Patient-Specific Deep Learning Auto-Segmentation for MR-Guided Adaptive Radiotherapy of Prostate Cancer

Abstract

MR guided conventional prostate radiotherapy requires agreement of bladder and rectum position, size, and shape with reference plan anatomy. In many instances, treatments may be delayed to allow optimal filling of these organs at risk (OARs). To account for such changes, we propose a patient specific deep learning based auto-segmentation (PS-DLAS) solution to implement pseudo (adapt-to-shape) ATS planning as a replacement for (adapt-to-position) ATP for conventionally fractionated treatments, and to generate auto-segmented contours for ATS SBRT treatments. A commercially available DLAS tool that allows to customize/re-train DLAS models to an individual patient was implemented and automated in 3 iterative steps: 1) training of initial PS-DLAS model based on the first daily MRI with manual contours for the patient, 2) using the trained model to auto-segment the next daily MRI (acquired at next fraction), and 3) re-training/updating the model with newly available daily MRI sets and the verified contours. Steps 1) and 3) were performed offline while the step 2 was online. The solution was tested using daily MRI sets collected for 6 prostate cancer patients treated with MR guided adaptive radiotherapy (MRgART) either in 5-fraction SBRT or conventionally fractionated with adapt-to-shape (ATS) or adapt-to-position (ATP) workflows on a 1.5T MR-Linac. The quality of the auto-segmented OAR contours, including bladder, penile bulb, prostate, rectum, and seminal vesicles, obtained in step 2) were assessed, where the acceptable contour slices (no editing required) were identified. Additionally, ATP was simulated using a pseudo ATS workflow on patient 6, allowing the PS-DLAS to be used on daily MRI sets and the obtained contours to be used in the daily plan optimization. The time saving and the plan quality of using the PS-DLAS were compared to those from the current standard clinical workflow. The times for the offline model training in steps 1) or 3) were < 50 minutes and the times of applying the updated PS-DLAS to auto-segment a daily MRI set in step 2) were within 1 minute using a hardware of Intel(R) Xeon(R) Gold 5222 CPU. Among the auto-segmented contour slices, 87%, 92%, 88%, 86% and 84% were found to be acceptable for bladder, penile bulb, prostate, rectum, and seminal vesicles, respectively. The time savings of using the PS-DLAS were about 7 minutes, compared to the manual contouring in the current ATS workflow. Manual editing was almost always required in 1-3 most superior slices for each organ. For the patient tested with the pseudo ATS, time saving was 10 minutes with comparable plan quality to the standard ATP workflow. In this study, we demonstrated use of a PS-DLAS approach to reduce contouring time in MR-guided ATS prostate radiotherapy. In addition, these results suggest a PS-DLAS approach combined with pseudo-ATS may reduce overall treatment times in conventionally fractionated MRgRT by eliminating time required for matching of OAR filling.