AI Model of Using Stratified Deep Learning to Delineate the Organs at Risk (OARs) for Thoracic Radiation Therapy

Abstract

<h3>Purpose/Objective(s)</h3> Guided by the atlas of RTOG1106, the aim of this study was to develop an Artificial Intelligent (AI) model to improve the consistency and efficiency for Organs at Risk (OARs) delineation. <h3>Materials/Methods</h3> A total of 101 lung patients with contrast-enhanced CT images were used for this study. Following the atlas of the RTOG1106 as the gold standard, OARs were first delineated manually by experienced oncologists. NnUNet was adopted as the basic deep learning network. A 4-fold cross validation was applied to evaluate the quantitative measures of the segmentation accuracy, with respect to the gold OAR reference of RTOG1106 atlas. Accuracy was assessed by Dice Score (DSC) and a user-study of 20 patients by three radiation oncologists, with score rendered by the editing time needed to meet the criteria of the RTOG1106 standards: major deviation (>10 mins), minor (<10 mins) and no revision. As control, consistency of the manual delineation was recorded from OARs delineated by 21 trained oncologists on a RTOG1106 dry case. <h3>Results</h3> An automated AI model was generated for 5 key thoracic OARs. The quantitative accuracy is summarized in Table 1. This model had a mean DSC of 93.5% (range 85.7-97.9%) on all OARs except the brachial plexus of 48.5%, which is anatomically complex. The average delineation time of the AI model was 2 min (range 1.6-2.5) for each patient. Of the user-study, 15 (75%) and 4 (20%) AI contours needed minor and no revision, respectively. In comparison, 12 (60%) and 2 (10%) manual contours required minor and no revision, respectively. Notably, only 1 (5%) AI contours needed major revision, while it was 6 (30%) for manual contours. Of the 21 radiation oncologists, the mean manual delineation time were 50 min (range 20 min-240 min), and 13 of them (62%) spent greater than 30 mins (max up to 4 hrs) in this patient. <h3>Conclusion</h3> We developed a stratified deep-learning AI segmentation for RTOG1106 OAR atlas. This model appeared to be more accurate than manual delineation, except for brachialplex which requires further model refinement. As the RTOG1106 atlas is widely used in the clinic, this AI model may be used to guide consistent contouring in clinical trial and time-saving in clinical practice.