Organ Specific Deep Learning-Based Correction of Inaccurate Auto-Segmentation on Abdominal MRI

Abstract

Deep learning-based automatic segmentation (DLAS) techniques offer limited success for abdominal organs on MRI, requiring substantial editing time. We have previously developed a deep learning based automatic contour correction (ACC) technique that can correct for inaccurate DLAS contours of bowels on MRI, reducing the manual editing time in MR-guided online adaptive radiation therapy (MRgART). This study aims to develop deep learning-based ACC models for pancreas and duodenum that are particularly difficult to contour either manually or with DLAS. Dense UNet, a deep learning algorithm that combines UNet with dense blocks, was trained to create ACC models. Organ-specific models were trained for pancreas and duodenum contours obtained from a research DLAS tool on MRIs from a 1.5T MR-Linac. The training dataset contained MRI slices paired with DLAS contours from 54 abdominal MRL sets along with ground truth contours and 540 additional augmented sets created by shifting, rotating, and scaling each organ along with the contours and varying the noise and bias field for each patient set. Each DLAS contour was classified into the acceptable (no additional edits required), minor edit (only simple edits required), or major edit category based on the expected editing effort determined using a contour classification model developed in a separate study. The ACC models were trained for the slices requiring minor edit and major edit separately. Performance of the obtained models were tested using an independent 11 MRI sets in term of the change of contour category based on the contour classification model. After applying the duodenum ACC model to the testing datasets, 16% (27/165) and 5% (8/178) of the minor and major edits' slices, respectively, improved to acceptable and 31% (54/178) of the major edit slices improved to minor edits. Furthermore, the total percentage of acceptable contours grew from 10% (36/378) to 19% (71/378) and the percentage of the major edit slices reduced from 47% (178/378) to 30% (115/378). After applying the pancreas ACC model to the testing datasets, 32% (47/143) and 1% (1/96) of the minor and major edits' slices, respectively, improved to acceptable and 49% (47/96) of the major edit slices improved to minor edit slices. Furthermore, the total percentage of acceptable contours grew from 14% (38/277) to 31% (86/277) and the percentage of major edit slices reduced from 35% (96/277) to 17% (48/277). Deep learning based automatic contour corrections can substantially improve inaccurate DLAS contours of pancreas and duodenum on MRI that would otherwise require time-consuming edits, resulting in less manual intervention and increased efficiency during MRgART.