Guided Deep Learning Auto-Correction of Suboptimal Auto-Segmentation of Complex Structures for MRI-Guided Adaptive Radiotherapy

Abstract

Fast and accurate auto-segmentation is desirable for MR-guided adaptive radiotherapy (MRgART), however is challenging particularly for complex structures, e.g., bowels. Anticipating that auto-segmentation, particularly deep learning (DL)-based, will be substantially improved in the near future, the auto-segmentation for these complex structures would become acceptable or suboptimal, implying a need for correcting such suboptimal contours. This study aims to develop a guided DL method to automatically and quickly correct for the suboptimal auto-segmented bowel contours for MRgART of abdominal tumors.Auto-segmentation contours of both small and large bowels created by standard deep convolutional neural network models on 46 T2-MRI sets were used as initial contours to demonstrate the feasibility of the proposed method. As preprocessing, MRI images were standardized (bias corrected, contrast enhanced and normalized) and cropped into multiple 2D subregions using the initial auto-segmented bowel contours, with each including a bowel loop. A total of 1320 suboptimal initial subregion contours with Dice similarity coefficients (DSC) in the range of 0.8∼0.9, as compared to the ground truth (manually delineated contours), were selected to train and test a U-net model for contour auto-correction. The model was trained using 75% of the dataset with two input channels: 1) the pre-processed MRI; 2) initial suboptimal contours used as a guidance as these contours were close to the ground truth. Data augmentation (rotation, flipping and scaling) was performed during the training, with the training objective of minimizing the binary cross entropy loss between the U-net output and ground truth. The obtained U-net auto-correction model was tested using the rest 25% of the dataset (325 subregions from 11 scans) and its performance was measured using DSC and 95% Hausdorff distance (95% HD).The U-net model performance is summarized in the table below. The model effectively corrected for the suboptimal bowel contours with more than 1/3 contours reaching DCS > 0.9 after correction. The execution time to correct for a testing contour was less than 1 second on a GTX 1060 GPU.A U-net model was developed to automatically and quickly correct for suboptimal auto-segmented bowel contours. With further development using large datasets, the U-net model could be integrated in conjunction with an auto-segmentation pipeline to improve the accuracy of suboptimal contours for complex anatomy, minimizing the labor-intensive and time-consuming manual editing and accelerating the process of MRgART.