A deep learning-driven automated treatment planning framework for patient treated with radiotherapy in cervical cancer

A deep learning-driven automated treatment planning framework for patient treated with radiotherapy in cervical cancer | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article A deep learning-driven automated treatment planning framework for patient treated with radiotherapy in cervical cancer Boda Ning, Xiuyan Liang, Zhenguo Cui, Yingfa Li, Qi Liu, Shuaining Ma, and 4 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8749620/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 14 Apr, 2026 Read the published version in Radiation Oncology → Version 1 posted 7 You are reading this latest preprint version Abstract Background and purpose: The rapid and efficient generation of high-quality, dose-consistency volumetric modulated arc therapy (VMAT) plans remains challenging in radiotherapy. This study proposes a deep learning (DL) end-to-end (E2E) auto-planning framework and validate its practicality and feasibility for clinical implementation. Materials and methods: A total of 458 cervical cancer VMAT plans were enrolled and split into training, validation, and test cohorts. An E2E auto-planning framework with a two-stage cascaded deep learning (DL) network was developed: Stage 1 predicted coarse dose from CT and structure masks, and Stage 2 refined it using four beam-band priors and a composite loss. Dose-volume histogram (DVH) endpoints from refined predicted dose were converted into Monaco objectives via a scripting module for iterative optimization. Performance was evaluated with Dose, DVH, and snDVH scores, ablations, and comparisons with manual plans in terms of quality, clinical evaluation and deliverability. Results: The proposed DL method achieved the best performance, with Dose score, DVH score and snDVH score of 2.114 ± 0.218 Gy, 1.194 ± 0.295 Gy and 2.027 ± 0.586, respectively. Compared with manual plans, E2E auto-plans preserved target volume coverage while reducing all DVH metrics for bladder, rectum, small intestine, and spinal cord by 2% - 35% (all p < 0.05). The gamma passing rate of E2E auto-plans was higher than manual plans in the 3%/3 mm gamma criterion (98.1% vs 97.9%). Conclusion: The proposed auto-planning framework demonstrated a high level of automation and clinical applicability, offering a reliable and promising tool to support radiotherapy workflows. Auto-planning Deep learning Dose prediction Treatment planning system script Radiotherapy Figures Figure 1 Figure 2 Figure 3 Introduction Intensity-modulated radiotherapy (IMRT) and volumetric modulated arc therapy (VMAT) are widely used for cervical cancer (CC) because they enable highly conformal target coverage while improving sparing of surrounding organs at risk (OARs) through inverse planning and modulation [ 1 – 3 ]. In routine treatment, prescription and clinical priorities are determined by radiation oncologists, whereas plan generation relies on medical physicists to iteratively adjust optimization objectives and parameters until institutional and guideline-based criteria are satisfied [ 4 ]. However, both the quality and efficiency of manual planning largely depend on planner experience, and plan quality can vary across individual planners, posing a major challenge to workflow efficiency in busy clinics [ 5 ]. To improve efficiency and reduce variability, automated treatment planning (ATP) has been extensively studied [ 6 ]. Knowledge-based planning (KBP) is widely used to improve planning efficiency and consistency by modeling planning objectives to guide the treatment planning workflow, thereby minimizing planner intervention [ 7 – 9 ]. Dose-volume histogram (DVH)-based prediction is one established KBP approach, in which predicted DVH endpoints are used as planning goals; however, accuracy can be limited because the detailed space information of three-dimensional (3D) dose distribution is not explicitly modeled [ 10 , 11 ]. In parallel, deep learning (DL)–based dose prediction has advanced rapidly, enabling fast generation of patient-specific 3D dose distributions from planning CT and structure masks after training on prior high-quality plans [ 12 – 19 ]. Nevertheless, even highly accurate predicted dose distributions are not inherently deliverable, and the gap between “dose prediction” and “clinically executable plans” remains a key barrier to achieve a closed-loop workflow. Recent studies have begun to bridge this gap by converting DL-predicted dose distributions into deliverable IMRT or VMAT plans [ 20 – 22 ]. However, whether predicted doses can be translated robustly into treatment planning system (TPS) objectives that converge reliably to high-quality, deliverable plans requires further validation. This need is most evident in steep dose fall-off regions and at target–OAR interfaces, where anatomy-only prediction may not adequately reflect the directionality and gradient constraints imposed by inverse planning, limiting robust translation into stable TPS objectives and consistent plan quality. We therefore propose a two-stage cascaded DL framework that incorporates directional beam band priors and a DVH-oriented composite loss, and converts the refined dose into patient-specific, Monaco-executable objectives via in-house TPS scripting for automated inverse optimization. This study aimed to establish and validate an end-to-end (E2E) VMAT auto-planning workflow for CC and to assess dose prediction accuracy, plan quality, and deliverability. Materials and methods Patients and clinical plans This retrospective study was approved by the Institutional Review Board of Institution A. A total of 458 CC patients treated with full-arc VMAT between 2023 and 2025 were enrolled according to the exclusion criteria (Fig. S1 ) and randomly assigned to training, validation, and test cohorts at an 8:1:1 ratio. For each patient, a simulation CT was acquired on a Philips Brilliance Big Bore scanner (Philips Healthcare, Best, The Netherlands) with a slice thickness of 5 mm. Target volumes and OARs were contoured by radiation oncologists in accordance with Radiation Therapy Oncology Group guidelines using AccuContour (v3.2; Manteia Technologies Co., Ltd.). All plans were prescribed 45 Gy in 25 fractions and were generated under the institutional planning protocol by medical physicists, with target coverage ensured and OAR sparing prioritized. Plan optimization and final dose calculation were performed in Monaco (clinical version 6.2.2; Elekta AB, Stockholm, Sweden) TPS. The planning objectives and dose constraints for targets and OARs are summarized in Table S1 . Construction of beam band masks An analysis of full-arc VMAT plans for CC patients revealed that the vast majority of dose delivery pathways were predominantly concentrated along several beam-entry directions (Fig. 1 a(1)). This directional concentration may reflect inverse optimization mechanism, where OARs dose constraints encouraged beam delivery to avoid traversing critical organs or, when avoidance is not feasible, to lower the beam field weight from those directions. Consequently, four directional beam-band masks were generated on each CT slice. Each beam band mask consisted of two parallel boundaries that are tangent to the plan target volume (PTV) on the respective slice (Fig. 1 a(2)). These masks were used as auxiliary inputs to the Stage 2 refinement network. Dose prediction model architecture Before DL network training, CT images must be preprocessed to standardize input dimensions (Appendix A). We developed a cascaded DL framework for E2E prediction of patient-specific 3D dose distributions, and the workflow is shown in Fig. 1 b. The framework comprises two consecutive stages. In Stage 1, a 3D U-Net was trained using the planning CT together with binary structure masks of the PTV, selected OARs, and the external body contour as multi-channel inputs [ 23 ]. The network outputs a coarse 3D dose map and is supervised against the reference dose using a mean absolute error (MAE) loss. This stage captured the global dose morphology from patient anatomy and structure masks. The second stage introduces a refinement step conditioned on directional beam band information. Specifically, four directional beam band masks were generated to approximate predominant entrance paths in clinical VMAT plans and used as auxiliary geometric priors. Subsequently, the coarse dose from Stage 1 was concatenated with the CT, structure masks, and the four band masks to form 15 input channels, and a 3D ResU-Net was trained to produce the refined predicted dose [ 24 ]. The stage 2 was supervised using a composite objective comprising four loss functions, including L body , L band , L grad and L DVH , which jointly encourage global accuracy, band-region fidelity, gradient consistency, and DVH-related agreement. Full definitions of each loss function were provided in Appendix B. The second-stage overall loss was defined as a weighted sum: $$\:\begin{array}{c}{\text{L}}_{\text{total}}\text{}\text{=}{\text{}\lambda}_{\text{body}\text{}}{\text{L}}_{\text{body}}\text{}\text{+}\text{}{\lambda}_{\text{band}\text{}}{\text{L}}_{\text{band}}\text{}\text{+}\text{}{\lambda}_{\text{grad}\text{}}{\text{L}}_{\text{grad}}\text{}\text{+}{\text{}\lambda}_{\text{DVH}\text{}}{\text{L}}_{\text{DVH}}\#\text{(}\text{1}\text{)}\end{array}$$ Detailed network architecture and the training environment and strategy were provided in Appendix C. The full code used in this work is openly available at GitHub ( https://github.com/Rick-Ds/Two_stage_E2E_DosePrediction ). Auto-plan generation To evaluate whether the refined predicted dose could be translated into deliverable treatment plans of comparable quality to the clinical reference, we performed a TPS script–based auto-planning study, with the workflow shown in Fig. 1 c. All procedures were implemented in the Monaco TPS. An in-house script first configured the beam geometry and initialized an automated plan. DVH endpoints were extracted from the refined predicted dose for each region of interest (ROI) and defined as the expected target DVH metrics ( M 0 ). Based on M 0 , plan-specific optimization objectives and constraints were assembled and converted into TPS-compatible parameters to drive inverse optimization. After each optimization finished, DVH metrics of the current plan ( M c ) were compared with the targets M 0 . If unmet, only the corresponding objectives were updated for the next iteration: PTV or OARs dose constraint values were increased or decreased by 2%, followed by re-optimization. We limited the maximum number of iterations to no more than five. If the objective cannot be achieved, the TPS scripting module retains an iteration result closest to the goal. A summary of the constraint functions was provided in Table S2. Evaluation and statistical analysis To demonstrate the superiority of the proposed E2E cascaded DL framework for dose prediction, we benchmarked it against top-performing OpenKBP models and subsequent state-of-the-art architectures, including VNet, C3D, DoseNet, HD-UNet, Attention-aware 3D U-Net, and TransDose [ 12 , 25 – 29 ]. For fairness, all methods were trained and evaluated on the same dataset with an identical training protocol. Performance was quantified using the official OpenKBP Dose score and DVH score, together with a custom scale-normalized DVH score (snDVH score) [ 30 ]. Detailed definitions were provided in Appendix D. Beyond dose prediction, we compared auto-plans with the corresponding clinical reference plans using dose distributions and DVH endpoints. Deliverability was evaluated via linac-based verification of the final auto-plans, and quality assurance results were recorded. To compare our method with reference approaches, paired sample t-tests were performed in SPSS (version 22.0). Statistical significance was defined as a two-tailed p value < 0.05. Ablation experiments To examine the interpretability and practical impact of the chosen number of beam bands, we generated band masks with different angular counts and performed dedicated ablation analyses. In addition, to assess the contribution of each loss component, we used the cascaded network as the baseline framework and conducted a series of ablation experiments in which key terms were added sequentially: (1) the global dose consistency loss L body ; (2) the band-focused dose loss L band ; (3) the band-gradient loss L grad ; and (4) the DVH-based numerical loss L DVH .​ Results In the beam-band ablation, a distinct elbow was observed at k = 4. Dose coverage within band-overlap regions increased from 0.39 at k = 1 to 0.95 at k = 4, whereas further increases in k produced only marginal gains of about 0.01 per additional band. Peak GPU memory increased approximately linearly with k, and k = 4 was therefore adopted as a practical trade-off for subsequent experiments (Supplementary Fig. S2). Benchmarking against six representative DL dose prediction models (Table 1 ), the proposed method achieved the best overall performance across Dose score, DVH score, and snDVH score, reaching 2.114 ± 0.218 Gy, 1.194 ± 0.295 Gy, and 2.027 ± 0.586, respectively. In addition, we assessed the incremental contribution of the four loss functions. As the key components were introduced sequentially, all three metrics showed an overall downward trend in Table 2 ; with the final composite loss, improvements were statistically significant across metrics except that the change in Dose score before versus after adding L DVH did not reach significance. Notably, adding L band and L grad reduced the Dose score from 2.407 to 2.176, indicating improved global and band-focused accuracy. After L DVH was incorporated, DVH score and snDVH score were further reduced by approximately 22% (from 1.526 to 1.194) and 18% (from 2.475 to 2.027), respectively, suggesting that the DVH-based constraint improved ROI-relevant endpoint agreement. In contrast, the change in Dose score after adding L DVH (2.176 to 2.114) was modest and did not show a clear additional gain. Figure 2 further illustrates the prediction accuracy of the proposed method for DVH metrics across the relevant ROIs. Overall, the predicted and reference dose distributions showed close agreement for PTV indices including D max , D mean , D 2% , D 98% , V 95% and HI, and for OAR-related endpoints including D mean , V 30 , and the Dmax of the small intestine and spinal cord. Table 1 Quantitative comparison between SOTA methods and our method in terms of Dose score, DVH score and snDVH Score. Model Dose Score [Gy] DVH Score [Gy] snDVH Score VNet 2.336 ± 0.335* 1.917 ± 0.426* 3.068 ± 0.748* C3D 2.319 ± 0.259* 1.906 ± 0.400* 2.794 ± 0.663* DoseUNet 2.227 ± 0.280* 1.866 ± 0.364* 2.646 ± 0.672* HD UNet 2.204 ± 0.283* 1.651 ± 0.333* 2.535 ± 0.703* Attention-aware 3D UNet 2.180 ± 0.337 1.565 ± 0.508* 2.582 ± 0.641* TransDose 2.136 ± 0.211 1.395 ± 0.293* 2.424 ± 0.666* Proposed 2.114 ± 0.218 1.194 ± 0.295 2.027 ± 0.586 *: P value < 0.05; P-value: calculated by Paired t-test; Table 2 Experimental results of ablation studies on our proposed method. Method Dose Score [Gy] DVH Score [Gy] snDVH Score Baseline 2.407 ± 0.230* 1.983 ± 0.337* 3.037 ± 0.595* Baseline + L band 2.245 ± 0.262* 1.641 ± 0.331* 2.763 ± 0.562* Baseline + L band + L grad 2.176 ± 0.290 1.526 ± 0.362* 2.475 ± 0.653* Baseline + L band + L grad + L DVH 2.114 ± 0.218 1.194 ± 0.295 2.027 ± 0.586 *: P value < 0.05; P-value: calculated by Paired t-test; The quantitative evaluation of the clinical reference plans and the E2E auto-plans were presented in Table 3 . Target coverage was comparable, with no significant differences in PTV V 95% (p = 0.195) or V 100% (p = 0.650), indicating that clinical coverage requirements were maintained. Small but significant reductions were observed in PTV D 2% (p = 0.001) and HI (p = 0.002), and they reflected a reduction in high-dose regions of target and an improvement in dose homogeneity. For OARs, there was no statistically significant difference in marrow DVH metrics between the two plan types, with similar D mean and V 30Gy (p = 0.714 and p = 0.213). Notably, all DVH metrics for the bladder, rectum, small intestine, and spinal cord were significantly reduced in the E2E auto-plans, with relative decreases of 2%–35% compared with the reference plans (all p < 0.05), whereas both femoral heads received higher doses, with an approximately 1 Gy increase in D mean and a 4%–6% increase in V 30Gy compared with the reference plans. Qualitative examples were shown in Fig. 3 . Across three representative cases, the high-dose regions were largely consistent between the E2E auto-plans and reference plans, while improved conformity was observed in the 20–30 Gy dose range. The corresponding DVHs showed nearly overlapping PTV coverage, accompanied by systematic shifts of the bladder, rectum, small bowel, and spinal cord curves toward lower doses within clinically relevant ranges, indicating improved OAR sparing without compromising target coverage, highlighting the dosimetric advantages of the E2E auto-planning approach. Table 3 Quantitative evaluation of DVH metrics of PTV and OARs of GT plan and E2E auto-plan. Structures DVH metrics GT plan E2E auto-plan P-value PTV Dmax(Gy) 49.74 ± 0.47 49.59 ± 0.49 .088 Dmean(Gy) 46.48 ± 0.20 46.34 ± 0.19 .000 V95(%) 99.70 ± 0.30 99.80 ± 0.20 .195 V100(%) 94.70 ± 0.30 94.70 ± 0.30 .650 D98%(Gy) 44.29 ± 0.26 44.35 ± 0.21 .062 D2%(Gy) 48.00 ± 0.32 47.82 ± 0.32 .001 CI 0.70 ± 0.03 0.70 ± 0.03 .070 HI 0.08 ± 0.01 0.07 ± 0.01 .002 Bladder Dmean(Gy) 30.25 ± 0.97 29.58 ± 1.70 .014 V30(%) 47.30 ± 2.30 44.80 ± 5.70 .002 V40(%) 28.90 ± 2.80 24.90 ± 5.20 .000 V45(%) 16.80 ± 3.10 13.60 ± 3.70 .000 Rectum Dmean(Gy) 28.20 ± 1.54 26.81 ± 1.90 .000 V30(%) 46.70 ± 3.80 40.40 ± 6.70 .000 V40(%) 20.70 ± 3.80 13.40 ± 4.30 .000 V45(%) 5.20 ± 3.00 2.00 ± 2.20 .000 Small_Intestine Dmax(Gy) 49.08 ± 0.51 48.77 ± 0.54 .005 Dmean(Gy) 19.12 ± 3.95 18.10 ± 3.89 .000 V30(%) 25.10 ± 6.90 21.00 ± 7.00 .000 SpinalCord Dmax(Gy) 31.98 ± 2.81 29.53 ± 2.61 .000 Dmean(Gy) 13.53 ± 4.16 12.34 ± 3.82 .000 Marrow Dmean(Gy) 33.63 ± 1.07 33.56 ± 1.34 .714 V30(%) 65.50 ± 4.80 64.30 ± 6.10 .213 Left Femoral Head Dmean(Gy) 23.49 ± 3.25 24.50 ± 2.93 .010 V30(%) 18.90 ± 10.50 24.40 ± 10.80 .001 V40(%) 0.40 ± 0.70 1.00 ± 1.30 .002 Right Femoral Head Dmean(Gy) 23.86 ± 2.57 25.05 ± 2.21 .006 V30(%) 18.40 ± 9.90 22.90 ± 8.90 .004 V40(%) 0.60 ± 0.90 1.00 ± 1.30 .014 P-value: calculated by Paired t-test; P value < 0.05 was considered statistically significant; Deliverability verification using gamma analysis with a 3%/3 mm criterion yielded comparable passing rates: the E2E auto-plans achieved a mean of 98.1% (range: from 96.7% to 99.0%), while the reference plans achieved 97.9% (range: from 96.8% to 99.3%) in table S3, indicating similar deliverability to clinical plans. Discussion We developed and validated a closed-loop VMAT auto-planning workflow that converts patient-specific anatomy into deliverable plans. Target coverage was preserved while DVH metrics for key OARs improved, and deliverability was supported by delivery verification. Dose prediction was conditioned on interpretable, direction-aware beam band priors and trained with a composite loss function to better capture steep dose fall-off and target–OAR abutment interfaces. Predicted dose was translated, via TPS script module, into Monaco-executable individualized objectives and iteratively optimized, enabling robust transfer from prediction to actionable constraints and reducing repeated manual tuning, thereby providing a practical route for integrating prediction into E2E clinical VMAT planning. Unlike other DL medical image regression tasks such as auto-segmentation and image reconstruction, where inputs are often relatively direct, dose prediction typically requires deliberate input design [ 31 ]. This is because plan generation is shaped not only by anatomy, but also by clinical preferences and physical realities such as beam arrangement and energy deposition constraints [ 32 , 33 ]. Sun et al. proposed a physics voxel–based optimization strategy and reported mean-dose reductions of 3.1, 6.2 and 4.5 Gy in the bladder and bilateral femoral heads compared with manual plans [ 34 ]. In the work by Teng et al., fixed-field characteristics specific to IMRT were leveraged: they simulated ray paths according to PTV position and beam angles to construct beam masks as network inputs, and demonstrated reduced prediction errors across individual beam angles [ 13 ]. Xiong et al. designed normalized distance-aware beam plates and mass density maps as physics-informed priors, and their ablation experiments showed a 5.8% reduction in MAE after incorporating such priors [ 35 ]. Considering the dynamic modulation and continuous gantry rotation intrinsic in VMAT, we constructed four directional beam band masks to approximate the directionality of dose incidence and used them as geometric priors as network inputs. In ablation experiments, adding the band masks reduced the Dose score from 2.407 to 2.245 (a relative decrease of 6.7%), and the full configuration further reduced it to 2.176 (a relative decrease of 9.6%). These results indicate that explicitly encoding direction-aware band priors in the model input yields a consistent improvement in dose prediction performance. Table 1 showed that the proposed two-stage cascaded dose prediction framework outperformed the other evaluated models. A key reason was that many prior DL approaches, despite careful architectural design, were still trained primarily with MAE or other similar mean error-related objectives. Such objectives encourage learning a mapping towards the conditional mean or median, which can produce overly smooth dose distributions and attenuate boundary details [ 36 , 37 ]. This is undesirable when high-gradient transitions are clinically critical. Nguyen et al. explored differentiable DVH-based and adversarially inspired losses, demonstrating their utility for training dose models and generating Pareto-optimal radiotherapy dose distributions [ 38 ]. In our framework, MAE was retained in the first stage to learn the global dose distribution, while the second stage introduces a composite objective ( L band + L grad + L DVH ) to refine band-region behaviour, local gradient transitions, and DVH consistency. This design aims to improve prediction accuracy while preserving global coherence. From a task-decomposition perspective, the cascade allows the model to learn the low-frequency/global dose pattern first and then correct high-frequency/local discrepancies, which can reduce the learning difficulty of the refinement stage and improve training robustness. Teng et al. adopted a related strategy by decomposing a global coarse dose into multiple field doses, refining them on a per-field basis, and then aggregating to obtain the final 3D dose distribution, achieving improved predictive performance [ 13 ]. In addition, our cohort was comparatively large, and organ-wise dose distributions span a broad range (Supplementary Fig. S3), which provided richer supervision for the refinement stage during error correction. Several studies had explored how to translate predicted dose into actionable planning. Shen et al. discretized predicted dose in specific regions to construct optimization objectives and used a two-step optimization scheme [ 20 ]. Choi et al. automatically extracted structure-specific objective values and weights from predicted dose and wrote them into a TPS to generate plans [ 39 ]. Church et al. combined dose mimicking with a residual U-Net to predict deliverability-related elements and converted them into DICOM-RT plans for evaluation [ 40 ]. Relative to these studies, the present work emphasized three aspects. First, it targets the VMAT setting, where dynamic modulation and continuous rotation impose directionality and high-gradient characteristics that were not well represented by anatomy alone. Second, the predicted dose was transformed through TPS scripting into Monaco-executable individualized constraints and iteratively optimized to produce the final plan, rather than remaining solely at the level of a predicted distribution. Third, we incorporated gamma analysis to support plan deliverability. Moreover, Table 3 indicated that, with target coverage maintained, DVH endpoints for most key OARs decreased overall. Notably, Dmean and V30% for the bilateral femoral heads increased by approximately 1 Gy and 5%, respectively. This might reflect a redistribution of dose contributions across directions during optimization. Additionally, femoral heads were relatively small structures and their DVH metrics could be more sensitive to localized dose variations. Importantly, femoral head doses in the E2E auto-plans still remained within clinically acceptable ranges. In future work, we would further refine the optimization strategy for small-volume bony structures such as the femoral heads, introducing additional constraints and/or objective rebalancing to prevent metric drift and ideally achieve neutral or reduced values. This study had several limitations. First, the diversity of the cohort and prescription configurations remains limited; generalizability across different prescriptions, planning preferences, and broader patient populations required further validation. We plan to expand the sample size and include additional prescription schemes in CC cohorts. Second, plan quality is partly contingent on the accuracy of dose prediction; we will continue to improve the predictive network and loss design to enhance robustness in steep-gradient and target–OAR adjacent regions. Third, although we achieved E2E auto-planning within the Monaco TPS, cross-TPS portability has not been established and will require dedicated interface development and consistency assessments. Finally, our experiments focused on a single tumor site; validation across multiple tumor sites is essential. A multi-site auto-planning framework, adapted to varying clinical requirements, is expected to offer broader applicability and improved robustness, and this will facilitate more reliable integration of the proposed workflow into clinical radiotherapy practice. Conclusion We propose a closed-loop framework that converts anatomy into deliverable VMAT plans by integrating direction-aware beam-band priors, DVH-guided learning, and Monaco TPS scripting for optimization. The method preserves target coverage, improves OARs’ DVH endpoints, and demonstrates deliverability with gamma verification. This approach reduces manual trial-and-error and generates deliverable plans, which has the potential to substantially improve the efficiency and consistency of CC radiotherapy planning, with promising prospects for rapid auto-plan generation in routine practice. Abbreviations IMRT Intensity-modulated radiotherapy VMAT Volumetric modulated arc therapy CC Cervical cancer OARs Organs at risk ATP Automated treatment planning KBP Knowledge-based planning DVH Dose-volume histogram 3D Three-dimensional DL Deep learning TPS Treatment planning system E2E End-to-end PTV Plan target volume MAE Mean absolute error ROI Region of interest snDVH score Scale-normalized DVH score. Declarations Ethics approval and consent to participate: This study was approved by the Ethics Committee (Institutional Review Board) of Harbin Medical University Cancer Hospital, Harbin, China (ECCR No. KY2023-83, 2023-12-18). The study was conducted in accordance with the Declaration of Helsinki. The requirement for informed consent was waived by the Ethics Committee/IRB due to the retrospective nature of the study. Consent for publication: Not applicable. Competing interests: The authors declare that they have no competing interests. Funding: This research was supported partially by the Joint Fund Cultivation Project of Heilongjiang Provincial Natural Science Foundation (No. PL2025A001), the National Natural Science Foundation of China (12375341). Author Contribution Boda Ning: Software, Data Curation, Writing - Original Draft, Validation, Formal analysis, Investigation. Xiuyan Liang: Data Curation, Formal analysis, Writing - Review & Editing. Zhenguo Cui: Data Curation, Conceptualization. Yingfa Li: Data Curation, Software. Qi Liu: Formal analysis. Shuaining Ma: Resources. Xiting Chen: Investigation. Shanshan Yang: Data Curation, Methodology, Supervision. Yanling Bai: Conceptualization, Methodology, Writing - Review & Editing, Supervision, Funding acquisition. Deyang Yu: Conceptualization, Methodology, Writing - Review & Editing, Supervision, Project administration, Funding acquisition. Acknowledgements: Not applicable. Clinical trial number : Not applicable. Data Availability The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request. References Bertelsen A, Hansen CR, Johansen J, Brink C. Single Arc Volumetric Modulated Arc Therapy of head and neck cancer. Radiother Oncol. 2010;95:142–8. https://doi.org/10.1016/j.radonc.2010.01.011 . Lee N, Puri DR, Blanco AI, Chao KS. Intensity-modulated radiation therapy in head and neck cancers: an update. Head Neck. 2007;29:387–400. https://doi.org/10.1002/hed.20332 . Verginadis II, Citrin DE, Ky B, Feigenberg SJ, Georgakilas AG, Hill-Kayser CE, et al. Radiotherapy toxicities: mechanisms, management, and future directions. Lancet. 2025;405:338–52. https://doi.org/10.1016/s0140-6736(24)02319-5 . Ge Y, Wu QJ. Knowledge-based planning for intensity-modulated radiation therapy: A review of data-driven approaches. Med Phys. 2019;46:2760–75. https://doi.org/10.1002/mp.13526 . Jia X, Rong Y, Wu Q, Cardenas CE, Court LE, Hrinivich WT, et al. NRG Oncology Assessment of Artificial Intelligence for Automatic Treatment Planning in Radiation Therapy Clinical Trials: Present and Future. Int J Radiat Oncol Biol Phys. 2025;123:282–95. https://doi.org/10.1016/j.ijrobp.2025.03.045 . Nguyen D, Lin MH, Sher D, Lu W, Jia X, Jiang S. Advances in Automated Treatment Planning. Semin Radiat Oncol. 2022;32:343–50. https://doi.org/10.1016/j.semradonc.2022.06.004 . Tol JP, Delaney AR, Dahele M, Slotman BJ, Verbakel WF. Evaluation of a knowledge-based planning solution for head and neck cancer. Int J Radiat Oncol Biol Phys. 2015;91:612–20. https://doi.org/10.1016/j.ijrobp.2014.11.014 . Hussein M, South CP, Barry MA, Adams EJ, Jordan TJ, Stewart AJ, et al. Clinical validation and benchmarking of knowledge-based IMRT and VMAT treatment planning in pelvic anatomy. Radiother Oncol. 2016;120:473–9. https://doi.org/10.1016/j.radonc.2016.06.022 . Li N, Carmona R, Sirak I, Kasaova L, Followill D, Michalski J, et al. Highly Efficient Training, Refinement, and Validation of a Knowledge-based Planning Quality-Control System for Radiation Therapy Clinical Trials. Int J Radiat Oncol Biol Phys. 2017;97:164–72. https://doi.org/10.1016/j.ijrobp.2016.10.005 . Wang M, Zhang Q, Lam S, Cai J, Yang R. A Review on Application of Deep Learning Algorithms in External Beam Radiotherapy Automated Treatment Planning. Front Oncol. 2020;10:580919. https://doi.org/10.3389/fonc.2020.580919 . Leino A, Heikkilä J, Virén T, Honkanen JTJ, Seppälä J, Korkalainen H. Deep learning-based prediction of the dose-volume histograms for volumetric modulated arc therapy of left-sided breast cancer. Med Phys. 2024;51:7986–97. https://doi.org/10.1002/mp.17410 . Jiao Z, Peng X, Wang Y, Xiao J, Nie D, Wu X, et al. TransDose: Transformer-based radiotherapy dose prediction from CT images guided by super-pixel-level GCN classification. Med Image Anal. 2023;89:102902. https://doi.org/10.1016/j.media.2023.102902 . Teng L, Wang B, Xu X, Zhang J, Mei L, Feng Q, et al. Beam-wise dose composition learning for head and neck cancer dose prediction in radiotherapy. Med Image Anal. 2024;92:103045. https://doi.org/10.1016/j.media.2023.103045 . Xie H, Zhang H, Chen Z, Tan T. Precision dose prediction for breast cancer patients undergoing IMRT: The Swin-UMamba-Channel Model. Comput Med Imaging Graph. 2024;116:102409. https://doi.org/10.1016/j.compmedimag.2024.102409 . Zhang Y, Li C, Zhong L, Chen Z, Yang W, Wang X. DoseDiff: Distance-Aware Diffusion Model for Dose Prediction in Radiotherapy. IEEE Trans Med Imaging. 2024;43:3621–33. https://doi.org/10.1109/tmi.2024.3383423 . Cao X, Zhao Y, Li S, Zhang F, Yang Z, Yang X. Beam field guided diffusion model for liver cancer radiotherapy dose distribution prediction. Med Phys. 2025;52:e17989. https://doi.org/10.1002/mp.17989 . Chang HH, Harms J, Cardan RA, Fiveash JB, Popple RA, Cardenas CE, nnDoseNet. Intuitive and flexible deep learning framework to train and evaluate radiotherapy dose prediction models. Comput Biol Med. 2025;198:111237. https://doi.org/10.1016/j.compbiomed.2025.111237 . Oppitz H, Eckl M, Siebenlist K, Boda-Heggemann J, Abo-Madyan Y, Giordano FA, et al. Knowledge-based automated radiation therapy treatment planning utilizing dose prediction with a 2.5D-U-Net. Phys Med. 2025;139:105199. https://doi.org/10.1016/j.ejmp.2025.105199 . Xie K, Hua Q, Gao L, Sun J, Lin T, Sui J, et al. SegMambaDP: a long-range sequential modeling mamba network for 3D dose prediction in radiotherapy. BMC Med Imaging. 2025;25:527. https://doi.org/10.1186/s12880-025-02055-8 . Shen Y, Tang X, Lin S, Jin X, Ding J, Shao M. Automatic dose prediction using deep learning and plan optimization with finite-element control for intensity modulated radiation therapy. Med Phys. 2024;51:545–55. https://doi.org/10.1002/mp.16743 . Xiao Y, Tanaka S, Kadoya N, Sato K, Kimura Y, Umezawa R, et al. Evaluation of deliverable artificial intelligence-based automated volumetric arc radiation therapy planning for whole pelvic radiation in gynecologic cancer. Sci Rep. 2025;15:15219. https://doi.org/10.1038/s41598-025-99717-y . Huang W, Liu T, Shen Y, Xiang Z, Wang D, Fu W, et al. Automatic radiotherapy planning for deliverable plans using deep learning dose prediction and dose rings optimization in cervical cancer. J Appl Clin Med Phys. 2025;26:e70353. https://doi.org/10.1002/acm2.70353 . Çiçek Ö, Abdulkadir A, Lienkamp SS, Brox T, Ronneberger O. 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation. Springer; 2016. pp. 424–32. Zhang Z, Liu Q, Wang Y. Road Extraction by Deep Residual U-Net. IEEE Geosci Remote Sens Lett. 2018;15:749–53. https://doi.org/10.1109/LGRS.2018.2802944 . Milletari F, Navab N, Ahmadi S-A. V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation. Fourth International Conference on 3D Vision (3DV): IEEE; 2016. pp. 565 – 71. Liu S, Zhang J, Li T, Yan H, Liu J, Technical Note. A cascade 3D U-Net for dose prediction in radiotherapy. Med Phys. 2021;48:5574–82. https://doi.org/10.1002/mp.15034 . Kearney V, Chan JW, Haaf S, Descovich M, Solberg TD. DoseNet: a volumetric dose prediction algorithm using 3D fully-convolutional neural networks. Phys Med Biol. 2018;63:235022. https://doi.org/10.1088/1361-6560/aaef74 . Nguyen D, Jia X, Sher D, Lin MH, Iqbal Z, Liu H, et al. 3D radiotherapy dose prediction on head and neck cancer patients with a hierarchically densely connected U-net deep learning architecture. Phys Med Biol. 2019;64:065020. https://doi.org/10.1088/1361-6560/ab039b . Osman AFI, Tamam NM, Yousif YAM. A comparative study of deep learning-based knowledge-based planning methods for 3D dose distribution prediction of head and neck. J Appl Clin Med Phys. 2023;24:e14015. https://doi.org/10.1002/acm2.14015 . Babier A, Zhang B, Mahmood R, Moore KL, Purdie TG, McNiven AL, et al. OpenKBP: The open-access knowledge-based planning grand challenge and dataset. Med Phys. 2021;48:5549–61. https://doi.org/10.1002/mp.14845 . Hémon C, Texier B, Lafond C, Nunes JC, Barateau A. Towards trustworthy AI in radiotherapy: a comprehensive review of uncertainty-aware techniques. Phys Med Biol. 2025;71. https://doi.org/10.1088/1361-6560/ae2a9f . Yue M, Xue X, Wang Z, Lambo RL, Zhao W, Xie Y, et al. Dose prediction via distance-guided deep learning: Initial development for nasopharyngeal carcinoma radiotherapy. Radiother Oncol. 2022;170:198–204. https://doi.org/10.1016/j.radonc.2022.03.012 . Barragán-Montero AM, Nguyen D, Lu W, Lin MH, Norouzi-Kandalan R, Geets X, et al. Three-dimensional dose prediction for lung IMRT patients with deep neural networks: robust learning from heterogeneous beam configurations. Med Phys. 2019;46:3679–91. https://doi.org/10.1002/mp.13597 . Sun Z, Xia X, Fan J, Zhao J, Zhang K, Wang J, et al. A hybrid optimization strategy for deliverable intensity-modulated radiotherapy plan generation using deep learning-based dose prediction. Med Phys. 2022;49:1344–56. https://doi.org/10.1002/mp.15462 . Xiong T, Ren G, Chen Z, Huang YH, Ma Z, Li Z, et al. A generalizable dose prediction model for automatic radiotherapy planning based on physics-informed priors and large-kernel convolutions. Med Phys. 2026;53:e70272. https://doi.org/10.1002/mp.70272 . Yang Q, Yan P, Zhang Y, Yu H, Shi Y, Mou X, et al. Low-Dose CT Image Denoising Using a Generative Adversarial Network With Wasserstein Distance and Perceptual Loss. IEEE Trans Med Imaging. 2018;37:1348–57. https://doi.org/10.1109/tmi.2018.2827462 . Han M, Shim H, Baek J. Low-dose CT denoising via convolutional neural network with an observer loss function. Med Phys. 2021;48:5727–42. https://doi.org/10.1002/mp.15161 . Nguyen D, McBeth R, Sadeghnejad Barkousaraie A, Bohara G, Shen C, Jia X, et al. Incorporating human and learned domain knowledge into training deep neural networks: A differentiable dose-volume histogram and adversarial inspired framework for generating Pareto optimal dose distributions in radiation therapy. Med Phys. 2020;47:837–49. https://doi.org/10.1002/mp.13955 . Choi B, Shrestha DK, Attia A, Stish BJ, Leenstra J, Rwigema JC, et al. Deep learning-based dose prediction for prostate cancer with empty bladder protocol: a framework for efficient and personalized radiotherapy planning. Front Oncol. 2025;15:1690416. https://doi.org/10.3389/fonc.2025.1690416 . Church C, Yap M, Bessrour M, Lamey M, Granville D. Automated plan generation for prostate radiotherapy patients using deep learning and scripted optimization. Phys Imaging Radiat Oncol. 2024;32:100641. https://doi.org/10.1016/j.phro.2024.100641 . Additional Declarations No competing interests reported. Supplementary Files SupplementaryMaterials.docx Cite Share Download PDF Status: Published Journal Publication published 14 Apr, 2026 Read the published version in Radiation Oncology → Version 1 posted Editorial decision: Revision requested 26 Feb, 2026 Reviews received at journal 26 Feb, 2026 Reviewers agreed at journal 09 Feb, 2026 Reviewers invited by journal 09 Feb, 2026 Editor assigned by journal 04 Feb, 2026 Submission checks completed at journal 04 Feb, 2026 First submitted to journal 31 Jan, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8749620","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":588104902,"identity":"b2d79bcb-53ae-414b-a029-eb03feefdde6","order_by":0,"name":"Boda Ning","email":"","orcid":"","institution":"Harbin Medical University Cancer Hospital","correspondingAuthor":false,"prefix":"","firstName":"Boda","middleName":"","lastName":"Ning","suffix":""},{"id":588104904,"identity":"09156b34-3380-4df0-a7c7-033beabc1441","order_by":1,"name":"Xiuyan Liang","email":"","orcid":"","institution":"Harbin Medical University Cancer Hospital","correspondingAuthor":false,"prefix":"","firstName":"Xiuyan","middleName":"","lastName":"Liang","suffix":""},{"id":588104905,"identity":"c188cca5-24b6-4880-baf3-29c5e67ef092","order_by":2,"name":"Zhenguo Cui","email":"","orcid":"","institution":"Harbin Medical University Cancer Hospital","correspondingAuthor":false,"prefix":"","firstName":"Zhenguo","middleName":"","lastName":"Cui","suffix":""},{"id":588104906,"identity":"77a1ee54-238d-4a75-99c9-cb6ae6c1895d","order_by":3,"name":"Yingfa Li","email":"","orcid":"","institution":"Capital Medical University","correspondingAuthor":false,"prefix":"","firstName":"Yingfa","middleName":"","lastName":"Li","suffix":""},{"id":588104907,"identity":"881512f5-b868-4839-828c-5040bf533fcb","order_by":4,"name":"Qi Liu","email":"","orcid":"","institution":"Harbin Medical University Cancer Hospital","correspondingAuthor":false,"prefix":"","firstName":"Qi","middleName":"","lastName":"Liu","suffix":""},{"id":588104908,"identity":"51c51cc7-8438-4493-93c9-56b67f5e508d","order_by":5,"name":"Shuaining Ma","email":"","orcid":"","institution":"Harbin Medical University Cancer Hospital","correspondingAuthor":false,"prefix":"","firstName":"Shuaining","middleName":"","lastName":"Ma","suffix":""},{"id":588104909,"identity":"44849529-75d4-4a1c-9742-0aaba8640bc0","order_by":6,"name":"Xiting Chen","email":"","orcid":"","institution":"Harbin Medical University Cancer Hospital","correspondingAuthor":false,"prefix":"","firstName":"Xiting","middleName":"","lastName":"Chen","suffix":""},{"id":588104910,"identity":"3732cf20-757c-4d97-9421-b8aa2016acdd","order_by":7,"name":"Shanshan Yang","email":"","orcid":"","institution":"Harbin Medical University Cancer Hospital","correspondingAuthor":false,"prefix":"","firstName":"Shanshan","middleName":"","lastName":"Yang","suffix":""},{"id":588104911,"identity":"f2c8546b-9ca9-42ae-9485-789f38d49a23","order_by":8,"name":"Yanling Bai","email":"","orcid":"","institution":"Harbin Medical University Cancer Hospital","correspondingAuthor":false,"prefix":"","firstName":"Yanling","middleName":"","lastName":"Bai","suffix":""},{"id":588104912,"identity":"86f2ae2b-d203-43d8-82eb-ddd3a3f04226","order_by":9,"name":"Deyang Yu","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA5klEQVRIiWNgGAWjYHACAzDJL8HABuEfIFaL5AyStRjcIFaLwY3kjY95au7Ybb7d++zRzTYGOb4bCYyfC/BqSSs25jn2LHnbnePmxrltDMaSNxKYpWfg1ZJjJs3DdjjZ7EYamzRQS+KGGwlszDz4tZj/5vl3ONl4BkRLPTFazJh52w7bGUhAtCQYENIieeZZseTcvsMJEjfS2I1zzkkYzjzzsFkanxa+48kbP7z5dtieH+iwxzllNvJAkYOf8WlROMDAwARUkNgA4UsAMWMDHg0MDPJAacYfDAz2eFWNglEwCkbByAYAoAJOtgmxQPMAAAAASUVORK5CYII=","orcid":"","institution":"Harbin Medical University Cancer Hospital","correspondingAuthor":true,"prefix":"","firstName":"Deyang","middleName":"","lastName":"Yu","suffix":""}],"badges":[],"createdAt":"2026-01-31 11:56:08","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8749620/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8749620/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1186/s13014-026-02842-9","type":"published","date":"2026-04-14T15:57:35+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":102440983,"identity":"d8cda378-16ec-42ce-84fa-cd2a80d0ee36","added_by":"auto","created_at":"2026-02-11 16:53:05","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":8170341,"visible":true,"origin":"","legend":"\u003cp\u003eOverview of the proposed E2E VMAT auto-planning workflow. (a) Beam band mask generation. Four directional beam band masks (0°, 45°, 90°, and 135°) were constructed to approximate VMAT entrance directionality. (b) Two-stage dose prediction framework. A Stage-1 network predicts a coarse dose from CT and structure masks, followed by a Stage-2 refinement network conditioned on beam band masks and optimized with a composite loss. (c) TPS-based auto-planning. DVH endpoints from the refined dose define patient-specific optimization targets, which are converted via an in-house Monaco scripting module into executable objectives and iteratively optimized to generate a deliverable plan.\u003c/p\u003e","description":"","filename":"figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-8749620/v1/14cdab0ef1d9489258bd6d13.png"},{"id":102441169,"identity":"ab9973ff-1512-4055-9cd8-b47fb0bad7ad","added_by":"auto","created_at":"2026-02-11 16:54:00","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":3517249,"visible":true,"origin":"","legend":"\u003cp\u003eBox plot of critical DVH metrics between the refined predicted dose (yellow) and the clinical plan dose (blue) across PTV and OARs.\u003c/p\u003e","description":"","filename":"figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-8749620/v1/af14b5a1a4859340f6963cf3.png"},{"id":102441009,"identity":"32f1755e-9064-49b0-a74e-dc9a3d032495","added_by":"auto","created_at":"2026-02-11 16:53:19","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":10840599,"visible":true,"origin":"","legend":"\u003cp\u003eRepresentative case comparisons between the clinical reference plan and the E2E auto-plan. (a)-(c) For three cases, axial/coronal/sagittal dose distributions are shown for the ground-truth clinical plan (GT), the TPS-generated auto-plan (Auto), and the voxel-wise difference map (Diff; Auto-GT). (d)-(f) Corresponding DVH curves for PTV and OARs. Solid lines denote GT and dashed lines denote Auto.\u003c/p\u003e","description":"","filename":"figure3.png","url":"https://assets-eu.researchsquare.com/files/rs-8749620/v1/16e3ed65c0c0147d8103a6e1.png"},{"id":107350756,"identity":"0f1d0a24-b967-47d7-9777-ba71ebf6953a","added_by":"auto","created_at":"2026-04-20 16:03:17","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":21584336,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8749620/v1/5f1f692b-6add-4a27-85d4-39824a67d6ab.pdf"},{"id":102440984,"identity":"2ba13758-4deb-4336-9b01-5908aa0a32e4","added_by":"auto","created_at":"2026-02-11 16:53:05","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":668913,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryMaterials.docx","url":"https://assets-eu.researchsquare.com/files/rs-8749620/v1/45de9b6352c59696d9c8eddb.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"A deep learning-driven automated treatment planning framework for patient treated with radiotherapy in cervical cancer","fulltext":[{"header":"Introduction","content":"\u003cp\u003eIntensity-modulated radiotherapy (IMRT) and volumetric modulated arc therapy (VMAT) are widely used for cervical cancer (CC) because they enable highly conformal target coverage while improving sparing of surrounding organs at risk (OARs) through inverse planning and modulation [\u003cspan additionalcitationids=\"CR2\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. In routine treatment, prescription and clinical priorities are determined by radiation oncologists, whereas plan generation relies on medical physicists to iteratively adjust optimization objectives and parameters until institutional and guideline-based criteria are satisfied [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. However, both the quality and efficiency of manual planning largely depend on planner experience, and plan quality can vary across individual planners, posing a major challenge to workflow efficiency in busy clinics [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eTo improve efficiency and reduce variability, automated treatment planning (ATP) has been extensively studied [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. Knowledge-based planning (KBP) is widely used to improve planning efficiency and consistency by modeling planning objectives to guide the treatment planning workflow, thereby minimizing planner intervention [\u003cspan additionalcitationids=\"CR8\" citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Dose-volume histogram (DVH)-based prediction is one established KBP approach, in which predicted DVH endpoints are used as planning goals; however, accuracy can be limited because the detailed space information of three-dimensional (3D) dose distribution is not explicitly modeled [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. In parallel, deep learning (DL)\u0026ndash;based dose prediction has advanced rapidly, enabling fast generation of patient-specific 3D dose distributions from planning CT and structure masks after training on prior high-quality plans [\u003cspan additionalcitationids=\"CR13 CR14 CR15 CR16 CR17 CR18\" citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. Nevertheless, even highly accurate predicted dose distributions are not inherently deliverable, and the gap between \u0026ldquo;dose prediction\u0026rdquo; and \u0026ldquo;clinically executable plans\u0026rdquo; remains a key barrier to achieve a closed-loop workflow.\u003c/p\u003e \u003cp\u003eRecent studies have begun to bridge this gap by converting DL-predicted dose distributions into deliverable IMRT or VMAT plans [\u003cspan additionalcitationids=\"CR21\" citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. However, whether predicted doses can be translated robustly into treatment planning system (TPS) objectives that converge reliably to high-quality, deliverable plans requires further validation. This need is most evident in steep dose fall-off regions and at target\u0026ndash;OAR interfaces, where anatomy-only prediction may not adequately reflect the directionality and gradient constraints imposed by inverse planning, limiting robust translation into stable TPS objectives and consistent plan quality. We therefore propose a two-stage cascaded DL framework that incorporates directional beam band priors and a DVH-oriented composite loss, and converts the refined dose into patient-specific, Monaco-executable objectives via in-house TPS scripting for automated inverse optimization. This study aimed to establish and validate an end-to-end (E2E) VMAT auto-planning workflow for CC and to assess dose prediction accuracy, plan quality, and deliverability.\u003c/p\u003e"},{"header":"Materials and methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003ePatients and clinical plans\u003c/h2\u003e \u003cp\u003eThis retrospective study was approved by the Institutional Review Board of Institution A. A total of 458 CC patients treated with full-arc VMAT between 2023 and 2025 were enrolled according to the exclusion criteria (Fig. \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e) and randomly assigned to training, validation, and test cohorts at an 8:1:1 ratio. For each patient, a simulation CT was acquired on a Philips Brilliance Big Bore scanner (Philips Healthcare, Best, The Netherlands) with a slice thickness of 5 mm. Target volumes and OARs were contoured by radiation oncologists in accordance with Radiation Therapy Oncology Group guidelines using AccuContour (v3.2; Manteia Technologies Co., Ltd.). All plans were prescribed 45 Gy in 25 fractions and were generated under the institutional planning protocol by medical physicists, with target coverage ensured and OAR sparing prioritized. Plan optimization and final dose calculation were performed in Monaco (clinical version 6.2.2; Elekta AB, Stockholm, Sweden) TPS. The planning objectives and dose constraints for targets and OARs are summarized in Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eConstruction of beam band masks\u003c/h3\u003e\n\u003cp\u003eAn analysis of full-arc VMAT plans for CC patients revealed that the vast majority of dose delivery pathways were predominantly concentrated along several beam-entry directions (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003ea(1)). This directional concentration may reflect inverse optimization mechanism, where OARs dose constraints encouraged beam delivery to avoid traversing critical organs or, when avoidance is not feasible, to lower the beam field weight from those directions. Consequently, four directional beam-band masks were generated on each CT slice. Each beam band mask consisted of two parallel boundaries that are tangent to the plan target volume (PTV) on the respective slice (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003ea(2)). These masks were used as auxiliary inputs to the Stage 2 refinement network.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e\n\u003ch3\u003eDose prediction model architecture\u003c/h3\u003e\n\u003cp\u003eBefore DL network training, CT images must be preprocessed to standardize input dimensions (Appendix A). We developed a cascaded DL framework for E2E prediction of patient-specific 3D dose distributions, and the workflow is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eb. The framework comprises two consecutive stages. In Stage 1, a 3D U-Net was trained using the planning CT together with binary structure masks of the PTV, selected OARs, and the external body contour as multi-channel inputs [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]. The network outputs a coarse 3D dose map and is supervised against the reference dose using a mean absolute error (MAE) loss. This stage captured the global dose morphology from patient anatomy and structure masks.\u003c/p\u003e \u003cp\u003eThe second stage introduces a refinement step conditioned on directional beam band information. Specifically, four directional beam band masks were generated to approximate predominant entrance paths in clinical VMAT plans and used as auxiliary geometric priors. Subsequently, the coarse dose from Stage 1 was concatenated with the CT, structure masks, and the four band masks to form 15 input channels, and a 3D ResU-Net was trained to produce the refined predicted dose [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. The stage 2 was supervised using a composite objective comprising four loss functions, including \u003cem\u003eL\u003c/em\u003e\u003csub\u003e\u003cem\u003ebody\u003c/em\u003e\u003c/sub\u003e, \u003cem\u003eL\u003c/em\u003e\u003csub\u003e\u003cem\u003eband\u003c/em\u003e\u003c/sub\u003e, \u003cem\u003eL\u003c/em\u003e\u003csub\u003e\u003cem\u003egrad\u003c/em\u003e\u003c/sub\u003e and \u003cem\u003eL\u003c/em\u003e\u003csub\u003e\u003cem\u003eDVH\u003c/em\u003e\u003c/sub\u003e, which jointly encourage global accuracy, band-region fidelity, gradient consistency, and DVH-related agreement. Full definitions of each loss function were provided in Appendix B. The second-stage overall loss was defined as a weighted sum:\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:\\begin{array}{c}{\\text{L}}_{\\text{total}}\\text{}\\text{=}{\\text{}\\lambda}_{\\text{body}\\text{}}{\\text{L}}_{\\text{body}}\\text{}\\text{+}\\text{}{\\lambda}_{\\text{band}\\text{}}{\\text{L}}_{\\text{band}}\\text{}\\text{+}\\text{}{\\lambda}_{\\text{grad}\\text{}}{\\text{L}}_{\\text{grad}}\\text{}\\text{+}{\\text{}\\lambda}_{\\text{DVH}\\text{}}{\\text{L}}_{\\text{DVH}}\\#\\text{(}\\text{1}\\text{)}\\end{array}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eDetailed network architecture and the training environment and strategy were provided in Appendix C. The full code used in this work is openly available at GitHub (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/Rick-Ds/Two_stage_E2E_DosePrediction\u003c/span\u003e\u003cspan address=\"https://github.com/Rick-Ds/Two_stage_E2E_DosePrediction\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e).\u003c/p\u003e\n\u003ch3\u003eAuto-plan generation\u003c/h3\u003e\n\u003cp\u003eTo evaluate whether the refined predicted dose could be translated into deliverable treatment plans of comparable quality to the clinical reference, we performed a TPS script\u0026ndash;based auto-planning study, with the workflow shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003ec. All procedures were implemented in the Monaco TPS. An in-house script first configured the beam geometry and initialized an automated plan. DVH endpoints were extracted from the refined predicted dose for each region of interest (ROI) and defined as the expected target DVH metrics (\u003cem\u003eM\u003c/em\u003e\u003csub\u003e\u003cem\u003e0\u003c/em\u003e\u003c/sub\u003e). Based on \u003cem\u003eM\u003c/em\u003e\u003csub\u003e\u003cem\u003e0\u003c/em\u003e\u003c/sub\u003e, plan-specific optimization objectives and constraints were assembled and converted into TPS-compatible parameters to drive inverse optimization. After each optimization finished, DVH metrics of the current plan (\u003cem\u003eM\u003c/em\u003e\u003csub\u003e\u003cem\u003ec\u003c/em\u003e\u003c/sub\u003e) were compared with the targets \u003cem\u003eM\u003c/em\u003e\u003csub\u003e\u003cem\u003e0\u003c/em\u003e\u003c/sub\u003e. If unmet, only the corresponding objectives were updated for the next iteration: PTV or OARs dose constraint values were increased or decreased by 2%, followed by re-optimization. We limited the maximum number of iterations to no more than five. If the objective cannot be achieved, the TPS scripting module retains an iteration result closest to the goal. A summary of the constraint functions was provided in Table S2.\u003c/p\u003e\n\u003ch3\u003eEvaluation and statistical analysis\u003c/h3\u003e\n\u003cp\u003eTo demonstrate the superiority of the proposed E2E cascaded DL framework for dose prediction, we benchmarked it against top-performing OpenKBP models and subsequent state-of-the-art architectures, including VNet, C3D, DoseNet, HD-UNet, Attention-aware 3D U-Net, and TransDose [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan additionalcitationids=\"CR26 CR27 CR28\" citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e]. For fairness, all methods were trained and evaluated on the same dataset with an identical training protocol. Performance was quantified using the official OpenKBP Dose score and DVH score, together with a custom scale-normalized DVH score (snDVH score) [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. Detailed definitions were provided in Appendix D. Beyond dose prediction, we compared auto-plans with the corresponding clinical reference plans using dose distributions and DVH endpoints. Deliverability was evaluated via linac-based verification of the final auto-plans, and quality assurance results were recorded. To compare our method with reference approaches, paired sample t-tests were performed in SPSS (version 22.0). Statistical significance was defined as a two-tailed p value\u0026thinsp;\u0026lt;\u0026thinsp;0.05.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eAblation experiments\u003c/h2\u003e \u003cp\u003eTo examine the interpretability and practical impact of the chosen number of beam bands, we generated band masks with different angular counts and performed dedicated ablation analyses. In addition, to assess the contribution of each loss component, we used the cascaded network as the baseline framework and conducted a series of ablation experiments in which key terms were added sequentially: (1) the global dose consistency loss \u003cem\u003eL\u003c/em\u003e\u003csub\u003e\u003cem\u003ebody\u003c/em\u003e\u003c/sub\u003e; (2) the band-focused dose loss \u003cem\u003eL\u003c/em\u003e\u003csub\u003e\u003cem\u003eband\u003c/em\u003e\u003c/sub\u003e; (3) the band-gradient loss \u003cem\u003eL\u003c/em\u003e\u003csub\u003e\u003cem\u003egrad\u003c/em\u003e\u003c/sub\u003e; and (4) the DVH-based numerical loss \u003cem\u003eL\u003c/em\u003e\u003csub\u003e\u003cem\u003eDVH\u003c/em\u003e\u003c/sub\u003e.​\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cp\u003eIn the beam-band ablation, a distinct elbow was observed at k\u0026thinsp;=\u0026thinsp;4. Dose coverage within band-overlap regions increased from 0.39 at k\u0026thinsp;=\u0026thinsp;1 to 0.95 at k\u0026thinsp;=\u0026thinsp;4, whereas further increases in k produced only marginal gains of about 0.01 per additional band. Peak GPU memory increased approximately linearly with k, and k\u0026thinsp;=\u0026thinsp;4 was therefore adopted as a practical trade-off for subsequent experiments (Supplementary Fig. S2).\u003c/p\u003e \u003cp\u003eBenchmarking against six representative DL dose prediction models (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e), the proposed method achieved the best overall performance across Dose score, DVH score, and snDVH score, reaching 2.114\u0026thinsp;\u0026plusmn;\u0026thinsp;0.218 Gy, 1.194\u0026thinsp;\u0026plusmn;\u0026thinsp;0.295 Gy, and 2.027\u0026thinsp;\u0026plusmn;\u0026thinsp;0.586, respectively. In addition, we assessed the incremental contribution of the four loss functions. As the key components were introduced sequentially, all three metrics showed an overall downward trend in Table \u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e; with the final composite loss, improvements were statistically significant across metrics except that the change in Dose score before versus after adding \u003cem\u003eL\u003c/em\u003e\u003csub\u003e\u003cem\u003eDVH\u003c/em\u003e\u003c/sub\u003e did not reach significance. Notably, adding \u003cem\u003eL\u003c/em\u003e\u003csub\u003e\u003cem\u003eband\u003c/em\u003e\u003c/sub\u003e and \u003cem\u003eL\u003c/em\u003e\u003csub\u003e\u003cem\u003egrad\u003c/em\u003e\u003c/sub\u003e reduced the Dose score from 2.407 to 2.176, indicating improved global and band-focused accuracy. After \u003cem\u003eL\u003c/em\u003e\u003csub\u003e\u003cem\u003eDVH\u003c/em\u003e\u003c/sub\u003e was incorporated, DVH score and snDVH score were further reduced by approximately 22% (from 1.526 to 1.194) and 18% (from 2.475 to 2.027), respectively, suggesting that the DVH-based constraint improved ROI-relevant endpoint agreement. In contrast, the change in Dose score after adding \u003cem\u003eL\u003c/em\u003e\u003csub\u003e\u003cem\u003eDVH\u003c/em\u003e\u003c/sub\u003e (2.176 to 2.114) was modest and did not show a clear additional gain. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e further illustrates the prediction accuracy of the proposed method for DVH metrics across the relevant ROIs. Overall, the predicted and reference dose distributions showed close agreement for PTV indices including D\u003csub\u003emax\u003c/sub\u003e, D\u003csub\u003emean\u003c/sub\u003e, D\u003csub\u003e2%\u003c/sub\u003e, D\u003csub\u003e98%\u003c/sub\u003e, V\u003csub\u003e95%\u003c/sub\u003e and HI, and for OAR-related endpoints including D\u003csub\u003emean\u003c/sub\u003e, V\u003csub\u003e30\u003c/sub\u003e, and the Dmax of the small intestine and spinal cord.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eQuantitative comparison between SOTA methods and our method in terms of Dose score, DVH score and snDVH Score.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colspan=\"4\" nameend=\"c4\" namest=\"c1\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eDose Score [Gy]\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eDVH Score [Gy]\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003esnDVH Score\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"4\" nameend=\"c4\" namest=\"c1\"\u003e \u003cp\u003eVNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e2.336\u0026thinsp;\u0026plusmn;\u0026thinsp;0.335*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1.917\u0026thinsp;\u0026plusmn;\u0026thinsp;0.426*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e3.068\u0026thinsp;\u0026plusmn;\u0026thinsp;0.748*\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"4\" nameend=\"c4\" namest=\"c1\"\u003e \u003cp\u003eC3D\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e2.319\u0026thinsp;\u0026plusmn;\u0026thinsp;0.259*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1.906\u0026thinsp;\u0026plusmn;\u0026thinsp;0.400*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e2.794\u0026thinsp;\u0026plusmn;\u0026thinsp;0.663*\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"4\" nameend=\"c4\" namest=\"c1\"\u003e \u003cp\u003eDoseUNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e2.227\u0026thinsp;\u0026plusmn;\u0026thinsp;0.280*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1.866\u0026thinsp;\u0026plusmn;\u0026thinsp;0.364*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e2.646\u0026thinsp;\u0026plusmn;\u0026thinsp;0.672*\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"4\" nameend=\"c4\" namest=\"c1\"\u003e \u003cp\u003eHD UNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e2.204\u0026thinsp;\u0026plusmn;\u0026thinsp;0.283*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1.651\u0026thinsp;\u0026plusmn;\u0026thinsp;0.333*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e2.535\u0026thinsp;\u0026plusmn;\u0026thinsp;0.703*\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"4\" nameend=\"c4\" namest=\"c1\"\u003e \u003cp\u003eAttention-aware 3D UNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e2.180\u0026thinsp;\u0026plusmn;\u0026thinsp;0.337\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1.565\u0026thinsp;\u0026plusmn;\u0026thinsp;0.508*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e2.582\u0026thinsp;\u0026plusmn;\u0026thinsp;0.641*\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"4\" nameend=\"c4\" namest=\"c1\"\u003e \u003cp\u003eTransDose\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e2.136\u0026thinsp;\u0026plusmn;\u0026thinsp;0.211\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1.395\u0026thinsp;\u0026plusmn;\u0026thinsp;0.293*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e2.424\u0026thinsp;\u0026plusmn;\u0026thinsp;0.666*\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"4\" nameend=\"c4\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eProposed\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e2.114\u0026thinsp;\u0026plusmn;\u0026thinsp;0.218\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e1.194\u0026thinsp;\u0026plusmn;\u0026thinsp;0.295\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u003cb\u003e2.027\u0026thinsp;\u0026plusmn;\u0026thinsp;0.586\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"7\" nameend=\"c7\" namest=\"c1\"\u003e \u003cp\u003e*: P value\u0026thinsp;\u0026lt;\u0026thinsp;0.05;\u003c/p\u003e \u003cp\u003eP-value: calculated by Paired t-test;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eExperimental results of ablation studies on our proposed method.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMethod\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDose Score [Gy]\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDVH Score [Gy]\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003esnDVH Score\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBaseline\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.407\u0026thinsp;\u0026plusmn;\u0026thinsp;0.230*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.983\u0026thinsp;\u0026plusmn;\u0026thinsp;0.337*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3.037\u0026thinsp;\u0026plusmn;\u0026thinsp;0.595*\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBaseline\u0026thinsp;+\u0026thinsp;\u003cem\u003eL\u003c/em\u003e\u003csub\u003e\u003cem\u003eband\u003c/em\u003e\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.245\u0026thinsp;\u0026plusmn;\u0026thinsp;0.262*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.641\u0026thinsp;\u0026plusmn;\u0026thinsp;0.331*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.763\u0026thinsp;\u0026plusmn;\u0026thinsp;0.562*\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBaseline\u0026thinsp;+\u0026thinsp;\u003cem\u003eL\u003c/em\u003e\u003csub\u003e\u003cem\u003eband\u003c/em\u003e\u003c/sub\u003e + \u003cem\u003eL\u003c/em\u003e\u003csub\u003e\u003cem\u003egrad\u003c/em\u003e\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.176\u0026thinsp;\u0026plusmn;\u0026thinsp;0.290\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.526\u0026thinsp;\u0026plusmn;\u0026thinsp;0.362*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.475\u0026thinsp;\u0026plusmn;\u0026thinsp;0.653*\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBaseline\u0026thinsp;+\u0026thinsp;\u003cem\u003eL\u003c/em\u003e\u003csub\u003e\u003cem\u003eband\u003c/em\u003e\u003c/sub\u003e + \u003cem\u003eL\u003c/em\u003e\u003csub\u003e\u003cem\u003egrad\u003c/em\u003e\u003c/sub\u003e + \u003cem\u003eL\u003c/em\u003e\u003csub\u003e\u003cem\u003eDVH\u003c/em\u003e\u003c/sub\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.114\u0026thinsp;\u0026plusmn;\u0026thinsp;0.218\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.194\u0026thinsp;\u0026plusmn;\u0026thinsp;0.295\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.027\u0026thinsp;\u0026plusmn;\u0026thinsp;0.586\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"4\" nameend=\"c4\" namest=\"c1\"\u003e \u003cp\u003e*: P value\u0026thinsp;\u0026lt;\u0026thinsp;0.05;\u003c/p\u003e \u003cp\u003eP-value: calculated by Paired t-test;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe quantitative evaluation of the clinical reference plans and the E2E auto-plans were presented in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. Target coverage was comparable, with no significant differences in PTV V\u003csub\u003e95%\u003c/sub\u003e (p\u0026thinsp;=\u0026thinsp;0.195) or V\u003csub\u003e100%\u003c/sub\u003e (p\u0026thinsp;=\u0026thinsp;0.650), indicating that clinical coverage requirements were maintained. Small but significant reductions were observed in PTV D\u003csub\u003e2%\u003c/sub\u003e (p\u0026thinsp;=\u0026thinsp;0.001) and HI (p\u0026thinsp;=\u0026thinsp;0.002), and they reflected a reduction in high-dose regions of target and an improvement in dose homogeneity. For OARs, there was no statistically significant difference in marrow DVH metrics between the two plan types, with similar D\u003csub\u003emean\u003c/sub\u003e and V\u003csub\u003e30Gy\u003c/sub\u003e (p\u0026thinsp;=\u0026thinsp;0.714 and p\u0026thinsp;=\u0026thinsp;0.213). Notably, all DVH metrics for the bladder, rectum, small intestine, and spinal cord were significantly reduced in the E2E auto-plans, with relative decreases of 2%\u0026ndash;35% compared with the reference plans (all p\u0026thinsp;\u0026lt;\u0026thinsp;0.05), whereas both femoral heads received higher doses, with an approximately 1 Gy increase in D\u003csub\u003emean\u003c/sub\u003e and a 4%\u0026ndash;6% increase in V\u003csub\u003e30Gy\u003c/sub\u003e compared with the reference plans. Qualitative examples were shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. Across three representative cases, the high-dose regions were largely consistent between the E2E auto-plans and reference plans, while improved conformity was observed in the 20\u0026ndash;30 Gy dose range. The corresponding DVHs showed nearly overlapping PTV coverage, accompanied by systematic shifts of the bladder, rectum, small bowel, and spinal cord curves toward lower doses within clinically relevant ranges, indicating improved OAR sparing without compromising target coverage, highlighting the dosimetric advantages of the E2E auto-planning approach.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eQuantitative evaluation of DVH metrics of PTV and OARs of GT plan and E2E auto-plan.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStructures\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDVH metrics\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eGT plan\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eE2E auto-plan\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eP-value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"7\" rowspan=\"8\"\u003e \u003cp\u003ePTV\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDmax(Gy)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e49.74\u0026thinsp;\u0026plusmn;\u0026thinsp;0.47\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e49.59\u0026thinsp;\u0026plusmn;\u0026thinsp;0.49\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.088\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDmean(Gy)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e46.48\u0026thinsp;\u0026plusmn;\u0026thinsp;0.20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e46.34\u0026thinsp;\u0026plusmn;\u0026thinsp;0.19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.000\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eV95(%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e99.70\u0026thinsp;\u0026plusmn;\u0026thinsp;0.30\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e99.80\u0026thinsp;\u0026plusmn;\u0026thinsp;0.20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.195\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eV100(%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e94.70\u0026thinsp;\u0026plusmn;\u0026thinsp;0.30\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e94.70\u0026thinsp;\u0026plusmn;\u0026thinsp;0.30\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.650\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eD98%(Gy)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e44.29\u0026thinsp;\u0026plusmn;\u0026thinsp;0.26\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e44.35\u0026thinsp;\u0026plusmn;\u0026thinsp;0.21\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.062\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eD2%(Gy)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e48.00\u0026thinsp;\u0026plusmn;\u0026thinsp;0.32\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e47.82\u0026thinsp;\u0026plusmn;\u0026thinsp;0.32\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.70\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.70\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.070\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.08\u0026thinsp;\u0026plusmn;\u0026thinsp;0.01\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.07\u0026thinsp;\u0026plusmn;\u0026thinsp;0.01\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.002\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003eBladder\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDmean(Gy)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e30.25\u0026thinsp;\u0026plusmn;\u0026thinsp;0.97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e29.58\u0026thinsp;\u0026plusmn;\u0026thinsp;1.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.014\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eV30(%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e47.30\u0026thinsp;\u0026plusmn;\u0026thinsp;2.30\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e44.80\u0026thinsp;\u0026plusmn;\u0026thinsp;5.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.002\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eV40(%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e28.90\u0026thinsp;\u0026plusmn;\u0026thinsp;2.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e24.90\u0026thinsp;\u0026plusmn;\u0026thinsp;5.20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.000\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eV45(%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e16.80\u0026thinsp;\u0026plusmn;\u0026thinsp;3.10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e13.60\u0026thinsp;\u0026plusmn;\u0026thinsp;3.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.000\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003eRectum\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDmean(Gy)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e28.20\u0026thinsp;\u0026plusmn;\u0026thinsp;1.54\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e26.81\u0026thinsp;\u0026plusmn;\u0026thinsp;1.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.000\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eV30(%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e46.70\u0026thinsp;\u0026plusmn;\u0026thinsp;3.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e40.40\u0026thinsp;\u0026plusmn;\u0026thinsp;6.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.000\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eV40(%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e20.70\u0026thinsp;\u0026plusmn;\u0026thinsp;3.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e13.40\u0026thinsp;\u0026plusmn;\u0026thinsp;4.30\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.000\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eV45(%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e5.20\u0026thinsp;\u0026plusmn;\u0026thinsp;3.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.00\u0026thinsp;\u0026plusmn;\u0026thinsp;2.20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.000\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eSmall_Intestine\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDmax(Gy)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e49.08\u0026thinsp;\u0026plusmn;\u0026thinsp;0.51\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e48.77\u0026thinsp;\u0026plusmn;\u0026thinsp;0.54\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.005\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDmean(Gy)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e19.12\u0026thinsp;\u0026plusmn;\u0026thinsp;3.95\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e18.10\u0026thinsp;\u0026plusmn;\u0026thinsp;3.89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.000\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eV30(%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e25.10\u0026thinsp;\u0026plusmn;\u0026thinsp;6.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e21.00\u0026thinsp;\u0026plusmn;\u0026thinsp;7.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.000\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSpinalCord\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDmax(Gy)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e31.98\u0026thinsp;\u0026plusmn;\u0026thinsp;2.81\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e29.53\u0026thinsp;\u0026plusmn;\u0026thinsp;2.61\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.000\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDmean(Gy)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e13.53\u0026thinsp;\u0026plusmn;\u0026thinsp;4.16\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e12.34\u0026thinsp;\u0026plusmn;\u0026thinsp;3.82\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.000\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eMarrow\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDmean(Gy)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e33.63\u0026thinsp;\u0026plusmn;\u0026thinsp;1.07\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e33.56\u0026thinsp;\u0026plusmn;\u0026thinsp;1.34\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.714\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eV30(%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e65.50\u0026thinsp;\u0026plusmn;\u0026thinsp;4.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e64.30\u0026thinsp;\u0026plusmn;\u0026thinsp;6.10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.213\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eLeft\u003c/p\u003e \u003cp\u003eFemoral Head\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDmean(Gy)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e23.49\u0026thinsp;\u0026plusmn;\u0026thinsp;3.25\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e24.50\u0026thinsp;\u0026plusmn;\u0026thinsp;2.93\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.010\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eV30(%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e18.90\u0026thinsp;\u0026plusmn;\u0026thinsp;10.50\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e24.40\u0026thinsp;\u0026plusmn;\u0026thinsp;10.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eV40(%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.40\u0026thinsp;\u0026plusmn;\u0026thinsp;0.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.00\u0026thinsp;\u0026plusmn;\u0026thinsp;1.30\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.002\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eRight\u003c/p\u003e \u003cp\u003eFemoral Head\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDmean(Gy)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e23.86\u0026thinsp;\u0026plusmn;\u0026thinsp;2.57\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e25.05\u0026thinsp;\u0026plusmn;\u0026thinsp;2.21\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.006\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eV30(%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e18.40\u0026thinsp;\u0026plusmn;\u0026thinsp;9.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e22.90\u0026thinsp;\u0026plusmn;\u0026thinsp;8.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.004\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eV40(%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.60\u0026thinsp;\u0026plusmn;\u0026thinsp;0.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.00\u0026thinsp;\u0026plusmn;\u0026thinsp;1.30\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.014\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"5\"\u003eP-value: calculated by Paired t-test; P value\u0026thinsp;\u0026lt;\u0026thinsp;0.05 was considered statistically significant;\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eDeliverability verification using gamma analysis with a 3%/3 mm criterion yielded comparable passing rates: the E2E auto-plans achieved a mean of 98.1% (range: from 96.7% to 99.0%), while the reference plans achieved 97.9% (range: from 96.8% to 99.3%) in table S3, indicating similar deliverability to clinical plans.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eWe developed and validated a closed-loop VMAT auto-planning workflow that converts patient-specific anatomy into deliverable plans. Target coverage was preserved while DVH metrics for key OARs improved, and deliverability was supported by delivery verification. Dose prediction was conditioned on interpretable, direction-aware beam band priors and trained with a composite loss function to better capture steep dose fall-off and target\u0026ndash;OAR abutment interfaces. Predicted dose was translated, via TPS script module, into Monaco-executable individualized objectives and iteratively optimized, enabling robust transfer from prediction to actionable constraints and reducing repeated manual tuning, thereby providing a practical route for integrating prediction into E2E clinical VMAT planning.\u003c/p\u003e \u003cp\u003eUnlike other DL medical image regression tasks such as auto-segmentation and image reconstruction, where inputs are often relatively direct, dose prediction typically requires deliberate input design [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e]. This is because plan generation is shaped not only by anatomy, but also by clinical preferences and physical realities such as beam arrangement and energy deposition constraints [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e, \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e]. Sun et al. proposed a physics voxel\u0026ndash;based optimization strategy and reported mean-dose reductions of 3.1, 6.2 and 4.5 Gy in the bladder and bilateral femoral heads compared with manual plans [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]. In the work by Teng et al., fixed-field characteristics specific to IMRT were leveraged: they simulated ray paths according to PTV position and beam angles to construct beam masks as network inputs, and demonstrated reduced prediction errors across individual beam angles [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. Xiong et al. designed normalized distance-aware beam plates and mass density maps as physics-informed priors, and their ablation experiments showed a 5.8% reduction in MAE after incorporating such priors [\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e]. Considering the dynamic modulation and continuous gantry rotation intrinsic in VMAT, we constructed four directional beam band masks to approximate the directionality of dose incidence and used them as geometric priors as network inputs. In ablation experiments, adding the band masks reduced the Dose score from 2.407 to 2.245 (a relative decrease of 6.7%), and the full configuration further reduced it to 2.176 (a relative decrease of 9.6%). These results indicate that explicitly encoding direction-aware band priors in the model input yields a consistent improvement in dose prediction performance.\u003c/p\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e showed that the proposed two-stage cascaded dose prediction framework outperformed the other evaluated models. A key reason was that many prior DL approaches, despite careful architectural design, were still trained primarily with MAE or other similar mean error-related objectives. Such objectives encourage learning a mapping towards the conditional mean or median, which can produce overly smooth dose distributions and attenuate boundary details [\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e, \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e]. This is undesirable when high-gradient transitions are clinically critical. Nguyen et al. explored differentiable DVH-based and adversarially inspired losses, demonstrating their utility for training dose models and generating Pareto-optimal radiotherapy dose distributions [\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e]. In our framework, MAE was retained in the first stage to learn the global dose distribution, while the second stage introduces a composite objective (\u003cem\u003eL\u003c/em\u003e\u003csub\u003e\u003cem\u003eband\u003c/em\u003e\u003c/sub\u003e + \u003cem\u003eL\u003c/em\u003e\u003csub\u003e\u003cem\u003egrad\u003c/em\u003e\u003c/sub\u003e + \u003cem\u003eL\u003c/em\u003e\u003csub\u003e\u003cem\u003eDVH\u003c/em\u003e\u003c/sub\u003e) to refine band-region behaviour, local gradient transitions, and DVH consistency. This design aims to improve prediction accuracy while preserving global coherence. From a task-decomposition perspective, the cascade allows the model to learn the low-frequency/global dose pattern first and then correct high-frequency/local discrepancies, which can reduce the learning difficulty of the refinement stage and improve training robustness. Teng et al. adopted a related strategy by decomposing a global coarse dose into multiple field doses, refining them on a per-field basis, and then aggregating to obtain the final 3D dose distribution, achieving improved predictive performance [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. In addition, our cohort was comparatively large, and organ-wise dose distributions span a broad range (Supplementary Fig. S3), which provided richer supervision for the refinement stage during error correction.\u003c/p\u003e \u003cp\u003eSeveral studies had explored how to translate predicted dose into actionable planning. Shen et al. discretized predicted dose in specific regions to construct optimization objectives and used a two-step optimization scheme [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. Choi et al. automatically extracted structure-specific objective values and weights from predicted dose and wrote them into a TPS to generate plans [\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e]. Church et al. combined dose mimicking with a residual U-Net to predict deliverability-related elements and converted them into DICOM-RT plans for evaluation [\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e]. Relative to these studies, the present work emphasized three aspects. First, it targets the VMAT setting, where dynamic modulation and continuous rotation impose directionality and high-gradient characteristics that were not well represented by anatomy alone. Second, the predicted dose was transformed through TPS scripting into Monaco-executable individualized constraints and iteratively optimized to produce the final plan, rather than remaining solely at the level of a predicted distribution. Third, we incorporated gamma analysis to support plan deliverability. Moreover, Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e indicated that, with target coverage maintained, DVH endpoints for most key OARs decreased overall. Notably, Dmean and V30% for the bilateral femoral heads increased by approximately 1 Gy and 5%, respectively. This might reflect a redistribution of dose contributions across directions during optimization. Additionally, femoral heads were relatively small structures and their DVH metrics could be more sensitive to localized dose variations. Importantly, femoral head doses in the E2E auto-plans still remained within clinically acceptable ranges. In future work, we would further refine the optimization strategy for small-volume bony structures such as the femoral heads, introducing additional constraints and/or objective rebalancing to prevent metric drift and ideally achieve neutral or reduced values.\u003c/p\u003e \u003cp\u003eThis study had several limitations. First, the diversity of the cohort and prescription configurations remains limited; generalizability across different prescriptions, planning preferences, and broader patient populations required further validation. We plan to expand the sample size and include additional prescription schemes in CC cohorts. Second, plan quality is partly contingent on the accuracy of dose prediction; we will continue to improve the predictive network and loss design to enhance robustness in steep-gradient and target\u0026ndash;OAR adjacent regions. Third, although we achieved E2E auto-planning within the Monaco TPS, cross-TPS portability has not been established and will require dedicated interface development and consistency assessments. Finally, our experiments focused on a single tumor site; validation across multiple tumor sites is essential. A multi-site auto-planning framework, adapted to varying clinical requirements, is expected to offer broader applicability and improved robustness, and this will facilitate more reliable integration of the proposed workflow into clinical radiotherapy practice.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eWe propose a closed-loop framework that converts anatomy into deliverable VMAT plans by integrating direction-aware beam-band priors, DVH-guided learning, and Monaco TPS scripting for optimization. The method preserves target coverage, improves OARs\u0026rsquo; DVH endpoints, and demonstrates deliverability with gamma verification. This approach reduces manual trial-and-error and generates deliverable plans, which has the potential to substantially improve the efficiency and consistency of CC radiotherapy planning, with promising prospects for rapid auto-plan generation in routine practice.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cdiv class=\"DefinitionList\"\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eIMRT\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eIntensity-modulated radiotherapy\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eVMAT\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eVolumetric modulated arc therapy\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eCC\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eCervical cancer\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eOARs\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eOrgans at risk\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eATP\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eAutomated treatment planning\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eKBP\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eKnowledge-based planning\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eDVH\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eDose-volume histogram\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e3D\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eThree-dimensional\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eDL\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eDeep learning\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eTPS\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eTreatment planning system\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eE2E\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eEnd-to-end\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003ePTV\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003ePlan target volume\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eMAE\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eMean absolute error\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eROI\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eRegion of interest\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003esnDVH score\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eScale-normalized DVH score.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e \u003cstrong\u003eEthics approval and consent to participate:\u003c/strong\u003e \u003cp\u003e This study was approved by the Ethics Committee (Institutional Review Board) of Harbin Medical University Cancer Hospital, Harbin, China (ECCR No. KY2023-83, 2023-12-18). The study was conducted in accordance with the Declaration of Helsinki. The requirement for informed consent was waived by the Ethics Committee/IRB due to the retrospective nature of the study.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eConsent for publication:\u003c/strong\u003e \u003cp\u003eNot applicable.\u003c/p\u003e \u003c/p\u003e\u003cp\u003e \u003ch2\u003eCompeting interests:\u003c/h2\u003e \u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eFunding:\u003c/h2\u003e \u003cp\u003eThis research was supported partially by the Joint Fund Cultivation Project of Heilongjiang Provincial Natural Science Foundation (No. PL2025A001), the National Natural Science Foundation of China (12375341).\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eBoda Ning: Software, Data Curation, Writing - Original Draft, Validation, Formal analysis, Investigation. Xiuyan Liang: Data Curation, Formal analysis, Writing - Review \u0026amp; Editing. Zhenguo Cui: Data Curation, Conceptualization. Yingfa Li: Data Curation, Software. Qi Liu: Formal analysis. Shuaining Ma: Resources. Xiting Chen: Investigation. Shanshan Yang: Data Curation, Methodology, Supervision. Yanling Bai: Conceptualization, Methodology, Writing - Review \u0026amp; Editing, Supervision, Funding acquisition. Deyang Yu: Conceptualization, Methodology, Writing - Review \u0026amp; Editing, Supervision, Project administration, Funding acquisition.\u003c/p\u003e\u003ch2\u003eAcknowledgements:\u003c/h2\u003e \u003cp\u003eNot applicable.\u003c/p\u003e \u003cp\u003e \u003cb\u003eClinical trial number\u003c/b\u003e: Not applicable.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eBertelsen A, Hansen CR, Johansen J, Brink C. Single Arc Volumetric Modulated Arc Therapy of head and neck cancer. Radiother Oncol. 2010;95:142\u0026ndash;8. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.radonc.2010.01.011\u003c/span\u003e\u003cspan address=\"10.1016/j.radonc.2010.01.011\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLee N, Puri DR, Blanco AI, Chao KS. Intensity-modulated radiation therapy in head and neck cancers: an update. Head Neck. 2007;29:387\u0026ndash;400. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/hed.20332\u003c/span\u003e\u003cspan address=\"10.1002/hed.20332\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVerginadis II, Citrin DE, Ky B, Feigenberg SJ, Georgakilas AG, Hill-Kayser CE, et al. Radiotherapy toxicities: mechanisms, management, and future directions. Lancet. 2025;405:338\u0026ndash;52. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/s0140-6736(24)02319-5\u003c/span\u003e\u003cspan address=\"10.1016/s0140-6736(24)02319-5\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGe Y, Wu QJ. Knowledge-based planning for intensity-modulated radiation therapy: A review of data-driven approaches. Med Phys. 2019;46:2760\u0026ndash;75. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/mp.13526\u003c/span\u003e\u003cspan address=\"10.1002/mp.13526\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJia X, Rong Y, Wu Q, Cardenas CE, Court LE, Hrinivich WT, et al. NRG Oncology Assessment of Artificial Intelligence for Automatic Treatment Planning in Radiation Therapy Clinical Trials: Present and Future. Int J Radiat Oncol Biol Phys. 2025;123:282\u0026ndash;95. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.ijrobp.2025.03.045\u003c/span\u003e\u003cspan address=\"10.1016/j.ijrobp.2025.03.045\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNguyen D, Lin MH, Sher D, Lu W, Jia X, Jiang S. Advances in Automated Treatment Planning. Semin Radiat Oncol. 2022;32:343\u0026ndash;50. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.semradonc.2022.06.004\u003c/span\u003e\u003cspan address=\"10.1016/j.semradonc.2022.06.004\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTol JP, Delaney AR, Dahele M, Slotman BJ, Verbakel WF. Evaluation of a knowledge-based planning solution for head and neck cancer. Int J Radiat Oncol Biol Phys. 2015;91:612\u0026ndash;20. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.ijrobp.2014.11.014\u003c/span\u003e\u003cspan address=\"10.1016/j.ijrobp.2014.11.014\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHussein M, South CP, Barry MA, Adams EJ, Jordan TJ, Stewart AJ, et al. Clinical validation and benchmarking of knowledge-based IMRT and VMAT treatment planning in pelvic anatomy. Radiother Oncol. 2016;120:473\u0026ndash;9. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.radonc.2016.06.022\u003c/span\u003e\u003cspan address=\"10.1016/j.radonc.2016.06.022\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi N, Carmona R, Sirak I, Kasaova L, Followill D, Michalski J, et al. Highly Efficient Training, Refinement, and Validation of a Knowledge-based Planning Quality-Control System for Radiation Therapy Clinical Trials. Int J Radiat Oncol Biol Phys. 2017;97:164\u0026ndash;72. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.ijrobp.2016.10.005\u003c/span\u003e\u003cspan address=\"10.1016/j.ijrobp.2016.10.005\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang M, Zhang Q, Lam S, Cai J, Yang R. A Review on Application of Deep Learning Algorithms in External Beam Radiotherapy Automated Treatment Planning. Front Oncol. 2020;10:580919. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3389/fonc.2020.580919\u003c/span\u003e\u003cspan address=\"10.3389/fonc.2020.580919\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLeino A, Heikkil\u0026auml; J, Vir\u0026eacute;n T, Honkanen JTJ, Sepp\u0026auml;l\u0026auml; J, Korkalainen H. Deep learning-based prediction of the dose-volume histograms for volumetric modulated arc therapy of left-sided breast cancer. Med Phys. 2024;51:7986\u0026ndash;97. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/mp.17410\u003c/span\u003e\u003cspan address=\"10.1002/mp.17410\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJiao Z, Peng X, Wang Y, Xiao J, Nie D, Wu X, et al. TransDose: Transformer-based radiotherapy dose prediction from CT images guided by super-pixel-level GCN classification. Med Image Anal. 2023;89:102902. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.media.2023.102902\u003c/span\u003e\u003cspan address=\"10.1016/j.media.2023.102902\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTeng L, Wang B, Xu X, Zhang J, Mei L, Feng Q, et al. Beam-wise dose composition learning for head and neck cancer dose prediction in radiotherapy. Med Image Anal. 2024;92:103045. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.media.2023.103045\u003c/span\u003e\u003cspan address=\"10.1016/j.media.2023.103045\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXie H, Zhang H, Chen Z, Tan T. Precision dose prediction for breast cancer patients undergoing IMRT: The Swin-UMamba-Channel Model. Comput Med Imaging Graph. 2024;116:102409. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.compmedimag.2024.102409\u003c/span\u003e\u003cspan address=\"10.1016/j.compmedimag.2024.102409\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang Y, Li C, Zhong L, Chen Z, Yang W, Wang X. DoseDiff: Distance-Aware Diffusion Model for Dose Prediction in Radiotherapy. IEEE Trans Med Imaging. 2024;43:3621\u0026ndash;33. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/tmi.2024.3383423\u003c/span\u003e\u003cspan address=\"10.1109/tmi.2024.3383423\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCao X, Zhao Y, Li S, Zhang F, Yang Z, Yang X. Beam field guided diffusion model for liver cancer radiotherapy dose distribution prediction. Med Phys. 2025;52:e17989. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/mp.17989\u003c/span\u003e\u003cspan address=\"10.1002/mp.17989\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChang HH, Harms J, Cardan RA, Fiveash JB, Popple RA, Cardenas CE, nnDoseNet. Intuitive and flexible deep learning framework to train and evaluate radiotherapy dose prediction models. Comput Biol Med. 2025;198:111237. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.compbiomed.2025.111237\u003c/span\u003e\u003cspan address=\"10.1016/j.compbiomed.2025.111237\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOppitz H, Eckl M, Siebenlist K, Boda-Heggemann J, Abo-Madyan Y, Giordano FA, et al. Knowledge-based automated radiation therapy treatment planning utilizing dose prediction with a 2.5D-U-Net. Phys Med. 2025;139:105199. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.ejmp.2025.105199\u003c/span\u003e\u003cspan address=\"10.1016/j.ejmp.2025.105199\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXie K, Hua Q, Gao L, Sun J, Lin T, Sui J, et al. SegMambaDP: a long-range sequential modeling mamba network for 3D dose prediction in radiotherapy. BMC Med Imaging. 2025;25:527. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1186/s12880-025-02055-8\u003c/span\u003e\u003cspan address=\"10.1186/s12880-025-02055-8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShen Y, Tang X, Lin S, Jin X, Ding J, Shao M. Automatic dose prediction using deep learning and plan optimization with finite-element control for intensity modulated radiation therapy. Med Phys. 2024;51:545\u0026ndash;55. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/mp.16743\u003c/span\u003e\u003cspan address=\"10.1002/mp.16743\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXiao Y, Tanaka S, Kadoya N, Sato K, Kimura Y, Umezawa R, et al. Evaluation of deliverable artificial intelligence-based automated volumetric arc radiation therapy planning for whole pelvic radiation in gynecologic cancer. Sci Rep. 2025;15:15219. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41598-025-99717-y\u003c/span\u003e\u003cspan address=\"10.1038/s41598-025-99717-y\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHuang W, Liu T, Shen Y, Xiang Z, Wang D, Fu W, et al. Automatic radiotherapy planning for deliverable plans using deep learning dose prediction and dose rings optimization in cervical cancer. J Appl Clin Med Phys. 2025;26:e70353. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/acm2.70353\u003c/span\u003e\u003cspan address=\"10.1002/acm2.70353\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e\u0026Ccedil;i\u0026ccedil;ek \u0026Ouml;, Abdulkadir A, Lienkamp SS, Brox T, Ronneberger O. 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation. Springer; 2016. pp. 424\u0026ndash;32.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang Z, Liu Q, Wang Y. Road Extraction by Deep Residual U-Net. IEEE Geosci Remote Sens Lett. 2018;15:749\u0026ndash;53. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/LGRS.2018.2802944\u003c/span\u003e\u003cspan address=\"10.1109/LGRS.2018.2802944\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMilletari F, Navab N, Ahmadi S-A. V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation. Fourth International Conference on 3D Vision (3DV): IEEE; 2016. pp. 565\u0026thinsp;\u0026ndash;\u0026thinsp;71.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu S, Zhang J, Li T, Yan H, Liu J, Technical Note. A cascade 3D U-Net for dose prediction in radiotherapy. Med Phys. 2021;48:5574\u0026ndash;82. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/mp.15034\u003c/span\u003e\u003cspan address=\"10.1002/mp.15034\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKearney V, Chan JW, Haaf S, Descovich M, Solberg TD. DoseNet: a volumetric dose prediction algorithm using 3D fully-convolutional neural networks. Phys Med Biol. 2018;63:235022. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1088/1361-6560/aaef74\u003c/span\u003e\u003cspan address=\"10.1088/1361-6560/aaef74\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNguyen D, Jia X, Sher D, Lin MH, Iqbal Z, Liu H, et al. 3D radiotherapy dose prediction on head and neck cancer patients with a hierarchically densely connected U-net deep learning architecture. Phys Med Biol. 2019;64:065020. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1088/1361-6560/ab039b\u003c/span\u003e\u003cspan address=\"10.1088/1361-6560/ab039b\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOsman AFI, Tamam NM, Yousif YAM. A comparative study of deep learning-based knowledge-based planning methods for 3D dose distribution prediction of head and neck. J Appl Clin Med Phys. 2023;24:e14015. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/acm2.14015\u003c/span\u003e\u003cspan address=\"10.1002/acm2.14015\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBabier A, Zhang B, Mahmood R, Moore KL, Purdie TG, McNiven AL, et al. OpenKBP: The open-access knowledge-based planning grand challenge and dataset. Med Phys. 2021;48:5549\u0026ndash;61. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/mp.14845\u003c/span\u003e\u003cspan address=\"10.1002/mp.14845\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eH\u0026eacute;mon C, Texier B, Lafond C, Nunes JC, Barateau A. Towards trustworthy AI in radiotherapy: a comprehensive review of uncertainty-aware techniques. Phys Med Biol. 2025;71. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1088/1361-6560/ae2a9f\u003c/span\u003e\u003cspan address=\"10.1088/1361-6560/ae2a9f\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYue M, Xue X, Wang Z, Lambo RL, Zhao W, Xie Y, et al. Dose prediction via distance-guided deep learning: Initial development for nasopharyngeal carcinoma radiotherapy. Radiother Oncol. 2022;170:198\u0026ndash;204. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.radonc.2022.03.012\u003c/span\u003e\u003cspan address=\"10.1016/j.radonc.2022.03.012\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBarrag\u0026aacute;n-Montero AM, Nguyen D, Lu W, Lin MH, Norouzi-Kandalan R, Geets X, et al. Three-dimensional dose prediction for lung IMRT patients with deep neural networks: robust learning from heterogeneous beam configurations. Med Phys. 2019;46:3679\u0026ndash;91. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/mp.13597\u003c/span\u003e\u003cspan address=\"10.1002/mp.13597\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSun Z, Xia X, Fan J, Zhao J, Zhang K, Wang J, et al. A hybrid optimization strategy for deliverable intensity-modulated radiotherapy plan generation using deep learning-based dose prediction. Med Phys. 2022;49:1344\u0026ndash;56. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/mp.15462\u003c/span\u003e\u003cspan address=\"10.1002/mp.15462\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXiong T, Ren G, Chen Z, Huang YH, Ma Z, Li Z, et al. A generalizable dose prediction model for automatic radiotherapy planning based on physics-informed priors and large-kernel convolutions. Med Phys. 2026;53:e70272. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/mp.70272\u003c/span\u003e\u003cspan address=\"10.1002/mp.70272\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYang Q, Yan P, Zhang Y, Yu H, Shi Y, Mou X, et al. Low-Dose CT Image Denoising Using a Generative Adversarial Network With Wasserstein Distance and Perceptual Loss. IEEE Trans Med Imaging. 2018;37:1348\u0026ndash;57. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/tmi.2018.2827462\u003c/span\u003e\u003cspan address=\"10.1109/tmi.2018.2827462\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHan M, Shim H, Baek J. Low-dose CT denoising via convolutional neural network with an observer loss function. Med Phys. 2021;48:5727\u0026ndash;42. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/mp.15161\u003c/span\u003e\u003cspan address=\"10.1002/mp.15161\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNguyen D, McBeth R, Sadeghnejad Barkousaraie A, Bohara G, Shen C, Jia X, et al. Incorporating human and learned domain knowledge into training deep neural networks: A differentiable dose-volume histogram and adversarial inspired framework for generating Pareto optimal dose distributions in radiation therapy. Med Phys. 2020;47:837\u0026ndash;49. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/mp.13955\u003c/span\u003e\u003cspan address=\"10.1002/mp.13955\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChoi B, Shrestha DK, Attia A, Stish BJ, Leenstra J, Rwigema JC, et al. Deep learning-based dose prediction for prostate cancer with empty bladder protocol: a framework for efficient and personalized radiotherapy planning. Front Oncol. 2025;15:1690416. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3389/fonc.2025.1690416\u003c/span\u003e\u003cspan address=\"10.3389/fonc.2025.1690416\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChurch C, Yap M, Bessrour M, Lamey M, Granville D. Automated plan generation for prostate radiotherapy patients using deep learning and scripted optimization. Phys Imaging Radiat Oncol. 2024;32:100641. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.phro.2024.100641\u003c/span\u003e\u003cspan address=\"10.1016/j.phro.2024.100641\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"radiation-oncology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"raon","sideBox":"Learn more about [Radiation Oncology](http://ro-journal.biomedcentral.com/)","snPcode":"13014","submissionUrl":"https://submission.nature.com/new-submission/13014/3","title":"Radiation Oncology","twitterHandle":"@OncoBioMed","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Auto-planning, Deep learning, Dose prediction, Treatment planning system script, Radiotherapy","lastPublishedDoi":"10.21203/rs.3.rs-8749620/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8749620/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground and purpose:\u003c/strong\u003e The rapid and efficient generation of high-quality, dose-consistency volumetric modulated arc therapy (VMAT) plans remains challenging in radiotherapy. This study proposes a deep learning (DL) end-to-end (E2E) auto-planning framework and validate its practicality and feasibility for clinical implementation.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMaterials and methods: \u003c/strong\u003eA total of 458 cervical cancer VMAT plans were enrolled and split into training, validation, and test cohorts. An E2E auto-planning framework with a two-stage cascaded deep learning (DL) network was developed: Stage 1 predicted coarse dose from CT and structure masks, and Stage 2 refined it using four beam-band priors and a composite loss. Dose-volume histogram (DVH) endpoints from refined predicted dose were converted into Monaco objectives via a scripting module for iterative optimization. Performance was evaluated with Dose, DVH, and snDVH scores, ablations, and comparisons with manual plans in terms of quality, clinical evaluation and deliverability.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults: \u003c/strong\u003eThe proposed DL method achieved the best performance, with Dose score, DVH score and snDVH score of 2.114 ± 0.218 Gy, 1.194 ± 0.295 Gy and 2.027 ± 0.586, respectively. Compared with manual plans, E2E auto-plans preserved target volume coverage while reducing all DVH metrics for bladder, rectum, small intestine, and spinal cord by 2% - 35% (all p \u0026lt; 0.05). The gamma passing rate of E2E auto-plans was higher than manual plans in the 3%/3 mm gamma criterion (98.1% vs 97.9%).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusion:\u003c/strong\u003eThe proposed auto-planning framework demonstrated a high level of automation and clinical applicability, offering a reliable and promising tool to support radiotherapy workflows.\u003c/p\u003e","manuscriptTitle":"A deep learning-driven automated treatment planning framework for patient treated with radiotherapy in cervical cancer","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-02-11 16:52:41","doi":"10.21203/rs.3.rs-8749620/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-02-26T14:39:02+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-02-26T08:50:58+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"108289015639673830201231901297135484992","date":"2026-02-09T08:40:37+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-02-09T08:02:32+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-02-04T07:34:31+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-02-04T07:31:28+00:00","index":"","fulltext":""},{"type":"submitted","content":"Radiation Oncology","date":"2026-01-31T11:31:25+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"radiation-oncology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"raon","sideBox":"Learn more about [Radiation Oncology](http://ro-journal.biomedcentral.com/)","snPcode":"13014","submissionUrl":"https://submission.nature.com/new-submission/13014/3","title":"Radiation Oncology","twitterHandle":"@OncoBioMed","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"ee186ab4-2d40-4de3-9edc-2f53c3d8c5cc","owner":[],"postedDate":"February 11th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2026-04-20T16:01:20+00:00","versionOfRecord":{"articleIdentity":"rs-8749620","link":"https://doi.org/10.1186/s13014-026-02842-9","journal":{"identity":"radiation-oncology","isVorOnly":false,"title":"Radiation Oncology"},"publishedOn":"2026-04-14 15:57:35","publishedOnDateReadable":"April 14th, 2026"},"versionCreatedAt":"2026-02-11 16:52:41","video":"","vorDoi":"10.1186/s13014-026-02842-9","vorDoiUrl":"https://doi.org/10.1186/s13014-026-02842-9","workflowStages":[]},"version":"v1","identity":"rs-8749620","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8749620","identity":"rs-8749620","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}