Applying aperture-based intensity map in automated plan review of volumetric modulated arc therapy for lung cancer patients

Introduction

Radiotherapy is commonly used to treat cancer, and is an indispensable component in the treatment of most types of cancer. Presently, approximately 60% of cancer patients receive radiotherapy for definitive, adjuvant, or palliative treatment (1), and about 40% of cancer patients are treated with radiotherapy alone, or jointly with surgery and chemotherapy (1). Modern medical linear accelerators (LINACs) are precise in delivering the radiation dose to the target volume, while delivering only minimal doses to the normal tissues. As a result, the tumor tissue in the target volume is significantly destroyed, while the surrounding healthy tissue is mostly protected from irradiation (2). To ensure the safe delivery of the radiation dose to the patient the high-quality treatment plan is needed, as radiotherapy errors can result in serious harm to the patient and can even be lethal (3).

As treatment planning and dose delivery are complex in modern radiotherapy, enhanced quality control is required throughout the entire treatment workflow. The physics plan review was designed to meet this requirement, and is conducted by clinical professionals, such as medical physicists. A physics plan review comprises a comprehensive inspection of various data and aspects associated with the patient’s treatment plan. It is one of the most effective measures to ensure the quality control of patient treatment. Guidelines, such as those of the American Association of Physicists in Medicine Task Group 275 (4) and the Medical Physics Practice Guideline (5), are strictly followed. To date, the manual process is the most popular plan review method, but its efficiency is limited, and it is prone to errors due to inconsistencies in human subjectivities (3). The routine physics plan review covers the major radiotherapy modalities, including conventional three-dimensional conformal radiotherapy (3D-CRT) and intensity modulated radiotherapy (IMRT). The plan review content includes patient positioning, on-board imaging, treatment planning, and delivery (6). However, for certain advanced treatment modalities, such as tomotherapy and volumetric modulated arc therapy (VMAT), the current design (7) is insufficient and needs updating.

To assist in the manual review process, automated methods have been introduced in this field (8). These methods are mainly rule-based and mimic the way in which human operators review the checklist of treatment plan. In the clinic, these methods are implemented in the oncology information system (OIS), and the plan review process is conducted automatically. Dewhurst et al. (9) developed a semi-automatic system to assist in the inspection of treatment plans. Covington et al. (10) designed software, the Plan-Checker Tool, to automate the plan review process. Furhang et al. (11) also developed software to automate the review of intra- and inter-fraction treatment plans. Yang et al. developed dynamic scripts to automate the consistency check of treatment plans (12). These automatic methods improved the accuracy and efficiency of the physics plan review process significantly.

Previous studies have largely focused on traditional or old treatment modalities, such as 3D-CRT. For example, Azmandian et al. (13) proposed a K-means clustering method to detect planning errors of four-field box plans. Kisling et al. used a deep-learning model trained on beam apertures, and digitally reconstructed radiographs to verify the clinical acceptability of four-field box plans (14). Besides 3D-CRT, mainstream modern radiotherapy technologies include IMRT and VMAT, which are more complex in terms of their plan parameters. VMAT plans usually consist of hundreds of apertures for a 360° single-gantry rotation (one arc), while IMRT plans usually use a few apertures, corresponding to 5–10 gantry angles. For VMAT plans, the gantry motion speed and leaves must be synchronized perfectly to ensure the best delivery efficiency and accuracy (15). Therefore, for a VMAT plan review task, the checklist items not only include static features, such as monitor units (MUs) and apertures, but also include dynamic features, such as leaf positions at the adjacent gantry angles.

Characterizing an object with few relevant parameters is crucial for successful machine-learning applications. However, selecting a set of salient features from a VMAT plan consisting of thousands of parameters is challenging (16). In previous studies, a set of a few parameters were manually selected from an IMRT plan by an experienced medical physicist and used in a learning model (17). However, these parameters are insufficient for a VMAT plan that consists of hundreds of apertures. Thus, a more effective method is needed to perform the feature extraction automatically.

Anomaly detection is an important field in computer vision and pattern recognition (18). Several methods have been introduced to assist in the identification of anomalies in the plan review process. Most recently, deep-learning methods have predominated in this field and shown their power in certain challenging tasks (19). Among them, the AutoEncoder (AE) has been widely used in shape representation (20,21), credit fraudulence detection (22,23), network attacking monitoring (24,25), etc. It has also been applied in radiotherapy for modeling organ motion (26), machine fault detection (27), and treatment quality assurance (28). Its application in the physics review of hybrid plans (comprising two IMRT and two 3D-CRT fields) for breast cancer patients has also been investigated (29).

In the present study, a Vanilla AE combined with an aperture-based feature map was used to build a detection model for the physics review of a VMAT plan. In the method section, the generation of a feature map from the control points (CPs) of the VMAT plan is first described, and the training and evaluation of the learning model are then explained. In the results section, the performances of the Vanilla AE and the other four classic machine-learning models are compared and analyzed, and the reliability of the Vanilla AE with respect to two perturbation factors are assessed. In the discussion section, the advantages and limitations of the proposed method are discussed.

Methods

Data

In this study, the data is VMAT treatment plans consisting of two full arcs. Each arc had 90 CPs; the total number of CPs for the two arcs was 180. To protect the lungs, virtual blocks in both sides of the lungs were set, and the dose constraints for the lungs were enhanced. These plans were designed using the Pinnacle³ treatment planning system (Philips Medical Systems, Andover, MA, USA) and executed using Synergy LINAC (Elekta Oncology Systems, Stockholm, Sweden). The multi-leaf collimator (MLC) of Synergy has 40 pairs of leaves (leaf width: 1.0 cm) that are mainly used to shape the beam aperture for radiotherapy. For an IMRT plan, multiple beam apertures are used at a gantry angle; for a VMAT plan, a single-beam aperture is used at a gantry angle. Compared with those in IMRT, those in VMAT are equally spaced and densely distributed over an arc. Usually, a single arc of 360° is applied for VMAT plans with gantry angles ranging from –180° to 180° and angle spacing ranging from 2° to 4°. In a VMAT plan, a set of parameters, including the leaf positions, MU, and dose rate, are associated with each gantry angle or CP.

In this study, the data of 677 VMAT plans for lung cancer patients were collected and tested. These VMAT plans had been clinically approved for patient treatment. Among them, 652 were regular plans and 25 were irregular plans. The feature maps were extracted from these plans for the model training. The training set contained 80% regular plan samples, while the testing set contained 25 irregular plan samples and 20% regular plan samples. The regular plans completely met the clinical standards and had no room for improvement. The irregular plans also met the clinical standards but had room for improvement. For example, if the numbers, ranges, or directions of the multiple arcs in a VMAT plan are not set properly, these plans are still deliverable, but the treatment quality and efficiency may be compromised to certain degrees. In the worst case, this could cause a mechanical fault or treatment interruption. In addition, the leaf position and MU could be changed unintentionally due to a glitch or software bug, which could cause wrong beam apertures and MUs during beam delivery. Therefore, irregular plans need to be identified early to avoid further risks to patients.

The treatment records of the VMAT plans between 2018 and 2024 were collected from the National Cancer Hospital, Chinese Academy of Medical Sciences. This study was conducted in accordance with the Declaration of Helsinki (as revised in 2013). The Ethics Committee of the National Cancer Center/Cancer Hospital, Chinese Academy of Medical Sciences, and Peking Union Medical College approved this study (No. NCC2023C-675). The requirement of written informed consent was waived due to the retrospective design of the study.

Aperture-based feature map

The MU parameters are numbers that can be directly processed by the learning model. However, the beam aperture is formed by a set of leaf positions and converted to a two-dimensional intensity map. As Figure 1 shows, the shape of a beam aperture at k-th CP can be represented by the i-th leaf index and j-th position index on a grid. The width of leaf is thinner (e.g., 0.25 mm) at the center and thicker (e.g., 0.50 mm) at the border. For consistency, the various leaf widths are interpolated to a single thickness or resolution (e.g., 0.1 mm). The resolution of the leaf position is mainly determined by the step size of the leaf motion.

Figure 1 An example of three apertures at adjacent control points.

The aperture at each CP is defined by the positions of all pairs of leaves and is expressed as follows:

Ak(i,j)={1,(i,j)∈Regionk0,else[1]

where A_k is the k-th aperture defined by all pairs of leaves at their positions (i, j). The corresponding intensity map I_k is expressed as:

Ik(i,j)=dkAk(i,j)[2]

where d_k represents the MU at the k-th CP, and K is the total number of CPs. The intensity map is re-sampled to a finer resolution (pixel size: 0.1 mm). The flow chart for extracting the feature maps from a VMAT plan is illustrated in Figure 2. Finally, a series of 2D feature maps for all CPs of a VMAT plan are obtained and become the samples for model learning.

Figure 2 Flowchart of the process used to create a series of two-dimensional feature maps for a VMAT plan. VMAT, volumetric modulated arc therapy.

Vanilla AE

A Vanilla AE is a classic AE model developed for feature extraction and representation. The input is the original 2D feature maps from the 180 CPs of a VMAT plan. The output is the reconstructed 2D feature maps by the latent variables in the bottleneck. The dimensions of the 2D feature map are 80 by 80. The detailed network architecture of the Vanilla AE is illustrated in Figure 3.

Figure 3 The network architecture of the Vanilla AE. Conv, convolution; BN, batch normalization; ReLu, rectified linear unit; AE, AutoEncoder.

As Figure 3 shows, the encoder part of AE takes a stack of images as input and compresses them to a latent vector in the bottleneck. The encoder consists of four down convolution blocks and one linear block. Each down convolution block is implemented by two convolutions; that is, a 3×3 convolution by two strides, and a one 3×3 convolution by one stride. Each linear block is implemented by one linear layer. Both blocks are followed by a batch normalization layer and a rectified linear unit. The decoder consists of the same four down convolution blocks and one linear block as the encoder part of AE but in the reverse sequence. It reconstructs the latent vector at the bottleneck layer back into the original input. The mean-square error (MSE)-based loss function, L(…), is used to penalize the difference between the reconstructed and original inputs, and is expressed as follows:

L(I,I′)=1K∑k=1K[Ik−I′k]2[3]

where I_k and I′k(k=1,…,k) are the original and reconstructed feature map for an aperture. The MSE is used to represent the reconstruction error in the following sections.

Anomaly detection

Figure 4 shows the workflow used to identify any anomalies by the Vanilla AE model. The AE model first learns using the regular plans, and their reconstruction errors are then obtained. To ensure that 100% of the irregular plans are correctly detected and the least percentage of regular plans are falsely detected, the threshold alpha (α) is selected as the maximum value that causes a zero false positive rate (FPR). For a tested sample, its reconstruction error is compared with the threshold α for a decision.

Figure 4 Flowchart of anomaly detection for the VMAT plan review. The model learning and testing routes are shown by the solid and dotted lines, respectively. AE, AutoEncoder; VMAT, volumetric modulated arc therapy.

In the clinic, the AE model can become a component of the OIS. After treatment planning is completed, and the physics review task of the VMAT plan is submitted in the OIS, the features of the treatment plan are extracted and processed by the AE model automatically. Depending on the review result, the plan is revised or proceeds to the next step. As a semi-automated detection module had already been installed in our OIS, it was practical to implement and test the proposed AE model in the OIS in the same fashion as the existing module. It should be noted that the current semi-automated detection module completes each review task in about 5 seconds, while the manual process takes about 300 seconds. If the AE model were to be implemented in the clinic, the efficiency of the physics plan review process would be improved significantly.

Evaluation

The Vanilla AE was compared with other classical methods; that is, models based on the principal component analysis (PCA) (30), isolation forest (IF) (31), local outlier factor (LOF) (32), and hierarchical density-based spatial clustering of applications with noise (HDBSCAN) (33). To accommodate the input of these algorithms, the 2D feature maps were flattened to a one-dimensional intensity vector. The different metrics of the reconstruction errors or differences were employed by these models for anomaly detection. The distance between the intensity map reconstructed by the few principal components and the original intensity map were used in the PCA-based detection model. A five-fold cross-validation was employed for model learning. The area under the curve (AUC) of the receiver operating characteristic (ROC) was used to evaluate the performance of the models. Additionally, the FPR, accuracy, precision, and f1-score, etc. of the models were also assessed.

To evaluate the reliability of the AE model, two perturbation factors (leaf offset and MU variation) were assessed. Leaf offset and MU variation can result in an improper beam shape and dose being delivered to the patient, which carries a high risk and thus must be prevented. The MU perturbation was introduced by increasing its value by 10%, 20%, …, 100% of its maximum for all CPs. The leaf-position perturbation was introduced by shifting its position by 10%, 20%, …, 100% of its maximum for partial or all leaves. Two types of leaf offsets were mimicked. First, the central 10 pairs of leaves were shifted and the other 30 leaves were fixed. Second, the total 40 pairs of leaves were all shifted. The perturbations were applied to all regular plans, and their corresponding reconstruction errors were computed. If the average error in relation to a perturbation level was more than the detection threshold α, this perturbation level was detected by the model.

Results

Model comparison

Table 1 shows the detection performance of the Vanilla AE and the other four classic models. The Vanilla AE achieved the best performance with an AUC of 0.943. The AUCs of the PCA and HDBSCAN models were also comparable to that of the Vanilla AE. The AUCs of the IF and LOF models were relatively lower than the AUC of the AE model. With a zero FPR, the precision and accuracy of the Vanilla AE were 0.407 and 0.769, respectively, and the Vanilla AE outperformed the other four models. Compared with the PCA model, which had the best performance among the classic models, the precision and accuracy of the Vanilla AE showed improvements of 18.6% and 10.0%, respectively.

The ROC curves of the five models are compared in Figure 5. The performances of the Vanilla AE and PCA model were comparable and higher than the other models. The performance of the HDBSCAN model was slightly lower than the Vanilla AE and PCA model, but higher than the IF and LOF models. The AUC of the five models were all more than 0.86, which is high for model performance.

Figure 5 ROC curves of all five detection models. AE, AutoEncoder; PCA, principal component analysis; IF, isolation forest; LOF, local outlier factor; HDBSCAN, hierarchical density-based spatial clustering of applications with noise; ROC, receiver operating characteristic.

Distance distribution

The box plot of distance distributions of the Vanilla AE is presented in Figure 6. The median of the irregular plan class was higher than the maximum value of the regular plan class. The minimum of the irregular plan class was close to the 75th percentile of the regular plan class. The gap between the irregular and regular plan classes was significant. However, as the number of irregular plans was far less than that of the regular plans, the gap between these two groups was less clear visually.

Figure 6 The distance distribution for all testing samples with the Vanilla AE. The maximum and minimum of the distribution are indicated by the top and bottom horizontal lines, respectively. The outliers are those points beyond these two lines. The 75th and 25th percentile of the distribution are denoted by the top and bottom edges of the box, respectively. The median of the distribution is represented by the middle line in the box. AE, AutoEncoder.

For comparison, the box plots of the distance distributions of the other four classic models are shown in Figure 7A-7D. The box plot of the PCA model was similar to that of the Vanilla AE, and both plots showed a clear gap between the distributions of the irregular and regular plans. For the IF, LOF, and HDBSCAN models, the median of the irregular plan class was lower than the maximum of the regular plan class. The minimum of the irregular plan class was far less than the 75th percentile of the regular plan class. Among the five models, Vanilla AE showed the largest gap between the distributions of the irregular and regular plans.

Figure 7 The distance distribution of the testing samples using four classic anomaly detection models: (A) PCA; (B) IF; (C) LOF; and (D) HDBSCAN. PCA, principal component analysis; IF, isolation forest; LOF, local outlier factor; HDBSCAN, hierarchical density-based spatial clustering of applications with noise.

Perturbation test

The MSE plot of the Vanilla AE in relation to the different perturbation levels is shown in Figure 8. Leaf perturbation can be detected when the perturbation levels is 10% or more. For the central 10 pairs of leaves, the MSE of the Vanilla AE was 0.003 in relation to the 10% perturbation level when the detection threshold was 0.0028 (Figure 8A). For all 40 pairs of leaves, the MSE of the Vanilla AE was 0.01 in relation to the 10% perturbation level when the detection threshold was 0.002 (Figure 8B). For the MU, the MSE of the Vanilla AE was 0.003 in relation to the 20% perturbation level when the detection threshold was 0.0026 (Figure 8C). Among the two types of perturbations, the sensitivity in detecting the leaf offset was higher than that in detecting MU variation.

Figure 8 The MSE plots of the AE model in relation to the different levels of perturbations. (A) Leaf offset with the central 10 pairs of leaves. (B) Leaf offset with all 40 pairs of leaves. (C) Increasing MUs for all CPs. The blue line and vertical line indicate the mean and standard deviation of the MSE in relation to the different perturbation levels. The red dotted line indicates the detection threshold α. MSE, mean-square error; AE, AutoEncoder; MU, monitor unit; CPs, control points.

The detection probabilities of the Vanilla AE in relation to the different perturbation levels are summarized in Table 2. The detection probability was calculated based on the proportion of the number of plans with reconstruction errors more than α to the total number of plans with respect to a given perturbation level. For shifting the central 10 pairs of leaves, the probabilities in detecting 20% and 50% perturbation levels were 55.7% and 86.2%, respectively. For shifting all 40 pairs of leaves, the probabilities in detecting 10% and 50% perturbation levels were 98.4% and 100%, respectively. For the MU variation, the probabilities in detecting 30% and 60% perturbation levels were 57.2% and 82.4%, respectively. Among the two perturbation factors, the probability in detecting leaf offset was higher than that in detecting MU variation. This result is consistent with the observations shown in Figure 8.

Discussion

The performance of the Vanilla AE and the other classic models in automatic plan reviews was evaluated based on the proposed aperture map. Compared to the typical geometry parameters, such as source-to-surface distance and beam angle, the 2D aperture map provides abundant information associated with informational radiation beams. In addition to static geometrical information related to each beam, the dynamic motion information, such as leaf positions at adjacent CPs, can be examined quantitatively. One study generated a feature map by combining all apertures of CPs of an IMRT plan to identify the treatment modalities/sites used in treatment planning (16). It is sensitive in distinguishing IMRT plans with different treatment modalities/sites, but its ability to handle the complexity of leaf positions in a VMAT plan is insufficient. Therefore, a series of 2D feature maps are needed, and are more suitable than a single feature map.

Compared with classic detection models, such as the PCA, the Vanilla AE is more complex and effective. The nonlinear activation function is employed in the encoder/decoder of the AE model. This allows the AE model to approximate any nonlinear function with high precision. The more complex mapping relationships between the high-dimensional space and low-dimensional space can be learned effectively (17). The Vanilla AE achieved the best performance among the five tested detection models based on the majority of the evaluation metrics, including AUC, accuracy, and precision. Notably, besides the current application on VMAT plan, the Vanilla AE model can also be extended to other treatment sites and modalities.

The reliability of the Vanilla AE model was evaluated using the perturbation test. When applying perturbation to all 40 pairs of leaves, the detection model was highly sensitive. However, if the perturbation was applied to the central 10 pairs of leaves, the sensitivity of the Vanilla AE was decreased. For the perturbation to the MUs, the sensitivity of the Vanilla AE was even less than that of the leaf offset. Therefore, the detection sensitivity of the Vanilla AE for leaf offset was higher than that of the MU variation. This was also observed in the detection probability (Table 2). To achieve a detection probability rate higher than 90%, the required perturbation levels were 60% for shifting the central 10 pairs of leaves, 10% for shifting all 40 pairs of leaves, and 80% for MU variation. Therefore, compared to leaf offset, the perturbation to the MU variation was less detectable for the Vanilla AE model.

The current AE model identified anomalies more accurately than the classic models; however, it was unable to identify the reasons for these anomalies. The AE and classic detection models tested in this study largely rely on the reconstruction error or distance, which is the combination of multiple perturbations or offsets. However, there is no way to distinguish the causes of anomaly based solely on the magnitude of the reconstruction error. The magnitude of the reconstruction error is affected by the scale of perturbations or offsets. A small perturbation or offset may result in less change in the reconstruction error or a distance that may not be detected by AE models (Figure 8 and Table 2). With 10% perturbation of the leaf position, the AE model was able to detect anomalies for the shifts of 40 pairs of leaves but was unable to do so for the shifts of 10 pairs of leaves.

Combining an intensity map and the Vanilla AE in the physics review of VMAT plan in radiotherapy is viable, but several limitations in this study should be considered. First, this study only tested VMAT treatment plans for lung cancer patients. Thus, other treatment sites should be examined in the future. The effectiveness of the AE model in IMRT plan review was not tested in this study but will be tested in the future. Second, the Vanilla AE is a typical AE model and requires further improvements. For example, the introduction of attention and adversarial modules could improve the learning capability of the model (34,35). Additionally, network parameter tuning is time consuming, and the automation mechanism should be considered (36). Third, the ratio of irregular versus regular plans was highly imbalanced. This would have resulted in poor predictive performance of the model for the minority class (37). To address this issue, synthetic data should be generated using a generative deep-learning model to compensate for the minority class in future research (38).

Conclusions

The 2D feature map obtained from the beam aperture is a novel way to characterize the MLC properties of a VMAT plan. The Vanilla AE was more effective at detecting irregular plans from regular radiotherapy plans than the classic machine-learning models. The combination of the intensity map and Vanilla AE model increases the accuracy and efficiency of the automatic review of VMAT plans.

Acknowledgments

Funding: This work was supported by the Non-profit Central Research Institute Fund of Chinese Academy of Medical Sciences (No. 2024-RW320-05), the CAMS Innovation Fund for Medical Sciences (CIFMS) (No. 2023-I2M-C&T-B-076), the National High Level Hospital Clinical Research Funding (No. 2022-CICAMS-80102022203), the National Natural Science Foundation of China (No. 11975312), and the Beijing Municipal Natural Science Foundation (No. 7202170).

Footnote

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-24-1398/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki (as revised in 2013). The Ethics Committee of the National Cancer Center/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College approved this study (No. NCC2023C-675). The requirement of written informed consent was waived due to the retrospective design of the study.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Du S, Chen Y, Zhang Q, Sun J, Ma G, Wang J, Zhang X, Sun T, Zhang Y, Zhang J, Wu Q, Zeng Z. Modern Radiotherapy in the Multidisciplinary Management of Common Cancers. Clinical Cancer Bulletin 2022;1:81-94.
2. Gardner SJ, Kim J, Chetty IJ. Modern Radiation Therapy Planning and Delivery. Hematol Oncol Clin North Am 2019;33:947-62. [Crossref] [PubMed]
3. Ganesh T. Incident reporting and learning in radiation oncology: Need of the hour. J Med Phys 2014;39:203-5. [Crossref] [PubMed]
4. Ford E, Conroy L, Dong L, de Los Santos LF, Greener A, Gwe-Ya Kim G, Johnson J, Johnson P, Mechalakos JG, Napolitano B, Parker S, Schofield D, Smith K, Yorke E, Wells M. Strategies for effective physics plan and chart review in radiation therapy: Report of AAPM Task Group 275. Med Phys 2020;47:e236-72. [Crossref] [PubMed]
5. Xia P, Sintay BJ, Colussi VC, Chuang C, Lo YC, Schofield D, Wells M, Zhou S. Medical Physics Practice Guideline (MPPG) 11.a: Plan and chart review in external beam radiotherapy and brachytherapy. J Appl Clin Med Phys 2021;22:4-19. [Crossref] [PubMed]
6. Yang D, Wu Y, Brame RS, Yaddanapudi S, Rangaraj D, Li HH, Goddu SM, Mutic S. Technical note: electronic chart checks in a paperless radiation therapy clinic. Med Phys 2012;39:4726-32. [Crossref] [PubMed]
7. Cui T, Peng JX, Jin XL, Chu XG, Zhang J, Cui LZ, Xing XF. Improving intensity-modulated radiation therapy quality assurance by adopting statistical process control. Cancer Radiother 2022;26:427-32. [Crossref] [PubMed]
8. Gopan O, Zeng J, Novak A, Nyflot M, Ford E. The effectiveness of pretreatment physics plan review for detecting errors in radiation therapy. Med Phys 2016;43:5181. [Crossref] [PubMed]
9. Dewhurst JM, Lowe M, Hardy MJ, Boylan CJ, Whitehurst P, Rowbottom CG. AutoLock: a semiautomated system for radiotherapy treatment plan quality control. J Appl Clin Med Phys 2015;16:5396. [Crossref] [PubMed]
10. Covington EL, Chen X, Younge KC, Lee C, Matuszak MM, Kessler ML, Keranen W, Acosta E, Dougherty AM, Filpansick SE, Moran JM. Improving treatment plan evaluation with automation. J Appl Clin Med Phys 2016;17:16-31. [Crossref] [PubMed]
11. Furhang EE, Dolan J, Sillanpaa JK, Harrison LB. Automating the initial physics chart checking process. J Appl Clin Med Phys 2009;10:129-35. [Crossref] [PubMed]
12. Yang D, Moore KL. Automated radiotherapy treatment plan integrity verification. Med Phys 2012;39:1542-51. [Crossref] [PubMed]
13. Azmandian F, Kaeli D, Dy JG, Hutchinson E, Ancukiewicz M, Niemierko A, Jiang SB. Towards the development of an error checker for radiotherapy treatment plans: a preliminary study. Phys Med Biol 2007;52:6511-24. [Crossref] [PubMed]
14. Kisling K, Cardenas C, Anderson BM, Zhang L, Jhingran A, Simonds H, Balter P, Howell RM, Schmeler K, Beadle BM, Court L. Automatic Verification of Beam Apertures for Cervical Cancer Radiation Therapy. Pract Radiat Oncol 2020;10:e415-24. [Crossref] [PubMed]
15. Otto K. Volumetric modulated arc therapy: IMRT in a single gantry arc. Med Phys 2008;35:310-7. [Crossref] [PubMed]
16. Kump PM, Xia J, Yaddanapudi S, Bai E. An automated treatment plan alert system to safeguard cancer treatments in radiation therapy. Mach Learn Appl 2022;10:100437. [Crossref] [PubMed]
17. Huang P, Yan H, Song Z, Xu Y, Hu Z, Dai J. Combining autoencoder with clustering analysis for anomaly detection in radiotherapy plans. Quant Imaging Med Surg 2023;13:2328-38. [Crossref] [PubMed]
18. Hojjati H, Ho TKK, Armanfard N. Self-supervised anomaly detection in computer vision and beyond: A survey and outlook. Neural Netw 2024;172:106106. [Crossref] [PubMed]
19. Pang G, Shen C, Cao L, Van Den Hengel A. Deep Learning for Anomaly Detection. ACM Computing Surveys (CSUR) 2021;54:1-38. [Crossref]
20. ChalapathyRChawlaS.Deep Learning for Anomaly Detection: A Survey.2019. arXiv:1901.03407v2.
21. Pimentel MAF, Clifton DA, Clifton L, Tarassenko L. A review of novelty detection. Signal Processing 2014;99:215-49. [Crossref]
22. Misra S, Thakur S, Ghosh M, Saha SK. An Autoencoder Based Model for Detecting Fraudulent Credit Card Transaction. Procedia Computer Science 2020;167:254-62. [Crossref]
23. Fanai H, Abbasimehr H. A novel combined approach based on deep Autoencoder and deep classifiers for credit card fraud detection. Expert Systems with Applications 2023;217:119562. [Crossref]
24. Song Y, Hyun S, Cheong YG. Analysis of Autoencoders for Network Intrusion Detection. Sensors (Basel) 2021.
25. Lopes IO, Zou D, Abdulqadder IH, Ruambo FA, Yuan B, Jin H. Effective network intrusion detection via representation learning: A Denoising AutoEncoder approach. Computer Communications 2022;194:55-65. [Crossref]
26. Mezheritsky T, Romaguera LV, Le W, Kadoury S. Population-based 3D respiratory motion modelling from convolutional autoencoders for 2D ultrasound-guided radiotherapy. Med Image Anal 2022;75:102260. [Crossref] [PubMed]
27. Dou T, Clasie B, Depauw N, Shen T, Brett R, Lu HM, Flanz JB, Jee KW. A deep LSTM autoencoder-based framework for predictive maintenance of a proton radiotherapy delivery system. Artif Intell Med 2022;132:102387. [Crossref] [PubMed]
28. Wang L, Li J, Zhang S, Zhang X, Zhang Q, Chan MF, Yang R, Sui J. Multi-task autoencoder based classification-regression model for patient-specific VMAT QA. Phys Med Biol 2020;65:235023. [Crossref] [PubMed]
29. Huang P, Shang J, Xu Y, Hu Z, Zhang K, Dai J, Yan H. Anomaly detection in radiotherapy plans using deep autoencoder networks. Front Oncol 2023;13:1142947. [Crossref] [PubMed]
30. Pearson K. LIII. On lines and planes of closest fit to systems of points in space. Philosophical Magazine 1901;2:559-72.
31. Liu FT, Ting KM, Zhou ZH. Isolation Forest. 2008 Eighth IEEE International Conference on Data Mining: 15-19 Dec. 2008; 2008:413-422.
32. Goodfellow I, Bengio Y, Courville A: Deep Learning: MIT Press; 2016.
33. Campello RJGB, Moulavi D, Sander J. Density-Based Clustering Based on Hierarchical Density Estimates. In: Pei J, Tseng VS, Cao L, Motoda H, Xu G, editors. Advances in Knowledge Discovery and Data Mining: 2013; Berlin, Heidelberg: Springer Berlin Heidelberg; 2013:160-72.
34. Oluwasanmi A, Aftab MU, Baagyere E, Qin Z, Ahmad M, Mazzara M. Attention Autoencoder for Generative Latent Representational Learning in Anomaly Detection. Sensors (Basel) 2021.
35. Zhou X, Hu K, Wang H. Robustness meets accuracy in adversarial training for graph autoencoder. Neural Netw 2023;157:114-24. [Crossref] [PubMed]
36. Galván E, Mooney P. Neuroevolution in deep neural networks: Current trends and future challenges. IEEE Trans Artif Intell 2021;2:476-93. [Crossref]
37. Guo H, Li Y, Jennifer S, Gu M, Huang Y, Gong B. Learning from classimbalanced data: review of methods and applications. Expert Syst Appl 2017;73:220-39. [Crossref]
38. Liu J, Yan H, Cheng H, Liu J, Sun P, Wang B, Mao R, Du C, Luo S. CBCT-based synthetic CT generation using generative adversarial networks with disentangled representation. Quant Imaging Med Surg 2021;11:4820-34. [Crossref] [PubMed]