. Adaptive radiation therapy is an emerging concept that involves complex imaging operations. AI with its superior ability in image style transferring can facilitate the adaptive radiation therapy workow by synthesizing CT images from CBCT or/and MRI
8-24