Evaluation Of A Deep Learning-Based Auto-Segmentation Method For Quality Assurance Of Both Male And Female Pelvic Organ-At-Risk Contours In NCTN Clinical Trials

Abstract

We previously reported a cascaded atrous convolution and spatial pyramid pooling module-based auto-segmentation method for organ-at-risk (OAR) segmentation and its quality assurance (QA) for radiotherapy clinical trial cases submitted from various disease sites. In this study, we enhanced the technique and evaluated the enhanced method for QA of OAR contouring using data submitted from both a genitourinary (GU) trial (NRG-RTOG 0938) and a gynecology (GYN) trial (NRG GY006). Three convolutional neural network (CNN)-based models were trained for rectum, bladder, left femur, and right femur: One GU model used data of 192 patients from RTOG 0938, one GYN model used data of 111 patients from GY006 and the third model used mixed data from both trials. Another 20 patient data sets from each trial, a total of 40 patient data sets, were used for model validation. Using these models, two sets of OAR contours were automatically generated on each test patient: One set used the GU and GYN models on 20 male and 20 female patients, respectively and another set used the mixed model for all 40 test patients. The performance of the models was evaluated using dice indices calculated using MiM software. For GU patients, contours generated using GU model vs. mixed model, the Dice Similarity Coefficient values were 0.94±0.03 vs. 0.94±0.03 for the bladder, 0.80±0.05 vs. 0.81±0.06 for the rectum, 0.87±0.04 vs. 0.86±0.05 for the left femur, and 0.88±0.03 vs. 0.87±0.05 for the right femur, respectively. For GYN patients’ contours generated using GYN model vs. mixed model, corresponding values were 0.80±0.15 vs. 0.85±0.18 for the bladder, 0.70±0.11 vs. 0.65±0.11 for the rectum, 0.45±0.21 vs. 0.45±0.18 for the left femur, and 0.46±0.20 vs. 0.44±0.19 for the right femur, respectively. The GYN data had fewer cases and relatively low image quality due to image artifacts caused by the inserted applicator. There was no guideline provided for GY006 femur contours in the protocol, which resulted in large variations in submitted contours. These factors have affected the auto-segmentation accuracy of the GYN models and test results using the GYN data. The CNN model-based method generates OAR segmentation with reasonable accuracy and has the potential to be used for contour QA for future patients enrolled on both GU and GYN trials. The models built using mixed male and female data also demonstrate acceptable accuracy for pelvic OARs for both male and female patients.