Clinical implementation and evaluation of deep learning-assisted automatic radiotherapy treatment planning for lung cancer

Clinical implementation and evaluation of deep learning-assisted automatic radiotherapy treatment planning for lung 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 Clinical implementation and evaluation of deep learning-assisted automatic radiotherapy treatment planning for lung cancer Ningyu Wang, Jiawei Fan, Yingjie Xu, Lingling Yan, Deqi Chen, and 4 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-3872969/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background and purpose The purpose of this study is to investigate the clinical application and assessment of deep learning (DL)-assisted automatic radiotherapy planning for lung cancer patients. Materials and methods We developed a deep learning model for predicting patient-specific dose that was trained and validated on a dataset of 235 lung cancer patients, and the model was integrated into clinical workflow to assist planners in generating treatment plans. We retrospectively selected and recovered additional 50 clinically treated lung cancer patients’ manual volumetric modulated arc therapy (VMAT) plans with different target volumes and different treatment patterns. Subsequently, automatic plans were generated for each of these patients. Both automatic and manual plans were subsequently compared in terms of overall plan quality metric (PQM), target coverage and homogeneity, organ at risk (OAR) sparing, monitor units (MUs), and planning time. Additionally, qualitative reviews of automatic and manual plans were implemented by four expert reviewers to assess the clinical applicability of DL-assisted automatic plans. Results The average PQM score was 40.7 ± 13.1 for manual plans and 40.8 ± 13.5 for automatic plans, and they had equivalent overall plan quality. The targets coverage and homogeneity of the automatic plans were considered equivalent or superior when compared to manual plans. Both plans had their own advantages in OAR sparing, such as better sparing of lung for manual plans and better sparing of heart for automatic plans. It is worth to note that the average planning time of automatic plans was reduced from 103.1 ± 18.5 minutes to 32.6 ± 5.3 minutes (P<0.001) and the MUs were reduced from 789.9 ± 234.3 to 692.5 ± 210.7 (P<0.001). In qualitative evaluation, automatic plans were deemed to be clinically acceptable for treatment in 88% of reviews (176/200), and all were accepted after fine tuning. Most expert reviews indicated a preference for equivalence between automatic and manual plans when making their selection. Conclusion The DL-assisted lung cancer plans demonstrated comparable or superior quality to manual plans, improved planning and treatment efficiency, and significantly reduced planning time and MUs. It has the potential to enhance the workflow of radiotherapy departments, ultimately providing tangible benefit to lung cancer patients. Lung cancer Automatic planning Deep learning VMAT Expert review Figures Figure 1 Figure 2 Figure 3 1. Introduction Lung cancer is the most morbidity and mortality cancer globally, accounting for about 21.3% of all malignant tumors [ 1 ]. Radiotherapy is one of the important means of lung cancer treatment and it plays an increasingly prominent role. Approximately 70% of lung cancer patients need radiotherapy during the treatment process. Radiotherapy is applicable to different types of lung cancer at different stages. Middle- and early-stage lung cancer can be treated by radical radiotherapy or combined radiotherapy and chemotherapy, and advanced lung cancer can be treated by palliative radiotherapy. As a result, the number of lung cancer radiotherapy plans that planners need to design is increasing clinically. Moreover, the targets of lung cancer plans are complex and variable, which makes planning difficult. In the design process of volumetric modulated arc therapy (VMAT) plans, planners need to repeatedly adjust the objective functions, weights and other plan parameters based on personal experience. Therefore, it is an iterative trial-and-error process [ 2 ]. The process is time consuming and the quality of the final treatment plan may depend highly on the experience and skills of planners, the complexity of the case, and the available time [ 3 ]. In addition to this, the development of advanced treatment techniques, such as simultaneous integrated boost irradiation [ 4 ], will lead to a workload growth of planning design. There has been increasing interest in automatic plans in recent years, with the aim of improving plan quality and reducing planning time. Some studies have shown that automatic plans can considerably maintain the plan quality while dramatically reduce the planning workload compared to manual plans [ 5 ]. Automatic plan can potentially decrease the amount of effort involved in manual "trial-and-error" methods by around 50% [ 6 ]. Currently, there are script-based planning method [ 7 ], as well as automatic planning modules that have been integrated into Treatment planning system (TPS). Examples of these modules include RapidPlan in Eclipse [ 8 ], the volume-driven automatic planning platform in Pinnacle 3 [ 9 ] and Multi-Criteria Optimization in RayStation [ 10 ]. These planning modules, with knowledge-based planning (KBP) methods [ 11 , 12 ], utilize databases of patient plans to establish prediction models for estimating dose-volume histograms and dose maps. In recent years, deep learning (DL) methods have achieved significant success in predicting radiotherapy plans [ 13 , 14 , 15 ], leading to enhanced accuracy and efficiency. In contrast to traditional KBP methods with handcrafted features, deep learning methods can automatically learn image features from raw data, such as CT images, contours, and dose maps, which are tailored specifically to the prediction task [ 16 ]. A virtual treatment planner network (VTPN) was built by Shen et al [ 17 ] using DL to operate a TPS by adjusting treatment planning parameters in it. Wang et al [ 18 ] proposed a novel algorithm named as multi-objectives adjustment policy network (MOAPN) and trained to learn how to adjust multiple optimization objectives in commercial Eclipse TPS based on the multi-agent DL scheme. The existing prediction models are focused on single target and treatment pattern, and there is a lack of a generic prediction model for radiotherapy plans with multiple targets and different treatment patterns. This study is based on a deep learning approach to establish an intelligent radiotherapy method for lung cancer that is applicable to any individualized anatomical structure and prescribed dose pattern, under the condition that the target location and shape are diverse. In addition, many existing automatic planning methods are for intensity modulated radiation therapy (IMRT) technology [ 2 , 14 , 18 ], while this study is for VMAT technology. The purpose of this study is to realize an integration of DL as a standard workflow into the current TPS to test the feasibility of clinical deployment for lung cancer. We retrospectively recovered 50 previously manually designed, peer reviewed and delivered patient VMAT treatment plans. For each patient, we designed automatic VMAT plans. Both automatic and manual plans were compared from different perspectives to evaluate the plan quality and efficiency, and ultimately to investigate the feasibility of clinical implementation of DL-assisted automatic treatment planning. 2. Materials and methods 2.1 Development of the DL model and integration into TPS The DL method was trained on 235 high-quality, previously approved, peer reviewed and delivered patient treatment plans which were generated by an experienced team dedicated to the treatment of lung cancer at a large academic cancer center. We developed a DL model for lung cancer dose prediction based on the architecture in our previous study [ 19 ]. We retrained the DL model using these lung cancer datasets. The inputs to the DL model were the contours, prescriptions of targets and the outputs were the doses for lung cancer cases. The details of datasets, architecture, inputs and outputs, training and validation process of the DL model can refer to Supplementary Material 1. The trained DL model was used for this retrospective study phase. For a new patient in retrospective dataset, the model automatically learned features from patient and delineated Planning Target Volumes (PTVs), OARs and Body to predict patient-specific doses. And the model was integrated into clinical Pinnacle 3 TPS to achieve setting inverse optimization objectives of targets and OARs, with the help of optimization and dose calculation engines from TPS to realize automatic treatment planning. The integration workflow demonstrated in Fig. 1 was implemented in clinical practice. It mainly involved three parts: i) DICOM-RT data preparation and transfer, ii) data preprocessing, dose prediction, and data post-processing, iii) data transfer and automatic design plan. Intelligent planning system was implemented in third-party servers and integrated into clinical Pinnacle 3 TPS. 2.2 Patient data Between January 2022 and May 2022, a cohort of additional 50 treated plans with different tumor stages and prescribed doses were randomly and retrospectively selected. All patients were staged according to the American Joint Committee on Cancer (AJCC) Manual for Staging of Cancer, 8th edition. Patients had previously been scanned using either a Philips Brilliance Big Bore or Siemens Syngo CT scanner. These CT images were transferred to the Pinnacle 3 TPS (version 16.2, Philips Medical Systems, Fitchburg, WI). The targets were then manually delineated by the radiation oncologist on these CT images with the help of a contrast-enhanced diagnostic CT. Generally, Clinical Target Volume (CTV) was expanded by 5mm from Gross Target Volume (GTV) and PTV was expanded by 5mm from CTV. If Planning Gross Target Volume (PGTV) needed to be delineated, a 5 mm margin between the PTV and PGTV were considered, though these margins could have been subjected to adaptations for each specific clinical case. The OARs included lung, heart, cord, cord PRV, trachea, esophagus, liver and so on. The dose constraint applied for planning are provided in Supplementary Material 2. 2.3 Planning process The retrospectively selected plans were previously manual designed and approved by several planners with 3–10 years working experiences and were all peer reviewed and delivered. The manual designing process of these plans included several steps: I) the planner first defined CT density table, removed the CT couch, and created an isocenter of the beams in the Pinnacle 3 TPS. II) A Pinnacle 3 script was then used for generating dose shaping structures (DSS), including rings of the targets (Ring), normal tissue (NT), blocks, the fan areas above and below the targets (Fan up and Fan down) (typical example of DSS is shown in Supplementary Material 3). III) Beam angles were determined by the planners according to the position of targets in relation to OARs. Usually, round-trip arcs were created, ranging from 181° to 30° for tumors located in right lung and from 330° to 180° for tumors located in left lung. IV) Other setting parameters were as follows: beam energy 6 MV, dose computational grid resolution 4 mm, control point spacing 4°, and the leaf motion constrained to 0.5 cm/deg. V) The inverse optimization objectives of targets, OARs and DSS were manually defined by planners according to the clinical requirements, and the trivial process of adjusting the optimization parameters were repeated until the plan can be used for clinic. These processes were heavily relied on the experience of the planners. VI) The senior physicists and radiotherapy physicians determined the final delivered plan. Unlike the manual plan designing process, the deep learning-assisted plan can automatically predict patient-specific inverse optimization objectives according to their prescriptions and geometries. For these retrospective selected plans, we redesigned the automatic plans. The process of I)-IV) was basically same as manual plan design. In step II), manual plan called the script of DSS, and automatic plan called the smartplan script including DSS and objective functions (the descriptions of objective functions are provided in Supplementary Material 4). Primarily, the biggest difference between manual and automatic plan was in step V). The process of V) in automatic plan was set based on the patient-specific prediction results. And the automatic plans were obtained by optimizing twice based on the predicted optimization objectives. 2.4 Evaluation Metrics 2.4.1 Quantitative evaluation Dosimetric analysis According to ICRU 83 [ 20 ], dosimetric evaluations for targets coverage and OARs sparing of two kinds of plans were performed with Dose Volume Histogram (DVH). The homogeneity index (HI) [ 21 ], D98, D95, D50, and D2, were used for target coverage. The definition of HI is as follows: $$HI=({D}_{2}-{D}_{98})/{D}_{50}\times 100\%$$ where D98, D95, D50, and D2 are the dose value corresponding to 98%, 95%, 50%, and 2% of the targets on the DVH, respectively. The OARs metrics used in comparison included: V5, V20, V30 and Dmean for Lung all; V5 for contralateral lung; V30, V40 and Dmean for heart; Dmax for cord, cord PRV, trachea and esophagus; Dmean for liver. The paired t-test was used for comparison with significance determined at the level of P < 0.05. Planning Time and MUs In addition, the planning time and MUs for two kinds of plans were recorded and evaluated. The automatic planning time was recorded directly during the design process, which included the time of step I)-V) in section 2.3. The manual planning time was estimated based on the effective working time, which started when the plans were initiated to design and finished when the plans were determined to be submitted for physical review and clinical review. The biggest difference in time between manual plan and automatic plan was in step V). Manual plan spent most of its time on manually adjusting optimization parameters and carrying out optimizing, while the automatic plan can be directly optimized based on predicted objective functions called by smartplan script. The MUs of both plans can be recorded directly in Pinnacle 3 TPS. PQM quality assessment To quantitatively assess the plan quality of automatic and manual plans, PQM [ 22 ] was used. Based on the concept of quality score proposed by Bohsung et al. [ 23 ] and the new PQM method with related submetrics for lung cancer [ 24 ] this study scored each structure of interest using a score function based on targets/OARs clinical constraints. When there was only PTV, the submetric of target was the HI of PTV. When PTV and PGTV were present, the submetrics of targets were the HI of PTV and PGTV. OAR-related submetrics included V5, V20, Dmean for lung; V30, V40, Dmean for heart and Dmax for cord, cord PRV. The lower limit, upper limit, and PQM value range for different submetrics are shown in Supplementary Material 5. For the target, the closer the HI value is to 0, the higher the PQM value is. For each OAR, the closer the dose index is to the upper limit, the lower the PQM value is, and the PQM value was 0 when dose index exceeds the upper limit. 2.4.2. Qualitative evaluation Blinded expert reviews [ 25 ] were implemented by four expert reviewers (one radiotherapy physician and three physicists) to assess the clinical applicability of automatic plan and there were 50 cases and 200 reviews. The reviewers independently assessed dosimetric preference towards automatic and manual plans as well as overall clinical acceptability for automatic plans. And the preference towards automatic and manual plans were evaluated in terms of overall review, target coverage, target conformity, target homogeneity, lung sparing, heart sparing, cord sparing and OARs sparing. The details of expert review form can refer to Supplementary Material 6. 3. Results 3.1 Patients Different stages and radiotherapy patterns corresponded to different prescribed doses. In 50 patients, simultaneous integrated boost technique was applied to 38 patients whose prescribed doses of PGTV ranged from 50 to 67.2 Gy in 20–30 fractions and prescribed doses of PTV were ranged from 40 to 54 Gy in 20–30 fractions. For the other 12 patients, only PTV was prescribed. And the doses of PTV were ranged from 45 to 60 Gy in 15–30 fractions. The patient characteristics are summarized in Table 1. Table 1 Clinical features of the 50 patients with lung cancer. Characteristics Number of patients % Tumor stages Ⅰ 2 4 Ⅱ 3 6 Ⅲ 37 74 Ⅳ 8 16 Radiotherapy mode Postoperative radiotherapy 7 14 Concurrent chemoradiotherapy 13 26 Radiotherapy alone 30 60 Prescribed dose (PTV single) 40~60 Gy, 15~30 fraction 12 24 Prescribed dose (PTV and PGTV) PTV: 40~54 Gy, 20~30 fraction PGTV: 50~67.2 Gy, 20~30 fraction 38 76 Gender Male 36 72 Female 14 28 Location Right 25 50 Left 25 50 3.2 Targets coverage and homogeneity Table 2 compares the target volume dose metrics between automatic plans and manual plans. For the PTV only, there was no significant difference for D98, D95, D50, D2 and HI values between two kinds of plans. For cases with PTV and PGTV targets, HI values of automatic plans were comparable to that of manual plans (P=0.34 for PTV, P=0.50 for PGTV). While HI values do not vary obviously, targets coverage of automatic plans were considered equivalent or better than the corresponding manual plans. The median dose (D50%) in PTV was on average improved by 1.0 Gy ± 0.2 Gy (P = 0.001) and the near-maximum dose (D2%) in PTV by 0.7 Gy ± 0.1 Gy (P = 0.02). The near-minimum dose (D98%) in PGTV was on average improved by 0.4 Gy ± 0.4 Gy (P = 0.01) and the median dose (D50%) in PGTV by 0.9 Gy ± 0.1 Gy (P = 0.001). Overall, both sets of plans achieved similar targets coverage and homogeneity. Table 2 Comparison of the targets dose metrics between manual and automatic plans. Structures Metrics Manual plan Automatic plan P-value PTV (single) HI 0.14±0.04 0.14±0.04 0.09 D98 (Gy) 49.8±4.9 48.8±5.0 0.28 D95 (Gy) 51.3±5.3 51.3±5.3 0.50 D50 (Gy) 54.3±5.6 54.8±5.4 0.26 D2 (Gy) 56.7±5.9 57.4±6.0 0.09 PTV (PTV and PGTV) HI 0.31±0.05 0.32±0.04 0.34 D98 (Gy) 47.8±2.8 47.9±2.3 0.66 D95 (Gy) 50.8±2.5 51.1±3.0 0.28 D50 (Gy) 58.7±3.7 59.7±3.9 0.001 D2 (Gy) 66.0±3.0 66.7±3.1 0.02 PGTV (PTV and PGTV) HI 0.12±0.02 0.11±0.03 0.50 D98 (Gy) 59.3±2.7 59.7±2.3 0.01 D95 (Gy) 60.7±2.6 60.8±4.9 0.32 D50 (Gy) 63.5±2.9 64.4±3.0 0.001 D2 (Gy) 66.7±3.1 67.1±3.1 0.24 3.3 OARs sparing Figure 2(a) shows a boxplot of DVH metrics between the automatic plans and manual plans for Lung and Heart. As is shown in Figure 2(b), the maximum and mean doses were considered for serial and parallel OARs, respectively. No significant difference was observed in the higher dose (V20, V30) of lung all and the lower dose (V5) of contralateral lung. But the V5 values of lung all in automatic plans were higher. Manual plans had better sparing for lung, while automatic plans had better sparing for heart. As can be seen from Table 3 and Figure 2, both V30 and Dmean were significantly lower in the automatic plans than in the manual plans (P=0.03). When considering other normal tissues sparing, including cord, trachea, esophagus and liver, it showed no significant difference. Moreover, Supplementary Material 7 demonstrates a typical case of better protection for OARs in automatic plan. Table 3 Comparison of the OARs dose metrics between manual and automatic plans. Structures Metrics Manual plan Automatic plan P-value Lung all V5 (%) 32.6±6.5 34.4±10.7 0.003 V20 (%) 17.4±5.6 17.7±6.2 0.30 V30 (%) 13.2±4.5 13.3±4.9 0.58 Dmean (Gy) 54.3±5.6 54.8±5.4 0.06 Contralateral lung V5 (%) 13.3±12.1 14.1±11.8 0.15 Heart V30 (%) 16.0±11.9 14.6±12.0 0.03 V40 (%) 9.5±7.8 8.8±8.0 0.19 Dmean (Gy) 12.4±8.0 11.9±7.9 0.03 Cord Dmax (Gy) 33.0±6.8 32.3±5.6 0.29 Cord PRV Dmax (Gy) 36.7±6.8 37.4±5.5 0.28 Trachea Dmax (Gy) 55.5±12.0 55.2±14.7 0.74 Esophagus Dmax (Gy) 52.7±11.9 53.6±12.0 0.05 Liver Dmean (Gy) 0.7±1.1 0.8±1.5 0.59 3.4 Planning qualities and time The overall plan qualities of two kinds of plans were similar, whereas automatic plans took much shorter time (P＜0.001) and had smaller MUs (P＜0.001) than manual plans. Table 4 presents the comparison results of PQM total value, planning time and MUs. The average PQM score was 40.7±13.1 for manual plans and 40.8±13.5 for automatic plans. Average planning time of automatic plans was (32.6 ± 5.3) min, which was a large improvement if compared with (103.1±18.5) min of manual plans. Meanwhile, MUs in automatic plans were on average reduced by 97.4 ± 23.6. Table 4 Comparison of PQM total values, planning time and monitor units between manual and automatic plans. Metrics Manual plan Automatic plan P-value PQM total value 40.7±13.1 40.8±13.5 0.83 Planning time (min) 103.1±18.5 32.6±5.3 ＜ 0.001 Monitor units 789.9±234.3 692.5±210.7 ＜ 0.001 3.4 Expert review and feedback Based on a tally of 200 reviews from four experts, automatic plans were deemed to be clinically acceptable for treatment in 88% of reviews (176/200). All preference reviews are shown in the Figure 3. In terms of overall review, automatic plans were deemed superior in 36 reviews (18%), equivalent in 98 reviews (49%), and inferior in 66 reviews (33%); In terms of target coverage, 94% of automatic and manual plans were equivalent. In terms of target conformity and homogeneity, experts’ favours for automatic and manual plans varied little, with more reviews deeming that they were equivalent. Across all reviews, automatic plans typically demonstrated better heart sparing and cord sparing compared with manual plans. But lung sparing of automatic plans were considered inferior to the manual plans. 4. Discussion In the process of designing VMAT plans, planners need to repeatedly adjust objective functions, weights and other plan parameters based on personal experience. As a result, the design process is time-consuming, laborious, and there is a large subjective uncertainty. Ultimately, plan designing is inefficient and varies widely in quality. Lung cancer, as the malignant tumor with the highest incidence, has a variety of target locations, shapes and prescribed doses, which makes the above problems more prominent. As can be seen in Table 1 , among the 50 randomly selected cases in this study, four stages were included:Ⅰ, Ⅱ, Ⅲ and Ⅳ; three kinds of treatment patterns were included: postoperative radiotherapy, concurrent chemoradiotherapy, radiotherapy alone. As a result, multiple prescribed dose patterns have been developed. 12 patients had only PTV (seven kinds of prescribed dose gradients) and 38 patients were treated with the simultaneous integrated boost technique (seven kinds of prescribed dose gradients). In this case, the predictive effect of automatic planning was more demanding. Therefore, the deep learning approach was adopted in this study to establish a prediction model of radiotherapy plan for lung cancer. Different from other automatic plan prediction models, this study used anatomical structures and prescribed doses to make predictions, which can be applied to lung cancer plans with different anatomical structures and prescribed doses. Finally, 50 lung cancer patients with different tumor stages and different treatment patterns were randomly selected for robust evaluation from different perspectives in this study. And the dosimetric parameters and plan efficiency parameters of VMAT automatic plans and manual plans were compared in Pinnacle 3 TPS, which aimed to evaluate the feasibility and superiority of automatic plans in lung cancer radiotherapy. From the analysis results of targets volume, there was no significant difference in the homogeneity of two kinds of plans. In terms of targets coverage, automatic plans were even superior to manual plans. According to the analysis results of OARs, automatic plans had no advantage in lung protection, but there was a significant reduction in V30 and Dmean of heart, possibly because that the lung received the most attention in clinical plans. The lower heart dose would likely benefit patients in a long term [ 26 ]. To conduct a comprehensive assessment of two kinds of plans, it is necessary to compare the overall plan quality according to a quantitative evaluation criterion. According to the results of PQM analysis with targets and multiple important OARs, the qualities of automatic plans were better than that of manual plans, although there was no significant difference. Above all, the time saved for planners will be appreciable [ 27 ]. If each lung cancer plan saved 70.5 ± 13.2 minutes, then 50 lung cancer plans would be able to save 2 to 3 working days. This is a substantial departmental burden that would greatly benefit from automation. Moreover, the MUs of automatic plans were much less than that of manual plans. The fewer MUs can improve the efficiency of plan implementation and reduce the treatment time, which could benefit the patients to reduce discomfort and variation during the treatment process [ 28 ]. In the qualitative evaluation, only 12% of automatic plans were not accepted. There were three main reasons for unacceptability. Firstly, for case that target was closer to spinal cord, the point dose of spinal cord might exceed 40 Gy, which was clinically unacceptable. In this case, manual plans would be further optimized and spinal cord dose would be reduced, whereas there was no this step in automatic plans. Secondly, for a few automatic plans, the dose hotspot (110% of the prescribed dose) fell on the esophagus. Thirdly, for complex plans with longer and larger target, the lung dose of automatic plans exceeded clinical constraints, such as V20 > 30% in lung. Fourthly, for case that target extended to liver, the prediction model of automatic plan did not take the liver into account, resulting in a higher dose of liver than that in the manual plan. It took a planner less than 10 min hands-on time to manually fine-tune and optimize these plans. Eventually they were all accepted. In the preference review towards automatic and manual plans in terms of different metrics, the reviews that considered the two to be equivalent were the most. In 33% of reviews the automatic plans were regarded as not superior to manual plans. Because specialized planner required consultation with physician during the process of designing clinical plans to establish optimization objectives or tradeoffs. For plans with irregular target shape and large target volume, clinical tradeoffs may have been made (for instance, compromising target conformity or homogeneity in favor of sparing a surrounding organ, or compromising heart dose in favor of sparing lung tissue). As a result, manual plans would be modified in detail according to the physicians' preference, and they went through multiple stages such as independent physical review and clinical review. Finally, the plan for clinical treatment was determined. That’s the reason why some manual plans were superior to automatic plans in qualitative evaluation. In summary, automatic plans were characterized by lower modulation complexity and less delivery time when compared to manual plans, without losing the advantage in targets coverage, homogeneity and OARs dose. But there were some limitations in the present study. Firstly, the prediction of lung dose was not very beneficial, resulting in slightly higher lung dose in automatic plans than in manual plans. In addition, after quantitative evaluation, it was found that different experts have different preferences, (e.g., some experts pay more attention to the lung sparing, some to the heart sparing, and some to the target conformity and homogeneity). The current prediction model cannot meet the preferences of all experts, so different automatic planning prediction models can be developed according to the preferences of different experts. Finally, the prediction of field angle was lacking currently in this study. In the future, the method will be further improved based on the above three points. As a result, for complex and difficult plans, planners still cannot entirely rely on automatic plans. And manual fine-tuning is needed on the basis of automatic plans, which is still much faster than generating manual plans from scratch. 5. Conclusions This study describes an automatic VMAT planning method based on deep learning and demonstrates its feasibility for lung cancer. The automatic planning method is suitable for lung cancer plans with different target volumes and different treatment patterns. It can be found by comparing with the manual plans, automatic treatment planning method might be an alternative, which improves the planning and treatment efficiency without compromising the plan quality. Abbreviations DL deep learning VMAT volumetric modulated arc therapy PQM plan quality metric OAR organ at risk MUs monitor units TPS Treatment planning system KBP knowledge-based planning VTPN virtual treatment planner network MOAPN multi-objectives adjustment policy network IMRT intensity modulated radiation therapy PTV Planning Target Volume AJCC American Joint Committee on Cancer CTV Clinical Target Volume GTV Gross Target Volume DSS dose shaping structure DVH Dose Volume Histogram HI homogeneity index Declarations Acknowledgments Not applicable. Funding This work was supported by the Special Research Fund for Central Universities, Peking Union Medical College, CAMS Innovation Fund for Medical Sciences (CIFMS) [2022-I2M-C&T-B-075]; the Shanghai Pujiang Programme [23PJD014]; the Beijing Hope Run Special Fund of Cancer Foundation of China (LC2021B01); National Natural Science Foundation of China (11875320). Author information Authors and Affiliations National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing 100021, China Ningyu Wang, Yingjie Xu, Lingling Yan, Deqi Chen, Wenqing Wang, Kuo Men, Jianrong Dai, Zhiqiang Liu Department of Radiation Oncology, Fudan University Shanghai Cancer Center, Shanghai 200032, China Jiawei Fan Department of Oncology, Shanghai Medical College Fudan University, Shanghai 200032, China Jiawei Fan Shanghai Key Laboratory of Radiation Oncology, Shanghai 200032, China Jiawei Fan Contributions Ningyu Wang: Conceptualization, Methodology, Formal analysis, Investigation, Data Curation, Writing - Original Draft, Visualization, Visualization Jiawei Fan: Methodology, Software, Yingjie Xu: Validation, Formal analysis Lingling Yan: Validation, Formal analysis Deqi Chen: Formal analysis, Data Curation Wenqing Wang: Validation, Formal analysis Kuo Men: Writing - Review & Editing, Supervision Jianrong Dai: Conceptualization, Methodology, Writing - Review & Editing, Visualization, Supervision, Funding acquisition Zhiqiang Liu: Conceptualization, Methodology, Software, Investigation, Writing - Review & Editing, Visualization, Funding acquisition, Project administration Corresponding author Correspondence to Zhiqiang Liu and Jianrong Dai. Ethics declarations Ethics approval and consent to participate This study was performed in line with the principles of the Declaration of Helsinki. Approval was granted by the Independent Ethics Committee of National Cancer Center/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College. All patients were informed and signed relevant informed consent documents. Consent for publication The authors consent to publish the manuscript in its current form. Competing interests The authors declare that they have no competing interests. References Bray F, Ferlay J, Soerjomataram I, Siegel RL, Torre LA, Jemal A. vol 68, pg 394,. Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries (2018). Ca-a Cancer Journal for Clinicians. 2020; 70:313-. Zhang D, Yuan Z, Hu P, Yang Y. Automatic treatment planning for cervical cancer radiation therapy using direct three-dimensional patient anatomy match. J Appl Clin Med Phys. 2022; 23. Della Gala G, Dirkx MLP, Hoekstra N, Fransen D, Lanconelli N, van de Pol M, et al. Fully automated VMAT treatment planning for advanced-stage NSCLC patients. Strahlenther Onkol. 2017;193:402–9. Swanick CW, Lin SH, Sutton J, Naik NS, Allen PK, Levy LB, et al. Use of Simultaneous Radiation Boost Achieves High Control Rates in Patients With Non-Small-Cell Lung Cancer Who Are Not Candidates for Surgery or Conventional Chemoradiation. Clin Lung Cancer. 2015;16:156–63. Hussein M, Heijmen BJM, Verellen D, Nisbet A. Automation in intensity modulated radiotherapy treatment planning-a review of recent innovations. Br J Radiol. 2018; 91. Buergy D, Sharfo AWM, Heijmen BJM, Voet PWJ, Breedveld S, Wenz F et al. Fully automated treatment planning of spinal metastases - A comparison to manual planning of Volumetric Modulated Arc Therapy for conventionally fractionated irradiation. Radiat Oncol. 2017; 12. Xhaferllari I, Wong E, Bzdusek K, Lock M, Chen JZ. Automated IMRT planning with regional optimization using planning scripts. J Appl Clin Med Phys. 2013;14:176–91. Fogliata A, Cozzi L, Reggiori G, Stravato A, Lobefalo F, Franzese C et al. RapidPlan knowledge based planning: iterative learning process and model ability to steer planning strategies. Radiat Oncol. 2019; 14. Wang S, Zheng DD, Zhang C, Ma RT, Bennion NR, Lei Y, et al. Automatic planning on hippocampal avoidance whole-brain radiotherapy. Med Dosim. 2017;42:63–8. Kamran SC, Mueller BS, Paetzold P, Dunlap J, Niemierko A, Bortfeld T, et al. Multi-criteria optimization achieves superior normal tissue sparing in a planning study of intensity-modulated radiation therapy for RTOG 1308-eligible non-small cell lung cancer patients. Radiother Oncol. 2016;118:515–20. Good D, Lo J, Lee WR, Wu QJ, Yin FF, Das SK. A Knowledge-Based Approach to Improving and Homogenizing Intensity Modulated Radiation Therapy Planning Quality Among Treatment Centers: An Example Application to Prostate Cancer Planning. Int J Radiat Oncol Biol Phys. 2013;87:176–81. Xu H, Lu J, Wang J, Fan J, Hu W. Implement a knowledge-based automated dose volume histogram prediction module in Pinnacle 3 treatment planning system for plan quality assurance and guidance. J Appl Clin Med Phys. 2019;20:134–40. Kearney V, Chan JW, Wang T, Perry A, Descovich M, Morin O, et al. DoseGAN: a generative adversarial network for synthetic dose prediction using attention-gated discrimination and generation. Sci Rep. 2020;10:11073. Fan J, Wang J, Chen Z, Hu C, Zhang Z, Hu W. Automatic treatment planning based on three-dimensional dose distribution predicted from deep learning technique. Med Phys. 2019;46:370–81. 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. Momin S, Fu Y, Lei Y, Roper J, Bradley JD, Curran WJ, et al. Knowledge-based radiation treatment planning: A data-driven method survey. J Appl Clin Med Phys. 2021;22:16–44. Shen CY, Nguyen D, Chen LY, Gonzalez Y, McBeth R, Qin N, et al. Operating a treatment planning system using a deep-reinforcement learning-based virtual treatment planner for prostate cancer intensity-modulated radiation therapy treatment planning. Med Phys. 2020;47:2329–36. Wang H, Bai X, Wang Y, Lu Y, Wang B. An integrated solution of deep reinforcement learning for automatic IMRT treatment planning in non-small-cell lung cancer. Front Oncol. 2023; 13. Liu Z, Fan J, Li M, Yan H, Hu Z, Huang P, et al. A deep learning method for prediction of three-dimensional dose distribution of helical tomotherapy. Med Phys. 2019;46:1972–83. Hodapp N, [The ICRU. Report 83: prescribing, recording and reporting photon-beam intensity-modulated radiation therapy (IMRT)]. Strahlenther Onkol. 2012;188:97–9. Kataria T, Sharma K, Subramani V, Karrthick KP, Bisht SS. Homogeneity Index: An objective tool for assessment of conformal radiation treatments. J Med Phys. 2012;37:207–13. Nelms BE, Robinson G, Markham J, Velasco K, Boyd S, Narayan S, et al. Variation in external beam treatment plan quality: An inter-institutional study of planners and planning systems. Practical radiation oncology. 2012;2:296–305. Bohsung J, Gillis S, Arrans R, Bakai A, De Wagter C, Knoos T, et al. IMRT treatment planning - A comparative inter-system and intor-centre planning exercise of the ESTRO QUASIMODO group. Radiother Oncol. 2005;76:354–61. Xia W, Liu Z, Yan L, Han F, Hu Z, Tian Y, et al. A longitudinal evaluation of improvements in treatment plan quality for lung cancer with volumetric modulated arc therapy. J Appl Clin Med Phys. 2020;21:33–43. McIntosh C, Conroy L, Tjong MC, Craig T, Bayley A, Catton C, et al. Clinical integration of machine learning for curative-intent radiation treatment of patients with prostate cancer. Nat Med. 2021;27:999–1005. Ling C, Han X, Zhai P, Xu H, Chen J, Wang J et al. A hybrid automated treatment planning solution for esophageal cancer. Radiat Oncol. 2019;14. Mitchell RA, Wai P, Colgan R, Kirby AM, Donovan EM. Improving the efficiency of breast radiotherapy treatment planning using a semi-automated approach. J Appl Clin Med Phys. 2017;18:18–24. Zhang Q, Peng Y, Song X, Yu H, Wang L, Zhang S. Dosimetric evaluation of automatic and manual plans for early nasopharyngeal carcinoma to radiotherapy. Med Dosim. 2020;45:E13–E20. Additional Declarations No competing interests reported. Supplementary Files SupplementaryFiles.docx Cite Share Download PDF Status: Posted Version 1 posted 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-3872969","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":268915055,"identity":"f41ac220-8827-49bd-95af-2ada72b1359f","order_by":0,"name":"Ningyu Wang","email":"","orcid":"","institution":"National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College","correspondingAuthor":false,"prefix":"","firstName":"Ningyu","middleName":"","lastName":"Wang","suffix":""},{"id":268915056,"identity":"b4f472f3-d1c8-4ea5-ad09-e4c67816bf58","order_by":1,"name":"Jiawei Fan","email":"","orcid":"","institution":"Department of Radiation Oncology, Fudan University Shanghai Cancer Center, Shanghai 200032, China","correspondingAuthor":false,"prefix":"","firstName":"Jiawei","middleName":"","lastName":"Fan","suffix":""},{"id":268915057,"identity":"4c716a89-2b21-49a7-8cc6-b3c3a7dc8dfe","order_by":2,"name":"Yingjie Xu","email":"","orcid":"","institution":"National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College","correspondingAuthor":false,"prefix":"","firstName":"Yingjie","middleName":"","lastName":"Xu","suffix":""},{"id":268915058,"identity":"519ec30b-482c-45f5-bb51-e9f01a69e544","order_by":3,"name":"Lingling Yan","email":"","orcid":"","institution":"National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College","correspondingAuthor":false,"prefix":"","firstName":"Lingling","middleName":"","lastName":"Yan","suffix":""},{"id":268915059,"identity":"c2a3c64b-4ac1-4bed-91b1-fd0f073cae85","order_by":4,"name":"Deqi Chen","email":"","orcid":"","institution":"National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College","correspondingAuthor":false,"prefix":"","firstName":"Deqi","middleName":"","lastName":"Chen","suffix":""},{"id":268915060,"identity":"7df7e1c8-dd35-4ade-bbf9-3245c870ff83","order_by":5,"name":"Wenqing Wang","email":"","orcid":"","institution":"National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College","correspondingAuthor":false,"prefix":"","firstName":"Wenqing","middleName":"","lastName":"Wang","suffix":""},{"id":268915061,"identity":"0907a6be-6b84-4f64-a018-fcd541121512","order_by":6,"name":"Kuo Men","email":"","orcid":"","institution":"National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College","correspondingAuthor":false,"prefix":"","firstName":"Kuo","middleName":"","lastName":"Men","suffix":""},{"id":268915062,"identity":"cf4ab019-40f5-4bc8-9db3-59693452e7fa","order_by":7,"name":"Jianrong Dai","email":"","orcid":"","institution":"National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College","correspondingAuthor":false,"prefix":"","firstName":"Jianrong","middleName":"","lastName":"Dai","suffix":""},{"id":268915063,"identity":"4e6e8b20-ee73-4080-b943-ac37f6887174","order_by":8,"name":"Zhiqiang Liu","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABBUlEQVRIiWNgGAWjYDACCcYGhgQQgx3I+MAgQYoWZsYGxhnEaYExmIGIhxh3yc9ubrzxoOaOXQMzc5u0TY2FPP+04w8YftQwyJvj0MI452CzRcKxZ8kNzIxt0jnHJAxn3M4xYOw5xmC4swG7FmaJxDaJBLbDyQwgLbkNEgkMt3MYGHiBHjQ4gF0LG1jLP6gWS6AW+dvpDxj/4tHCA9KS2HbYDqyFEajF4HaCATM+WyQkEpstEvsOJ7AxMzZb9gD9shHol8MyQMYGHFrkZ6Q/vPnj22F7fvb2hzd+1NTJy91Of/jwTY2NPC5bwDYBcWIbssgBBgJxCpK1x6tiFIyCUTAKRjYAAHqEVCO5Qpx/AAAAAElFTkSuQmCC","orcid":"","institution":"National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College","correspondingAuthor":true,"prefix":"","firstName":"Zhiqiang","middleName":"","lastName":"Liu","suffix":""}],"badges":[],"createdAt":"2024-01-17 12:59:15","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-3872969/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-3872969/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":50171516,"identity":"be833f01-47d7-4f8b-a7cb-ff7eb6f2f334","added_by":"auto","created_at":"2024-01-25 15:46:38","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":73229,"visible":true,"origin":"","legend":"\u003cp\u003eThe workflow of integrating the DL model into TPS.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-3872969/v1/c9f708f9e881fe007b8103ab.png"},{"id":50172084,"identity":"837186ee-d518-4e3f-98da-c27ec1ecfa6f","added_by":"auto","created_at":"2024-01-25 15:54:38","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":57158,"visible":true,"origin":"","legend":"\u003cp\u003eStatistical plots of the DVH metrics for lung and heart (a) and the maximum and mean doses for serial and parallel OARs, respectively (b).\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-3872969/v1/1acc62d4ca033a1d29a1a497.png"},{"id":50171518,"identity":"dcc9988e-3f85-46ec-8dee-998650d6cdcd","added_by":"auto","created_at":"2024-01-25 15:46:38","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":3379823,"visible":true,"origin":"","legend":"\u003cp\u003eStatistical plots of the preference of four experts towards automatic and manual plans in terms of different metrics.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-3872969/v1/36e141032087247517a70eb5.png"},{"id":50419446,"identity":"759f17bd-2f42-4ad6-9fb2-892b540006ef","added_by":"auto","created_at":"2024-01-31 09:09:57","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":710358,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-3872969/v1/dfaec734-71db-4833-be53-d2ee91c15fef.pdf"},{"id":50171519,"identity":"769bd98a-bb54-466e-81d5-487888aeb915","added_by":"auto","created_at":"2024-01-25 15:46:38","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":1793282,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFiles.docx","url":"https://assets-eu.researchsquare.com/files/rs-3872969/v1/e714c3082ae17d856ce10eaa.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Clinical implementation and evaluation of deep learning-assisted automatic radiotherapy treatment planning for lung cancer","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eLung cancer is the most morbidity and mortality cancer globally, accounting for about 21.3% of all malignant tumors [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Radiotherapy is one of the important means of lung cancer treatment and it plays an increasingly prominent role. Approximately 70% of lung cancer patients need radiotherapy during the treatment process. Radiotherapy is applicable to different types of lung cancer at different stages. Middle- and early-stage lung cancer can be treated by radical radiotherapy or combined radiotherapy and chemotherapy, and advanced lung cancer can be treated by palliative radiotherapy. As a result, the number of lung cancer radiotherapy plans that planners need to design is increasing clinically. Moreover, the targets of lung cancer plans are complex and variable, which makes planning difficult.\u003c/p\u003e \u003cp\u003eIn the design process of volumetric modulated arc therapy (VMAT) plans, planners need to repeatedly adjust the objective functions, weights and other plan parameters based on personal experience. Therefore, it is an iterative trial-and-error process [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. The process is time consuming and the quality of the final treatment plan may depend highly on the experience and skills of planners, the complexity of the case, and the available time [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. In addition to this, the development of advanced treatment techniques, such as simultaneous integrated boost irradiation [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e], will lead to a workload growth of planning design.\u003c/p\u003e \u003cp\u003eThere has been increasing interest in automatic plans in recent years, with the aim of improving plan quality and reducing planning time. Some studies have shown that automatic plans can considerably maintain the plan quality while dramatically reduce the planning workload compared to manual plans [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Automatic plan can potentially decrease the amount of effort involved in manual \"trial-and-error\" methods by around 50% [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. Currently, there are script-based planning method [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e], as well as automatic planning modules that have been integrated into Treatment planning system (TPS). Examples of these modules include RapidPlan in Eclipse [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e], the volume-driven automatic planning platform in Pinnacle\u003csup\u003e3\u003c/sup\u003e [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e] and Multi-Criteria Optimization in RayStation [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. These planning modules, with knowledge-based planning (KBP) methods [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e], utilize databases of patient plans to establish prediction models for estimating dose-volume histograms and dose maps. In recent years, deep learning (DL) methods have achieved significant success in predicting radiotherapy plans [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e, \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e], leading to enhanced accuracy and efficiency. In contrast to traditional KBP methods with handcrafted features, deep learning methods can automatically learn image features from raw data, such as CT images, contours, and dose maps, which are tailored specifically to the prediction task [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. A virtual treatment planner network (VTPN) was built by Shen et al [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e] using DL to operate a TPS by adjusting treatment planning parameters in it. Wang et al [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e] proposed a novel algorithm named as multi-objectives adjustment policy network (MOAPN) and trained to learn how to adjust multiple optimization objectives in commercial Eclipse TPS based on the multi-agent DL scheme. The existing prediction models are focused on single target and treatment pattern, and there is a lack of a generic prediction model for radiotherapy plans with multiple targets and different treatment patterns. This study is based on a deep learning approach to establish an intelligent radiotherapy method for lung cancer that is applicable to any individualized anatomical structure and prescribed dose pattern, under the condition that the target location and shape are diverse. In addition, many existing automatic planning methods are for intensity modulated radiation therapy (IMRT) technology [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e], while this study is for VMAT technology.\u003c/p\u003e \u003cp\u003eThe purpose of this study is to realize an integration of DL as a standard workflow into the current TPS to test the feasibility of clinical deployment for lung cancer. We retrospectively recovered 50 previously manually designed, peer reviewed and delivered patient VMAT treatment plans. For each patient, we designed automatic VMAT plans. Both automatic and manual plans were compared from different perspectives to evaluate the plan quality and efficiency, and ultimately to investigate the feasibility of clinical implementation of DL-assisted automatic treatment planning.\u003c/p\u003e"},{"header":"2. Materials and methods","content":"\u003cp\u003e2.1 Development of the DL model and integration into TPS\u003c/p\u003e\n\u003cp\u003eThe DL method was trained on 235 high-quality, previously approved, peer reviewed and delivered patient treatment plans which were generated by an experienced team dedicated to the treatment of lung cancer at a large academic cancer center. We developed a DL model for lung cancer dose prediction based on the architecture in our previous study [\u003cspan class=\"CitationRef\"\u003e19\u003c/span\u003e]. We retrained the DL model using these lung cancer datasets. The inputs to the DL model were the contours, prescriptions of targets and the outputs were the doses for lung cancer cases. The details of datasets, architecture, inputs and outputs, training and validation process of the DL model can refer to Supplementary Material 1.\u003c/p\u003e\n\u003cp\u003eThe trained DL model was used for this retrospective study phase. For a new patient in retrospective dataset, the model automatically learned features from patient and delineated Planning Target Volumes (PTVs), OARs and Body to predict patient-specific doses. And the model was integrated into clinical Pinnacle\u003csup\u003e3\u003c/sup\u003e TPS to achieve setting inverse optimization objectives of targets and OARs, with the help of optimization and dose calculation engines from TPS to realize automatic treatment planning. The integration workflow demonstrated in Fig. \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e was implemented in clinical practice. It mainly involved three parts: i) DICOM-RT data preparation and transfer, ii) data preprocessing, dose prediction, and data post-processing, iii) data transfer and automatic design plan. Intelligent planning system was implemented in third-party servers and integrated into clinical Pinnacle\u003csup\u003e3\u003c/sup\u003e TPS.\u003c/p\u003e\n\u003cp\u003e2.2 Patient data\u003c/p\u003e\n\u003cp\u003eBetween January 2022 and May 2022, a cohort of additional 50 treated plans with different tumor stages and prescribed doses were randomly and retrospectively selected. All patients were staged according to the American Joint Committee on Cancer (AJCC) Manual for Staging of Cancer, 8th edition. Patients had previously been scanned using either a Philips Brilliance Big Bore or Siemens Syngo CT scanner. These CT images were transferred to the Pinnacle\u003csup\u003e3\u003c/sup\u003e TPS (version 16.2, Philips Medical Systems, Fitchburg, WI). The targets were then manually delineated by the radiation oncologist on these CT images with the help of a contrast-enhanced diagnostic CT. Generally, Clinical Target Volume (CTV) was expanded by 5mm from Gross Target Volume (GTV) and PTV was expanded by 5mm from CTV. If Planning Gross Target Volume (PGTV) needed to be delineated, a 5 mm margin between the PTV and PGTV were considered, though these margins could have been subjected to adaptations for each specific clinical case. The OARs included lung, heart, cord, cord PRV, trachea, esophagus, liver and so on. The dose constraint applied for planning are provided in Supplementary Material 2.\u003c/p\u003e\n\u003cp\u003e2.3 Planning process\u003c/p\u003e\n\u003cp\u003eThe retrospectively selected plans were previously manual designed and approved by several planners with 3\u0026ndash;10 years working experiences and were all peer reviewed and delivered. The manual designing process of these plans included several steps: I) the planner first defined CT density table, removed the CT couch, and created an isocenter of the beams in the Pinnacle\u003csup\u003e3\u003c/sup\u003e TPS. II) A Pinnacle\u003csup\u003e3\u003c/sup\u003e script was then used for generating dose shaping structures (DSS), including rings of the targets (Ring), normal tissue (NT), blocks, the fan areas above and below the targets (Fan up and Fan down) (typical example of DSS is shown in Supplementary Material 3). III) Beam angles were determined by the planners according to the position of targets in relation to OARs. Usually, round-trip arcs were created, ranging from 181\u0026deg; to 30\u0026deg; for tumors located in right lung and from 330\u0026deg; to 180\u0026deg; for tumors located in left lung. IV) Other setting parameters were as follows: beam energy 6 MV, dose computational grid resolution 4 mm, control point spacing 4\u0026deg;, and the leaf motion constrained to 0.5 cm/deg. V) The inverse optimization objectives of targets, OARs and DSS were manually defined by planners according to the clinical requirements, and the trivial process of adjusting the optimization parameters were repeated until the plan can be used for clinic. These processes were heavily relied on the experience of the planners. VI) The senior physicists and radiotherapy physicians determined the final delivered plan.\u003c/p\u003e\n\u003cp\u003eUnlike the manual plan designing process, the deep learning-assisted plan can automatically predict patient-specific inverse optimization objectives according to their prescriptions and geometries. For these retrospective selected plans, we redesigned the automatic plans. The process of I)-IV) was basically same as manual plan design. In step II), manual plan called the script of DSS, and automatic plan called the smartplan script including DSS and objective functions (the descriptions of objective functions are provided in Supplementary Material 4). Primarily, the biggest difference between manual and automatic plan was in step V). The process of V) in automatic plan was set based on the patient-specific prediction results. And the automatic plans were obtained by optimizing twice based on the predicted optimization objectives.\u003c/p\u003e\n\u003cp\u003e\u003cspan\u003e\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003e2.4 Evaluation Metrics\u003c/p\u003e\u003cspan\u003e\n \u003cp\u003e2.4.1 Quantitative evaluation\u003c/p\u003e\n\u003c/span\u003e\n\u003cp\u003e\u003c/p\u003e\n\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\n \u003ch2\u003eDosimetric analysis\u003c/h2\u003e\n \u003cp\u003eAccording to ICRU 83 [\u003cspan class=\"CitationRef\"\u003e20\u003c/span\u003e], dosimetric evaluations for targets coverage and OARs sparing of two kinds of plans were performed with Dose Volume Histogram (DVH). The homogeneity index (HI) [\u003cspan class=\"CitationRef\"\u003e21\u003c/span\u003e], D98, D95, D50, and D2, were used for target coverage. The definition of HI is as follows:\u003c/p\u003e\n \u003cdiv id=\"Equa\" class=\"Equation\"\u003e\n \u003cdiv class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e$$HI=({D}_{2}-{D}_{98})/{D}_{50}\\times 100\\%$$\u003c/div\u003e\n \u003c/div\u003e\n \u003cp\u003ewhere D98, D95, D50, and D2 are the dose value corresponding to 98%, 95%, 50%, and 2% of the targets on the DVH, respectively. The OARs metrics used in comparison included: V5, V20, V30 and Dmean for Lung all; V5 for contralateral lung; V30, V40 and Dmean for heart; Dmax for cord, cord PRV, trachea and esophagus; Dmean for liver. The paired t-test was used for comparison with significance determined at the level of P\u0026thinsp;\u0026lt;\u0026thinsp;0.05.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\n \u003ch2\u003ePlanning Time and MUs\u003c/h2\u003e\n \u003cp\u003eIn addition, the planning time and MUs for two kinds of plans were recorded and evaluated. The automatic planning time was recorded directly during the design process, which included the time of step I)-V) in section 2.3. The manual planning time was estimated based on the effective working time, which started when the plans were initiated to design and finished when the plans were determined to be submitted for physical review and clinical review. The biggest difference in time between manual plan and automatic plan was in step V). Manual plan spent most of its time on manually adjusting optimization parameters and carrying out optimizing, while the automatic plan can be directly optimized based on predicted objective functions called by smartplan script. The MUs of both plans can be recorded directly in Pinnacle\u003csup\u003e3\u003c/sup\u003e TPS.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\n \u003ch2\u003ePQM quality assessment\u003c/h2\u003e\n \u003cp\u003eTo quantitatively assess the plan quality of automatic and manual plans, PQM [\u003cspan class=\"CitationRef\"\u003e22\u003c/span\u003e] was used. Based on the concept of quality score proposed by Bohsung et al. [\u003cspan class=\"CitationRef\"\u003e23\u003c/span\u003e] and the new PQM method with related submetrics for lung cancer [\u003cspan class=\"CitationRef\"\u003e24\u003c/span\u003e] this study scored each structure of interest using a score function based on targets/OARs clinical constraints. When there was only PTV, the submetric of target was the HI of PTV. When PTV and PGTV were present, the submetrics of targets were the HI of PTV and PGTV. OAR-related submetrics included V5, V20, Dmean for lung; V30, V40, Dmean for heart and Dmax for cord, cord PRV. The lower limit, upper limit, and PQM value range for different submetrics are shown in Supplementary Material 5. For the target, the closer the HI value is to 0, the higher the PQM value is. For each OAR, the closer the dose index is to the upper limit, the lower the PQM value is, and the PQM value was 0 when dose index exceeds the upper limit.\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003e2.4.2. Qualitative evaluation\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003eBlinded expert reviews [\u003cspan class=\"CitationRef\"\u003e25\u003c/span\u003e] were implemented by four expert reviewers (one radiotherapy physician and three physicists) to assess the clinical applicability of automatic plan and there were 50 cases and 200 reviews. The reviewers independently assessed dosimetric preference towards automatic and manual plans as well as overall clinical acceptability for automatic plans. And the preference towards automatic and manual plans were evaluated in terms of overall review, target coverage, target conformity, target homogeneity, lung sparing, heart sparing, cord sparing and OARs sparing. The details of expert review form can refer to Supplementary Material 6.\u003c/p\u003e\n\u003c/div\u003e"},{"header":"3. Results","content":"\u003cp\u003e3.1 Patients\u003c/p\u003e\n\u003cp\u003eDifferent stages and radiotherapy patterns corresponded to different prescribed doses. In 50 patients, simultaneous integrated boost technique was applied to 38 patients whose prescribed doses of PGTV ranged from 50 to 67.2 Gy in 20\u0026ndash;30 fractions and prescribed doses of PTV were ranged from 40 to 54 Gy in 20\u0026ndash;30 fractions. For the other 12 patients, only PTV was prescribed. And the doses of PTV were ranged from 45 to 60 Gy in 15\u0026ndash;30 fractions. The patient characteristics are summarized in Table 1.\u003c/p\u003e\n\u003cp\u003eTable 1 Clinical features of the 50 patients with lung cancer.\u003c/p\u003e\n\u003cdiv align=\"Left\"\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"661\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd width=\"46.74735249621785%\" colspan=\"2\"\u003e\n \u003cp\u003eCharacteristics\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.05446293494705%\"\u003e\n \u003cp\u003eNumber of patients\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"29.1981845688351%\"\u003e\n \u003cp\u003e%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"22.087745839636913%\"\u003e\n \u003cp\u003eTumor stages\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.659606656580937%\"\u003e\n \u003cp\u003eⅠ\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.05446293494705%\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"29.1981845688351%\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"22.087745839636913%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.659606656580937%\"\u003e\n \u003cp\u003eⅡ\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.05446293494705%\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"29.1981845688351%\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"22.087745839636913%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.659606656580937%\"\u003e\n \u003cp\u003eⅢ\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.05446293494705%\"\u003e\n \u003cp\u003e37\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"29.1981845688351%\"\u003e\n \u003cp\u003e74\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"22.087745839636913%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.659606656580937%\"\u003e\n \u003cp\u003eⅣ\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.05446293494705%\"\u003e\n \u003cp\u003e8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"29.1981845688351%\"\u003e\n \u003cp\u003e16\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"22.087745839636913%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003cp\u003eRadiotherapy mode\u003c/p\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.659606656580937%\"\u003e\n \u003cp\u003ePostoperative radiotherapy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.05446293494705%\"\u003e\n \u003cp\u003e7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"29.1981845688351%\"\u003e\n \u003cp\u003e14\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"22.087745839636913%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.659606656580937%\"\u003e\n \u003cp\u003eConcurrent\u003c/p\u003e\n \u003cp\u003echemoradiotherapy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.05446293494705%\"\u003e\n \u003cp\u003e13\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"29.1981845688351%\"\u003e\n \u003cp\u003e26\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"22.087745839636913%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.659606656580937%\"\u003e\n \u003cp\u003eRadiotherapy alone\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.05446293494705%\"\u003e\n \u003cp\u003e30\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"29.1981845688351%\"\u003e\n \u003cp\u003e60\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"22.087745839636913%\"\u003e\n \u003cp\u003ePrescribed dose\u003c/p\u003e\n \u003cp\u003e(PTV single)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.659606656580937%\"\u003e\n \u003cp\u003e40~60 Gy, 15~30 fraction\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.05446293494705%\"\u003e\n \u003cp\u003e12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"29.1981845688351%\"\u003e\n \u003cp\u003e24\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"22.087745839636913%\"\u003e\n \u003cp\u003ePrescribed dose\u003c/p\u003e\n \u003cp\u003e(PTV and PGTV)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.659606656580937%\"\u003e\n \u003cp\u003ePTV: 40~54 Gy,\u003c/p\u003e\n \u003cp\u003e20~30 fraction\u003c/p\u003e\n \u003cp\u003ePGTV: 50~67.2 Gy,\u003c/p\u003e\n \u003cp\u003e20~30 fraction\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.05446293494705%\"\u003e\n \u003cp\u003e38\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"29.1981845688351%\"\u003e\n \u003cp\u003e76\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"22.087745839636913%\"\u003e\n \u003cp\u003eGender\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.659606656580937%\"\u003e\n \u003cp\u003eMale\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.05446293494705%\"\u003e\n \u003cp\u003e36\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"29.1981845688351%\"\u003e\n \u003cp\u003e72\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"22.087745839636913%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.659606656580937%\"\u003e\n \u003cp\u003eFemale\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.05446293494705%\"\u003e\n \u003cp\u003e14\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"29.1981845688351%\"\u003e\n \u003cp\u003e28\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"22.087745839636913%\"\u003e\n \u003cp\u003eLocation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.659606656580937%\"\u003e\n \u003cp\u003eRight\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.05446293494705%\"\u003e\n \u003cp\u003e25\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"29.1981845688351%\"\u003e\n \u003cp\u003e50\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"22.087745839636913%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.659606656580937%\"\u003e\n \u003cp\u003eLeft\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.05446293494705%\"\u003e\n \u003cp\u003e25\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"29.1981845688351%\"\u003e\n \u003cp\u003e50\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e3.2 Targets coverage and homogeneity\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTable 2 compares the target volume dose metrics between automatic plans and manual plans. For the PTV only, there was no significant difference for D98, D95, D50, D2 and HI values between two kinds of plans. For cases with PTV and PGTV targets, HI values of automatic plans were comparable to that of manual plans (P=0.34 for PTV, P=0.50 for PGTV). While HI values do not vary obviously, targets coverage of automatic plans were considered equivalent or better than the corresponding manual plans. The median dose (D50%) in PTV was on average improved by 1.0 Gy \u0026plusmn; 0.2 Gy (P = 0.001) and the near-maximum dose (D2%) in PTV by 0.7 Gy \u0026plusmn; 0.1 Gy (P = 0.02). The near-minimum dose (D98%) in PGTV was on average improved by 0.4 Gy \u0026plusmn; 0.4 Gy (P = 0.01) and the median dose (D50%) in PGTV by 0.9 Gy \u0026plusmn; 0.1 Gy (P = 0.001). Overall, both sets of plans achieved similar targets coverage and homogeneity.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;Table 2 Comparison of the targets dose metrics between manual and automatic plans.\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"657\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd width=\"25.875190258751903%\"\u003e\n \u003cp\u003eStructures\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\"\u003e\n \u003cp\u003eMetrics\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.61339421613394%\"\u003e\n \u003cp\u003eManual plan\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\"\u003e\n \u003cp\u003eAutomatic plan\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.068493150684931%\"\u003e\n \u003cp\u003eP-value\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"25.875190258751903%\"\u003e\n \u003cp\u003ePTV (single)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\"\u003e\n \u003cp\u003eHI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.61339421613394%\" valign=\"top\"\u003e\n \u003cp\u003e0.14\u0026plusmn;0.04\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003e0.14\u0026plusmn;0.04\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.068493150684931%\"\u003e\n \u003cp\u003e0.09\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"25.875190258751903%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\"\u003e\n \u003cp\u003eD98 (Gy)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.61339421613394%\"\u003e\n \u003cp\u003e49.8\u0026plusmn;4.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003e48.8\u0026plusmn;5.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.068493150684931%\"\u003e\n \u003cp\u003e0.28\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"25.875190258751903%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\"\u003e\n \u003cp\u003eD95 (Gy)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.61339421613394%\" valign=\"top\"\u003e\n \u003cp\u003e51.3\u0026plusmn;5.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003e51.3\u0026plusmn;5.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.068493150684931%\"\u003e\n \u003cp\u003e0.50\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"25.875190258751903%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\"\u003e\n \u003cp\u003eD50 (Gy)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.61339421613394%\" valign=\"top\"\u003e\n \u003cp\u003e54.3\u0026plusmn;5.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003e54.8\u0026plusmn;5.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.068493150684931%\"\u003e\n \u003cp\u003e0.26\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"25.875190258751903%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\"\u003e\n \u003cp\u003eD2 (Gy)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.61339421613394%\" valign=\"top\"\u003e\n \u003cp\u003e56.7\u0026plusmn;5.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003e57.4\u0026plusmn;6.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.068493150684931%\"\u003e\n \u003cp\u003e0.09\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"25.875190258751903%\"\u003e\n \u003cp\u003ePTV (PTV and PGTV)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\"\u003e\n \u003cp\u003eHI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.61339421613394%\" valign=\"top\"\u003e\n \u003cp\u003e0.31\u0026plusmn;0.05\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003e0.32\u0026plusmn;0.04\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.068493150684931%\"\u003e\n \u003cp\u003e0.34\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"25.875190258751903%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\"\u003e\n \u003cp\u003eD98 (Gy)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.61339421613394%\" valign=\"top\"\u003e\n \u003cp\u003e47.8\u0026plusmn;2.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003e47.9\u0026plusmn;2.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.068493150684931%\"\u003e\n \u003cp\u003e0.66\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"25.875190258751903%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\"\u003e\n \u003cp\u003eD95 (Gy)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.61339421613394%\" valign=\"top\"\u003e\n \u003cp\u003e50.8\u0026plusmn;2.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003e51.1\u0026plusmn;3.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.068493150684931%\"\u003e\n \u003cp\u003e0.28\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"25.875190258751903%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\"\u003e\n \u003cp\u003eD50 (Gy)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.61339421613394%\" valign=\"top\"\u003e\n \u003cp\u003e58.7\u0026plusmn;3.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003e59.7\u0026plusmn;3.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.068493150684931%\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.001\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"25.875190258751903%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\"\u003e\n \u003cp\u003eD2 (Gy)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.61339421613394%\" valign=\"top\"\u003e\n \u003cp\u003e66.0\u0026plusmn;3.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003e66.7\u0026plusmn;3.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.068493150684931%\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.02\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"25.875190258751903%\"\u003e\n \u003cp\u003ePGTV (PTV and PGTV)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\"\u003e\n \u003cp\u003eHI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.61339421613394%\" valign=\"top\"\u003e\n \u003cp\u003e0.12\u0026plusmn;0.02\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003e0.11\u0026plusmn;0.03\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.068493150684931%\"\u003e\n \u003cp\u003e0.50\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"25.875190258751903%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\"\u003e\n \u003cp\u003eD98 (Gy)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.61339421613394%\" valign=\"top\"\u003e\n \u003cp\u003e59.3\u0026plusmn;2.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003e59.7\u0026plusmn;2.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.068493150684931%\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.01\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"25.875190258751903%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\"\u003e\n \u003cp\u003eD95 (Gy)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.61339421613394%\" valign=\"top\"\u003e\n \u003cp\u003e60.7\u0026plusmn;2.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003e60.8\u0026plusmn;4.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.068493150684931%\"\u003e\n \u003cp\u003e0.32\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"25.875190258751903%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\"\u003e\n \u003cp\u003eD50 (Gy)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.61339421613394%\" valign=\"top\"\u003e\n \u003cp\u003e63.5\u0026plusmn;2.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003e64.4\u0026plusmn;3.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.068493150684931%\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.001\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"25.875190258751903%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\"\u003e\n \u003cp\u003eD2 (Gy)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.61339421613394%\" valign=\"top\"\u003e\n \u003cp\u003e66.7\u0026plusmn;3.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003e67.1\u0026plusmn;3.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.068493150684931%\"\u003e\n \u003cp\u003e0.24\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e3.3 OARs sparing\u003c/p\u003e\n\u003cp\u003eFigure 2(a) shows a boxplot of DVH metrics between the automatic plans and manual plans for Lung and Heart. As is shown in Figure 2(b), the maximum and mean doses were considered for serial and parallel OARs, respectively. No significant difference was observed in the higher dose (V20, V30) of lung all and the lower dose (V5) of contralateral lung. But the V5 values of lung all in automatic plans were higher. Manual plans had better sparing for lung, while automatic plans had better sparing for heart. As can be seen from Table 3 and Figure 2, both V30 and Dmean were significantly lower in the automatic plans than in the manual plans (P=0.03). When considering other normal tissues sparing, including cord, trachea, esophagus and liver, it showed no significant difference. Moreover, Supplementary Material 7 demonstrates a typical case of better protection for OARs in automatic plan.\u003c/p\u003e\n\u003cp\u003eTable 3 Comparison of the OARs dose metrics between manual and automatic plans.\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"657\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd width=\"25.875190258751903%\"\u003e\n \u003cp\u003eStructures\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\"\u003e\n \u003cp\u003eMetrics\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.61339421613394%\"\u003e\n \u003cp\u003eManual plan\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\"\u003e\n \u003cp\u003eAutomatic plan\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.068493150684931%\"\u003e\n \u003cp\u003eP-value\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"25.875190258751903%\"\u003e\n \u003cp\u003eLung all\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\"\u003e\n \u003cp\u003eV5 (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.61339421613394%\"\u003e\n \u003cp\u003e32.6\u0026plusmn;6.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\"\u003e\n \u003cp\u003e34.4\u0026plusmn;10.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.068493150684931%\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.003\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"25.875190258751903%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003eV20 (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.61339421613394%\"\u003e\n \u003cp\u003e17.4\u0026plusmn;5.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003e17.7\u0026plusmn;6.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.068493150684931%\"\u003e\n \u003cp\u003e0.30\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"25.875190258751903%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003eV30 (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.61339421613394%\" valign=\"top\"\u003e\n \u003cp\u003e13.2\u0026plusmn;4.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003e13.3\u0026plusmn;4.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.068493150684931%\"\u003e\n \u003cp\u003e0.58\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"25.875190258751903%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003eDmean (Gy)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.61339421613394%\" valign=\"top\"\u003e\n \u003cp\u003e54.3\u0026plusmn;5.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003e54.8\u0026plusmn;5.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.068493150684931%\"\u003e\n \u003cp\u003e0.06\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"25.875190258751903%\"\u003e\n \u003cp\u003eContralateral lung\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\"\u003e\n \u003cp\u003eV5 (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.61339421613394%\" valign=\"top\"\u003e\n \u003cp\u003e13.3\u0026plusmn;12.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003e14.1\u0026plusmn;11.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.068493150684931%\"\u003e\n \u003cp\u003e0.15\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"25.875190258751903%\"\u003e\n \u003cp\u003eHeart\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003eV30 (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.61339421613394%\" valign=\"top\"\u003e\n \u003cp\u003e16.0\u0026plusmn;11.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003e14.6\u0026plusmn;12.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.068493150684931%\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.03\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"25.875190258751903%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003eV40 (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.61339421613394%\" valign=\"top\"\u003e\n \u003cp\u003e9.5\u0026plusmn;7.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003e8.8\u0026plusmn;8.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.068493150684931%\"\u003e\n \u003cp\u003e0.19\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"25.875190258751903%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003eDmean (Gy)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.61339421613394%\" valign=\"top\"\u003e\n \u003cp\u003e12.4\u0026plusmn;8.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003e11.9\u0026plusmn;7.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.068493150684931%\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.03\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"25.875190258751903%\"\u003e\n \u003cp\u003eCord\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003eDmax (Gy)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.61339421613394%\" valign=\"top\"\u003e\n \u003cp\u003e33.0\u0026plusmn;6.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003e32.3\u0026plusmn;5.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.068493150684931%\"\u003e\n \u003cp\u003e0.29\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"25.875190258751903%\"\u003e\n \u003cp\u003eCord PRV\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003eDmax (Gy)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.61339421613394%\" valign=\"top\"\u003e\n \u003cp\u003e36.7\u0026plusmn;6.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003e37.4\u0026plusmn;5.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.068493150684931%\"\u003e\n \u003cp\u003e0.28\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"25.875190258751903%\"\u003e\n \u003cp\u003eTrachea\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003eDmax (Gy)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.61339421613394%\" valign=\"top\"\u003e\n \u003cp\u003e55.5\u0026plusmn;12.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003e55.2\u0026plusmn;14.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.068493150684931%\"\u003e\n \u003cp\u003e0.74\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"25.875190258751903%\"\u003e\n \u003cp\u003eEsophagus\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003eDmax (Gy)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.61339421613394%\" valign=\"top\"\u003e\n \u003cp\u003e52.7\u0026plusmn;11.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003e53.6\u0026plusmn;12.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.068493150684931%\"\u003e\n \u003cp\u003e0.05\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"25.875190258751903%\"\u003e\n \u003cp\u003eLiver\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003eDmean (Gy)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.61339421613394%\" valign=\"top\"\u003e\n \u003cp\u003e0.7\u0026plusmn;1.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.72146118721461%\" valign=\"top\"\u003e\n \u003cp\u003e0.8\u0026plusmn;1.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.068493150684931%\"\u003e\n \u003cp\u003e0.59\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e3.4 Planning qualities and time\u003c/p\u003e\n\u003cp\u003eThe overall plan qualities of two kinds of plans were similar, whereas automatic plans took much shorter time (P＜0.001) and had smaller MUs (P＜0.001) than manual plans. Table 4 presents the comparison results of PQM total value, planning time and MUs. The average PQM score was 40.7\u0026plusmn;13.1 for manual plans and 40.8\u0026plusmn;13.5 for automatic plans. Average planning time of automatic plans was (32.6 \u0026plusmn; 5.3) min, which was a large improvement if compared with (103.1\u0026plusmn;18.5) min of manual plans. Meanwhile, MUs in automatic plans were on average reduced by 97.4 \u0026plusmn; 23.6.\u003c/p\u003e\n\u003cp\u003eTable 4 Comparison of PQM total values, planning time and monitor units between manual and automatic plans.\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"657\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd width=\"26.98170731707317%\"\u003e\n \u003cp\u003eMetrics\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.646341463414632%\"\u003e\n \u003cp\u003eManual plan\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"30.182926829268293%\"\u003e\n \u003cp\u003eAutomatic plan\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.1890243902439%\"\u003e\n \u003cp\u003eP-value\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"26.98170731707317%\"\u003e\n \u003cp\u003ePQM total value\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.646341463414632%\"\u003e\n \u003cp\u003e40.7\u0026plusmn;13.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"30.182926829268293%\"\u003e\n \u003cp\u003e40.8\u0026plusmn;13.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.1890243902439%\"\u003e\n \u003cp\u003e0.83\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"26.98170731707317%\"\u003e\n \u003cp\u003ePlanning time (min)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.646341463414632%\"\u003e\n \u003cp\u003e103.1\u0026plusmn;18.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"30.182926829268293%\" valign=\"top\"\u003e\n \u003cp\u003e32.6\u0026plusmn;5.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.1890243902439%\"\u003e\n \u003cp\u003e\u003cstrong\u003e＜\u003c/strong\u003e\u003cstrong\u003e0.001\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"26.98170731707317%\"\u003e\n \u003cp\u003eMonitor units\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.646341463414632%\" valign=\"top\"\u003e\n \u003cp\u003e789.9\u0026plusmn;234.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"30.182926829268293%\" valign=\"top\"\u003e\n \u003cp\u003e692.5\u0026plusmn;210.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.1890243902439%\"\u003e\n \u003cp\u003e\u003cstrong\u003e＜\u003c/strong\u003e\u003cstrong\u003e0.001\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e3.4 Expert review and feedback\u003c/p\u003e\n\u003cp\u003eBased on a tally of 200 reviews from four experts, automatic plans were deemed to be clinically acceptable for treatment in 88% of reviews (176/200). All preference reviews are shown in the Figure 3. In terms of overall review, automatic plans were deemed superior in 36 reviews (18%), equivalent in 98 reviews (49%), and inferior in 66 reviews (33%); In terms of target coverage, 94% of automatic and manual plans were equivalent. In terms of target conformity and homogeneity, experts\u0026rsquo; favours for automatic and manual plans varied little, with more reviews deeming that they were equivalent. Across all reviews, automatic plans typically demonstrated better heart sparing and cord sparing compared with manual plans. But lung sparing of automatic plans were considered inferior to the manual plans.\u003c/p\u003e"},{"header":"4. Discussion","content":"\u003cp\u003eIn the process of designing VMAT plans, planners need to repeatedly adjust objective functions, weights and other plan parameters based on personal experience. As a result, the design process is time-consuming, laborious, and there is a large subjective uncertainty. Ultimately, plan designing is inefficient and varies widely in quality. Lung cancer, as the malignant tumor with the highest incidence, has a variety of target locations, shapes and prescribed doses, which makes the above problems more prominent. As can be seen in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, among the 50 randomly selected cases in this study, four stages were included:Ⅰ, Ⅱ, Ⅲ and Ⅳ; three kinds of treatment patterns were included: postoperative radiotherapy, concurrent chemoradiotherapy, radiotherapy alone. As a result, multiple prescribed dose patterns have been developed. 12 patients had only PTV (seven kinds of prescribed dose gradients) and 38 patients were treated with the simultaneous integrated boost technique (seven kinds of prescribed dose gradients). In this case, the predictive effect of automatic planning was more demanding. Therefore, the deep learning approach was adopted in this study to establish a prediction model of radiotherapy plan for lung cancer. Different from other automatic plan prediction models, this study used anatomical structures and prescribed doses to make predictions, which can be applied to lung cancer plans with different anatomical structures and prescribed doses.\u003c/p\u003e \u003cp\u003eFinally, 50 lung cancer patients with different tumor stages and different treatment patterns were randomly selected for robust evaluation from different perspectives in this study. And the dosimetric parameters and plan efficiency parameters of VMAT automatic plans and manual plans were compared in Pinnacle\u003csup\u003e3\u003c/sup\u003e TPS, which aimed to evaluate the feasibility and superiority of automatic plans in lung cancer radiotherapy. From the analysis results of targets volume, there was no significant difference in the homogeneity of two kinds of plans. In terms of targets coverage, automatic plans were even superior to manual plans. According to the analysis results of OARs, automatic plans had no advantage in lung protection, but there was a significant reduction in V30 and Dmean of heart, possibly because that the lung received the most attention in clinical plans. The lower heart dose would likely benefit patients in a long term [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. To conduct a comprehensive assessment of two kinds of plans, it is necessary to compare the overall plan quality according to a quantitative evaluation criterion. According to the results of PQM analysis with targets and multiple important OARs, the qualities of automatic plans were better than that of manual plans, although there was no significant difference.\u003c/p\u003e \u003cp\u003eAbove all, the time saved for planners will be appreciable [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. If each lung cancer plan saved 70.5\u0026thinsp;\u0026plusmn;\u0026thinsp;13.2 minutes, then 50 lung cancer plans would be able to save 2 to 3 working days. This is a substantial departmental burden that would greatly benefit from automation. Moreover, the MUs of automatic plans were much less than that of manual plans. The fewer MUs can improve the efficiency of plan implementation and reduce the treatment time, which could benefit the patients to reduce discomfort and variation during the treatment process [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eIn the qualitative evaluation, only 12% of automatic plans were not accepted. There were three main reasons for unacceptability. Firstly, for case that target was closer to spinal cord, the point dose of spinal cord might exceed 40 Gy, which was clinically unacceptable. In this case, manual plans would be further optimized and spinal cord dose would be reduced, whereas there was no this step in automatic plans. Secondly, for a few automatic plans, the dose hotspot (110% of the prescribed dose) fell on the esophagus. Thirdly, for complex plans with longer and larger target, the lung dose of automatic plans exceeded clinical constraints, such as V20\u0026thinsp;\u0026gt;\u0026thinsp;30% in lung. Fourthly, for case that target extended to liver, the prediction model of automatic plan did not take the liver into account, resulting in a higher dose of liver than that in the manual plan. It took a planner less than 10 min hands-on time to manually fine-tune and optimize these plans. Eventually they were all accepted. In the preference review towards automatic and manual plans in terms of different metrics, the reviews that considered the two to be equivalent were the most. In 33% of reviews the automatic plans were regarded as not superior to manual plans. Because specialized planner required consultation with physician during the process of designing clinical plans to establish optimization objectives or tradeoffs. For plans with irregular target shape and large target volume, clinical tradeoffs may have been made (for instance, compromising target conformity or homogeneity in favor of sparing a surrounding organ, or compromising heart dose in favor of sparing lung tissue). As a result, manual plans would be modified in detail according to the physicians' preference, and they went through multiple stages such as independent physical review and clinical review. Finally, the plan for clinical treatment was determined. That\u0026rsquo;s the reason why some manual plans were superior to automatic plans in qualitative evaluation.\u003c/p\u003e \u003cp\u003eIn summary, automatic plans were characterized by lower modulation complexity and less delivery time when compared to manual plans, without losing the advantage in targets coverage, homogeneity and OARs dose. But there were some limitations in the present study. Firstly, the prediction of lung dose was not very beneficial, resulting in slightly higher lung dose in automatic plans than in manual plans. In addition, after quantitative evaluation, it was found that different experts have different preferences, (e.g., some experts pay more attention to the lung sparing, some to the heart sparing, and some to the target conformity and homogeneity). The current prediction model cannot meet the preferences of all experts, so different automatic planning prediction models can be developed according to the preferences of different experts. Finally, the prediction of field angle was lacking currently in this study. In the future, the method will be further improved based on the above three points. As a result, for complex and difficult plans, planners still cannot entirely rely on automatic plans. And manual fine-tuning is needed on the basis of automatic plans, which is still much faster than generating manual plans from scratch.\u003c/p\u003e"},{"header":"5. Conclusions","content":"\u003cp\u003eThis study describes an automatic VMAT planning method based on deep learning and demonstrates its feasibility for lung cancer. The automatic planning method is suitable for lung cancer plans with different target volumes and different treatment patterns. It can be found by comparing with the manual plans, automatic treatment planning method might be an alternative, which improves the planning and treatment efficiency without compromising the plan quality.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cdiv class=\"DefinitionList\"\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\"\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\"\u003ePQM\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eplan quality metric\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eOAR\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eorgan at risk\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eMUs\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003emonitor units\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\"\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\"\u003eVTPN\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003evirtual treatment planner network\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eMOAPN\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003emulti-objectives adjustment policy network\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eIMRT\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eintensity modulated radiation therapy\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\u003ePlanning Target Volume\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eAJCC\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eAmerican Joint Committee on Cancer\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eCTV\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eClinical Target Volume\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eGTV\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eGross Target Volume\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eDSS\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003edose shaping structure\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\"\u003eHI\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003ehomogeneity index\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgments\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work was supported by the Special Research Fund for Central Universities, Peking Union Medical College, CAMS Innovation Fund for Medical Sciences (CIFMS) [2022-I2M-C\u0026amp;T-B-075]; the Shanghai Pujiang Programme [23PJD014]; the Beijing Hope Run Special Fund of Cancer Foundation of China (LC2021B01); National Natural Science Foundation of China (11875320).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor information\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAuthors and Affiliations\u003c/p\u003e\n\u003cp\u003eNational Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing 100021, China\u003c/p\u003e\n\u003cp\u003eNingyu Wang, Yingjie Xu, Lingling Yan, Deqi Chen, Wenqing Wang, Kuo Men, Jianrong Dai, Zhiqiang Liu\u003c/p\u003e\n\u003cp\u003eDepartment of Radiation Oncology, Fudan University Shanghai Cancer Center, Shanghai 200032, China\u003c/p\u003e\n\u003cp\u003eJiawei Fan\u003c/p\u003e\n\u003cp\u003eDepartment of Oncology, Shanghai Medical College Fudan University, Shanghai 200032, China\u003c/p\u003e\n\u003cp\u003eJiawei Fan\u003c/p\u003e\n\u003cp\u003eShanghai Key Laboratory of Radiation Oncology, Shanghai 200032, China\u003c/p\u003e\n\u003cp\u003eJiawei Fan\u003c/p\u003e\n\u003cp\u003eContributions\u003c/p\u003e\n\u003cp\u003eNingyu Wang: Conceptualization, Methodology, Formal analysis, Investigation, Data Curation, Writing - Original Draft, Visualization, Visualization\u003c/p\u003e\n\u003cp\u003eJiawei Fan: Methodology, Software,\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eYingjie Xu: Validation, Formal analysis\u003c/p\u003e\n\u003cp\u003eLingling Yan: Validation, Formal analysis\u003c/p\u003e\n\u003cp\u003eDeqi Chen: Formal analysis, Data Curation\u003c/p\u003e\n\u003cp\u003eWenqing Wang: Validation, Formal analysis\u003c/p\u003e\n\u003cp\u003eKuo Men: Writing - Review \u0026amp; Editing, Supervision\u003c/p\u003e\n\u003cp\u003eJianrong Dai: Conceptualization, Methodology, Writing - Review \u0026amp; Editing, Visualization, Supervision, Funding acquisition\u003c/p\u003e\n\u003cp\u003eZhiqiang Liu: Conceptualization, Methodology, Software, Investigation, Writing - Review \u0026amp; Editing, Visualization, Funding acquisition, Project administration\u003c/p\u003e\n\u003cp\u003eCorresponding author\u003c/p\u003e\n\u003cp\u003eCorrespondence to Zhiqiang Liu and Jianrong Dai.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics declarations\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eEthics approval and consent to participate\u003c/p\u003e\n\u003cp\u003eThis study was performed in line with the principles of the Declaration of Helsinki. Approval was granted by the Independent Ethics Committee of National Cancer Center/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College. All patients were informed and signed relevant informed consent documents.\u003c/p\u003e\n\u003cp\u003eConsent for publication\u003c/p\u003e\n\u003cp\u003eThe authors consent to publish the manuscript in its current form.\u003c/p\u003e\n\u003cp\u003eCompeting interests\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eBray F, Ferlay J, Soerjomataram I, Siegel RL, Torre LA, Jemal A. vol 68, pg 394,. Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries (2018). Ca-a Cancer Journal for Clinicians. 2020; 70:313-.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang D, Yuan Z, Hu P, Yang Y. Automatic treatment planning for cervical cancer radiation therapy using direct three-dimensional patient anatomy match. J Appl Clin Med Phys. 2022; 23.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDella Gala G, Dirkx MLP, Hoekstra N, Fransen D, Lanconelli N, van de Pol M, et al. Fully automated VMAT treatment planning for advanced-stage NSCLC patients. Strahlenther Onkol. 2017;193:402\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSwanick CW, Lin SH, Sutton J, Naik NS, Allen PK, Levy LB, et al. Use of Simultaneous Radiation Boost Achieves High Control Rates in Patients With Non-Small-Cell Lung Cancer Who Are Not Candidates for Surgery or Conventional Chemoradiation. Clin Lung Cancer. 2015;16:156\u0026ndash;63.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHussein M, Heijmen BJM, Verellen D, Nisbet A. Automation in intensity modulated radiotherapy treatment planning-a review of recent innovations. Br J Radiol. 2018; 91.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBuergy D, Sharfo AWM, Heijmen BJM, Voet PWJ, Breedveld S, Wenz F et al. Fully automated treatment planning of spinal metastases - A comparison to manual planning of Volumetric Modulated Arc Therapy for conventionally fractionated irradiation. Radiat Oncol. 2017; 12.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXhaferllari I, Wong E, Bzdusek K, Lock M, Chen JZ. Automated IMRT planning with regional optimization using planning scripts. J Appl Clin Med Phys. 2013;14:176\u0026ndash;91.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFogliata A, Cozzi L, Reggiori G, Stravato A, Lobefalo F, Franzese C et al. RapidPlan knowledge based planning: iterative learning process and model ability to steer planning strategies. Radiat Oncol. 2019; 14.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang S, Zheng DD, Zhang C, Ma RT, Bennion NR, Lei Y, et al. Automatic planning on hippocampal avoidance whole-brain radiotherapy. Med Dosim. 2017;42:63\u0026ndash;8.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKamran SC, Mueller BS, Paetzold P, Dunlap J, Niemierko A, Bortfeld T, et al. Multi-criteria optimization achieves superior normal tissue sparing in a planning study of intensity-modulated radiation therapy for RTOG 1308-eligible non-small cell lung cancer patients. Radiother Oncol. 2016;118:515\u0026ndash;20.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGood D, Lo J, Lee WR, Wu QJ, Yin FF, Das SK. A Knowledge-Based Approach to Improving and Homogenizing Intensity Modulated Radiation Therapy Planning Quality Among Treatment Centers: An Example Application to Prostate Cancer Planning. Int J Radiat Oncol Biol Phys. 2013;87:176\u0026ndash;81.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXu H, Lu J, Wang J, Fan J, Hu W. Implement a knowledge-based automated dose volume histogram prediction module in Pinnacle\u003csup\u003e3\u003c/sup\u003e treatment planning system for plan quality assurance and guidance. J Appl Clin Med Phys. 2019;20:134\u0026ndash;40.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKearney V, Chan JW, Wang T, Perry A, Descovich M, Morin O, et al. DoseGAN: a generative adversarial network for synthetic dose prediction using attention-gated discrimination and generation. Sci Rep. 2020;10:11073.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFan J, Wang J, Chen Z, Hu C, Zhang Z, Hu W. Automatic treatment planning based on three-dimensional dose distribution predicted from deep learning technique. Med Phys. 2019;46:370\u0026ndash;81.\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.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMomin S, Fu Y, Lei Y, Roper J, Bradley JD, Curran WJ, et al. Knowledge-based radiation treatment planning: A data-driven method survey. J Appl Clin Med Phys. 2021;22:16\u0026ndash;44.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShen CY, Nguyen D, Chen LY, Gonzalez Y, McBeth R, Qin N, et al. Operating a treatment planning system using a deep-reinforcement learning-based virtual treatment planner for prostate cancer intensity-modulated radiation therapy treatment planning. Med Phys. 2020;47:2329\u0026ndash;36.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang H, Bai X, Wang Y, Lu Y, Wang B. An integrated solution of deep reinforcement learning for automatic IMRT treatment planning in non-small-cell lung cancer. Front Oncol. 2023; 13.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu Z, Fan J, Li M, Yan H, Hu Z, Huang P, et al. A deep learning method for prediction of three-dimensional dose distribution of helical tomotherapy. Med Phys. 2019;46:1972\u0026ndash;83.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHodapp N, [The ICRU. Report 83: prescribing, recording and reporting photon-beam intensity-modulated radiation therapy (IMRT)]. Strahlenther Onkol. 2012;188:97\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKataria T, Sharma K, Subramani V, Karrthick KP, Bisht SS. Homogeneity Index: An objective tool for assessment of conformal radiation treatments. J Med Phys. 2012;37:207\u0026ndash;13.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNelms BE, Robinson G, Markham J, Velasco K, Boyd S, Narayan S, et al. Variation in external beam treatment plan quality: An inter-institutional study of planners and planning systems. Practical radiation oncology. 2012;2:296\u0026ndash;305.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBohsung J, Gillis S, Arrans R, Bakai A, De Wagter C, Knoos T, et al. IMRT treatment planning - A comparative inter-system and intor-centre planning exercise of the ESTRO QUASIMODO group. Radiother Oncol. 2005;76:354\u0026ndash;61.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXia W, Liu Z, Yan L, Han F, Hu Z, Tian Y, et al. A longitudinal evaluation of improvements in treatment plan quality for lung cancer with volumetric modulated arc therapy. J Appl Clin Med Phys. 2020;21:33\u0026ndash;43.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMcIntosh C, Conroy L, Tjong MC, Craig T, Bayley A, Catton C, et al. Clinical integration of machine learning for curative-intent radiation treatment of patients with prostate cancer. Nat Med. 2021;27:999\u0026ndash;1005.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLing C, Han X, Zhai P, Xu H, Chen J, Wang J et al. A hybrid automated treatment planning solution for esophageal cancer. Radiat Oncol. 2019;14.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMitchell RA, Wai P, Colgan R, Kirby AM, Donovan EM. Improving the efficiency of breast radiotherapy treatment planning using a semi-automated approach. J Appl Clin Med Phys. 2017;18:18\u0026ndash;24.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang Q, Peng Y, Song X, Yu H, Wang L, Zhang S. Dosimetric evaluation of automatic and manual plans for early nasopharyngeal carcinoma to radiotherapy. Med Dosim. 2020;45:E13\u0026ndash;E20.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Lung cancer, Automatic planning, Deep learning, VMAT, Expert review","lastPublishedDoi":"10.21203/rs.3.rs-3872969/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-3872969/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cb\u003eBackground and purpose\u003c/b\u003e\u003c/p\u003e \u003cp\u003eThe purpose of this study is to investigate the clinical application and assessment of deep learning (DL)-assisted automatic radiotherapy planning for lung cancer patients.\u003c/p\u003e\u003cp\u003e\u003cb\u003eMaterials and methods\u003c/b\u003e\u003c/p\u003e \u003cp\u003eWe developed a deep learning model for predicting patient-specific dose that was trained and validated on a dataset of 235 lung cancer patients, and the model was integrated into clinical workflow to assist planners in generating treatment plans. We retrospectively selected and recovered additional 50 clinically treated lung cancer patients\u0026rsquo; manual volumetric modulated arc therapy (VMAT) plans with different target volumes and different treatment patterns. Subsequently, automatic plans were generated for each of these patients. Both automatic and manual plans were subsequently compared in terms of overall plan quality metric (PQM), target coverage and homogeneity, organ at risk (OAR) sparing, monitor units (MUs), and planning time. Additionally, qualitative reviews of automatic and manual plans were implemented by four expert reviewers to assess the clinical applicability of DL-assisted automatic plans.\u003c/p\u003e\u003cp\u003e\u003cb\u003eResults\u003c/b\u003e\u003c/p\u003e \u003cp\u003eThe average PQM score was 40.7\u0026thinsp;\u0026plusmn;\u0026thinsp;13.1 for manual plans and 40.8\u0026thinsp;\u0026plusmn;\u0026thinsp;13.5 for automatic plans, and they had equivalent overall plan quality. The targets coverage and homogeneity of the automatic plans were considered equivalent or superior when compared to manual plans. Both plans had their own advantages in OAR sparing, such as better sparing of lung for manual plans and better sparing of heart for automatic plans. It is worth to note that the average planning time of automatic plans was reduced from 103.1\u0026thinsp;\u0026plusmn;\u0026thinsp;18.5 minutes to 32.6\u0026thinsp;\u0026plusmn;\u0026thinsp;5.3 minutes (P\u0026lt;0.001) and the MUs were reduced from 789.9\u0026thinsp;\u0026plusmn;\u0026thinsp;234.3 to 692.5\u0026thinsp;\u0026plusmn;\u0026thinsp;210.7 (P\u0026lt;0.001). In qualitative evaluation, automatic plans were deemed to be clinically acceptable for treatment in 88% of reviews (176/200), and all were accepted after fine tuning. Most expert reviews indicated a preference for equivalence between automatic and manual plans when making their selection.\u003c/p\u003e\u003cp\u003e\u003cb\u003eConclusion\u003c/b\u003e\u003c/p\u003e \u003cp\u003eThe DL-assisted lung cancer plans demonstrated comparable or superior quality to manual plans, improved planning and treatment efficiency, and significantly reduced planning time and MUs. It has the potential to enhance the workflow of radiotherapy departments, ultimately providing tangible benefit to lung cancer patients.\u003c/p\u003e","manuscriptTitle":"Clinical implementation and evaluation of deep learning-assisted automatic radiotherapy treatment planning for lung cancer","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-01-25 15:46:33","doi":"10.21203/rs.3.rs-3872969/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"9c0804af-98ef-4cc5-88b2-f313cca736c4","owner":[],"postedDate":"January 25th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2024-01-31T09:01:48+00:00","versionOfRecord":[],"versionCreatedAt":"2024-01-25 15:46:33","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-3872969","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-3872969","identity":"rs-3872969","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}