Performance Evaluation of Multi-Contrast Dixon MRI and CT for Abdominal Fat and Muscle Segmentation Using a UNet CNN

We evaluate the performance of a deep learning framework for segmenting abdominal fat and muscle using multi-contrast Dixon magnetic resonance (MR) and computed tomography (CT) images. We aim to compare MR image segmentation by testing Dixon fat-only, water-only, and combination of both types of images and comparing the results with CT images. Nineteen subjects underwent abdominal CT and Dixon MR imaging on the same day. For each participant, three pairs of matched axial images from both CT and MR were selected at the intervertebral levels of L2-L3, L3-L4, and L4-L5 for analysis. References labels were generated through semi-automated segmentation of subcutaneous adipose tissue, visceral adipose tissue, and muscle areas. They were then used to train and evaluate a U-Net based Convolutional Neural Network (CNN) framework with a 3-fold cross-validation to compare the segmentation performance across CT, Dixon fat-only and water-only MR images. Combining the fat-only and water-only MR image inputs produced superior results in all labels. Our study demonstrates that CNN-based segmentation performance for abdominal fat and muscle improves with the inclusion of additional input channels, such as combining Dixon fat-only and water-only MR images. While CT results represent the gold standard in abdominal image segmentation, increasing the number of input image channels used in MR segmentation can approach, and even match, the results of CT.