Automatic Contour Refinement of Inaccurate Auto-Segmentation Using an Active Contour Model for MR-Guided Adaptive Radiotherapy

Abstract

<h3>Purpose/Objective(s)</h3> One of the bottleneck problems of MR-guided adaptive radiotherapy (MRgART) is the impractically long time required to segment the patient's anatomy of the day. Deep learning auto-segmentation (DLAS) has a limited success for complex structures such as those in abdomen. We have previously proposed an active contour model (ACM) to automatically refine the inaccurate contours from DLAS. This study aims to improve the performance of the ACM method by further optimizing the method and testing it using large and diversified abdominal MRI datasets. <h3>Materials/Methods</h3> An improved ACM method that does not require any manual parameter adjustment was implemented. The ACM utilizes the probability maps generated from DLAS models to establish 2D parameter maps and to initialize contour evolution. The performance of the ACM method was tested on the inaccurate DLAS contours of abdominal organs from 91 MRI sets (76 sets of T2-HASTE sequence from a 3T MR-simulator and 15 sets of BTFE sequence from a 1.5T MR-Linac). Bowels were divided into multiple subregions with each including a bowel loop. Levels of contour inaccuracies were divided into two groups: (1) major error group with Dice similarity coefficient (DSC) <0.5 or mean distance to agreement (MDA) >8 mm, and (2) minor error group with remaining contours of 0.5≤ DSC<0.8 or 3 mm<MDA≤8 mm. The performance of the ACM was evaluated based on various metrics, e.g., DSC, MDA, and the added path length (APL, a metric that is correlated with contour editing time). Manually delineated contours were used as ground truth. <h3>Results</h3> As shown in the table below, the average DSC, MDA and APL values for all the contours were improved after the ACM correction. By applying this ACM, approximately half of the contours in major error group were improved to minor error group, and 12-40% of the contours became practically acceptable per TG-132 recommendation (DSC>0.8 and MDA<3 mm). The largest improvement was seen in the major error group of pancreases, where APL reduced by one third, indicating substantial reductions in the subsequent contour editing effort. The execution time of this method to correct one contour was around 1 second. <h3>Conclusion</h3> The ACM method was successfully implemented to automictically and quickly correct for inaccurate contours from MRI-based DL auto-segmentation for complex anatomy, e.g., bowels and pancreas, in abdomen. The ACM method may be used to reduce the manual contour editing workload and accelerate the segmentation process for MRgART.