T. Zhang; Department of Radiation Oncology, Xijing Hospital, Fourth Military Medical University, Xian, Shaanxi, China
Purpose/Objective(s): Treatment planning is conventionally performed manually by professionally trained medical dosimetrists and physicists through time-and effort-consuming. With the rapid development of deep learning methods, deep Learning-based predicted dose distribution can be taken as a quality control tool for clinical treatment plan, by which the planners can know whether or where the dose distributions can be improved, and the physicians can immediately view 3D dose distributions to adjust OARs dose constraint requirements. The purpose of this study was to propose a clinical feasible optimization strategy for midpiece esophageal carcinoma volumetric modulated arc therapy (VMAT) plan, including dose prediction via a deep learning (DL) and optimization strategy based on dose prediction. Materials/
Methods: The adopted network for dose prediction is a modified 3D U-Net named as 3D YZY U-NET. A database of 75 volumetric modulated arc therapy (VMAT) plans which were all designed by two senior and experienced medical physicals for midpiece esophageal cancer patients, was utilized to produce training and validation datasets. 10 patients were used for testing. All of the plans used two completed-degree arcs with avoidance sectors(60°-120°,300°-240°) to lower the dose to the lungs. Dose prediction was performed using a contoured image of the planning target volume (PTV), planning gross target volume (PGTV), body, heart left, Lung, right lung and spinal cord. The Mean Absolute Error (MAE) of 3D dose distribution was used for quantitative evaluation of the dose prediction. 10 testing plans designed by a DL-assisted method were compared with clinically approved plans to evaluate clinical gain, according to dosimetric indices and dose volume histograms (DVH). Results: Similarities were found between the DVH of DL predicted models and clinical approved plan. The visual inspection indicates that the predicted and clinical DVHs of PGTV, PTV and OARs have an acceptable agreement for each patient. The MAE between the DL predicted models and clinical plans were 0.047, 0.041, 0.031, 0.056, 0.036, 0.039, 0.052(PGTV D95 normalized to 1) in the PTV, PGTV and OARs of the body, Heart, Left lung, right lung and spinal cord respectively. However, DL-assisted plans achieved clinically required dose coverage to the target volumes and achieved equal or better OAR sparing when compared to the clinical approved plans. Most notably, no significant differences (p < 0.05) were found in the conformity indices (CIs) between clinical plans and DL-based plans. All the DL-based plans could be delivered on commercial TPS in 15 minutes. Conclusion: A deep-learning method for dose prediction was developed and was demonstrated accurately in patient-specific dose for midpiece esophageal cancer. An optimization strategy based on dose prediction shows great benefits of clinical feasibility, result consistency and high efficiency in generating clinically acceptable plans for midpiece esophageal cancer.
