Deep Learning-based Auto-segmentation on CT and MRI for Abdominal Structures
A. Amjad,J. Xu,D. Thill,N. O'Connell,L. Buchanan,I. K. Jones,W. A. Hall,B. A. Erickson,A. Li
DOI: https://doi.org/10.1016/j.ijrobp.2020.07.2276
2020-01-01
Abstract:Rapid and consistent delineation of abdominal structures on CT and MRI is challenging in MR-guided adaptive radiation therapy (MRgART) for abdominal tumors. The purpose of this work is to assess the efficacy and performance of auto-segmentation on abdominal CT and MR images using a deep convolutional neural network (DCNN). The auto-segmentation package developed was based on a 3D DCNN with an architecture of Fully CN network, with end-to-end mapping. A robust model training was achieved by integrating, long-range U-Net skip, short-range residue connections and respective segmentation maps. Paired images of mid-position CTs and MRIs acquired using T2-weighted HASTE and T1-weighted DIXON protocols on a 3T MR system for 25 patients with abdominal tumors were used for model training. For each case, contours of 11 structures, i.e., A_Aorta, Bowel_Large, Bowel_Small, Duodenum, Kidney_L, Kidney_R, Liver, Pancreas, SpinalCord, Spleen and Stomach, were delineated manually following RTOG guidelines and independently checked to minimize segmentation inconsistencies. The performance of the auto-segmentation was assessed with 5 testing data sets in terms of execution time and clinical acceptability using Mean Distance to Agreement (MDA) and Dice Similarity Coefficient (DSC). Metrics tabulated below for the auto-segmented contours of 11 structures for 5 testing sets, revealed that the segmentation for all organs on both MRI series is slightly improved compared to CT, due to the enhanced soft-tissue contrast with MR. On average, across all image pairs, al1 structures excepts bowels showed excellent consistency. In addition, the first seven organs in the table showed clinical acceptability, under the TG 132 tolerances (MDA < 3mm, DSC > 0.8). The results reaffirmed expectation of large segmentation unpredictability for bowels (size and shape variations) and duodenum (caudal to the pylorus). Volume analysis of contours from CT-to-MR showcased negligible volume changes in Liver, Spleen, Kidneys, Stomach and large volumetric changes 80%, 30%, 51% and 58% for A_Aorta, Bowel_Large, Bowel_Small and SpinalCord, for same-date scan cases. Additionally, an anterior shift of contours from CT-to-MR was observed. The average segmentation time was 180 sec for CT, 70 sec for T1 and 50 sec for T2, on an Intel Xeon CPU (2.4 GHz, 64GB RAM) system. Rapid and consistent auto-segmentation of abdominal structures on CT and MRI can be achieved with a deep learning solution. This work justifies further development for much-needed auto-segmentation in MRgART for abdominal tumors.Abstract 209; TableA_AortaKidney_LKidney_RLiverSpinalCordSpleenPancreasDuodenumStomachCT1, 0.90.7, 0.951, 0.921.1, 0.970.58, 0.91.4, 0.942, 0.782.7, 0.755.7, 0.82T10.3, 0.960.5, 0.960.4, 0.970.5, 0.980.7, 0.830.3, 0.983, 0.814, 0.84, 0.9T20.7, 0.920.5, 0.950.8, 0.930.8, 0.971.2, 0.830.4, 0.971.2, 0.854, 0.811, 0.94 Open table in a new tab