Author: Sara Endo, Takeshi Fujisawa, Hidehiro Hojo, Masaki Nakamura, Hidenobu Tachibana 👨🔬
Affiliation: Department of Radiation Oncology, National Cancer Center Hospital East, Radiation Safety and Quality Assurance Division, National Cancer Center Hospital East 🌍
Purpose: To assess the clinical feasibility of deep learning (DL)-based automated contouring through qualitative and quantitative assessments.
Methods: Sixty cases were chosen, including 3 OARs in brain metastasis, 4 OARs in lung, and 2 Oars in cervix cases. The clinically approved treatment plans for them were extracted. SYNAPSE Radiotherapy (FUJIFILM Corp., Japan) implements a deep-learning approach of auto-contouring for organs-at-risk (OARs). As the qualitative assessments, our radiation oncologists visually assessed the agreements between DL-based auto contours and physician-drawn contours on a four-point scale. As the quantitative assessment, dose-volume constraints using the auto contours and physician-drawn contours were compared to evaluate whether the autocontour OAR dose was equivalent to that of the manual contours for the plans.
Results: In the qualitative evaluation by physicians, 85.5%, 55.0%, and 66.0% were clinically acceptable without modifications for the brain, lung, and cervix cases, respectively. 98.5%, 66.3%, and 77% were found to be clinically acceptable when including those requiring minimal manual adjustments for those cases, respectively. For stereotactic radiosurgery plans in brain metastasis cases, 99.1% and 99.1% of structures met dosimetric criteria for automatic and manual contours, respectively. For lung plans, 91.8% and 95.7% of structures met dosimetric criteria for automatic and manual contours, respectively; For cervix plans, 98.0% and 98.0% of structures met dosimetric criteria for automatic and manual contours, respectively. With a 5% margin, autocontours were equivalent to the manual contours for 98.1%, 84.7%, and 93.4% of dose-volume constraints for the SRS, lung, and cervix plan, respectively. There was a comparatively low agreement of 89.5 %, 87.5%, and 31.3% for the brainstem, esophagus, and heart, respectively.
Conclusion: The DL-based automated contouring system is clinically effective with minimal manual modifications. This study also shows that dosimetrically equivalent OAR contours can be created in most of the stereotactic radiosurgery, lung, and cervix plans.