Auto-Correction for Inaccurate Auto-Segmentation of Abdominal MRI by Combining Deep Learning and Active Contour Method

Abstract

Fast and accurate auto-segmentation is crucial during the magnetic resonance-guided adaptive radiation therapy (MRgART). Deep learning auto-segmentation (DLAS) approaches do not always result in clinically acceptable contours, especially for complex abdominal organs. We have previously reported that the inaccurate DLAS for bowels can be refined using deep learning (DL) and/or active contour method (ACM). This study aims to develop an automatic contour correction (ACC) tool by combining DL and ACM techniques to correct for inaccurate DLAS of pancreas and stomach on MRI. The ACC technique consisted of ACM and DL based on UNet system. The ACM utilizes the probability maps generated from DLAS models to establish 2D parameter maps and to initialize contour evolution, thus not requiring initial parameter adjustments. The Organ specific DL-UNet models were trained for pancreas and stomach contours obtained from a research DLAS tool on abdominal MRIs acquired during routine MRgART from a 1.5T MR-Linac using either turbo field-echo or balanced turbo field-echo sequences. The training dataset contained MR slices along with DLAS and ground truth contours from 54 abdominal MRL sets, and 540 additional augmented sets created by shifting and rotating. DLAS contours were classified based on the expected editing effort into the acceptable, minor edit, or major edit category using an in-house developed classification model. Performance of the obtained ACC models were tested on an independent dataset of 11 sets of abdominal MRIs. For pancreas, the DL-UNet model improved 17% (26/153) and 2% (2/95) of the minor and major edits' slices of the testing dataset, respectively, to acceptable and 39% (37/95) of the major edit slices improved to minor edits. The ACM model improved 3% (4/153) of the minor edit slices to acceptable and the 36% (34/95) of the major slices to minor edits. Using the ACC technique with DL and ACM combined, the percentage of acceptable contours increased from 10% (29/277) to 24% (66/277), and minor edits from 55% (153/277) to 61% (170/277), while the percentage of the major edit slices reduced from 35% (95/277) to 15% (41/277). For stomach, the DL model improved 8% (29/366) of the minor edit slices to acceptable and 50% (16/32) of the major edit slices to minor edit slices. The ACM resulted in 2% (6/366) of minor edit slices to acceptable and 41% (13/32) of major edit slices to minor category. Combining both the DL and ACM, the overall percentage of acceptable stomach contours grew from 13% (58/456) to 22% (101/456) and the percentage of major edit slices reduced from 7% (32/456) to 2% (11/456). The ACC method combining both DL and ACM models can substantially improve the quality of inaccurate DLAS contours of pancreas and stomach in a fully automated and fast manner, minimizing the subsequent manual editing time required for MRgART.