Liver Segmentation in Magnetic Resonance Imaging via Mean Shape Fitting with Fully Convolutional Neural Networks

Abstract

In this work, we propose a novel learning-based segmentation technique for delineating liver volumes in magnetic resonance images. The method utilizes the shape prior of the liver for improved accuracy. Instead of labeling the tissue via binary classification, our method completes the segmentation by deforming a label template of the liver average shape based on the learned image features. The average shape of the liver we used is estimated from a large set of expert-labeled computed tomography images. A fully convolutional neural network (FCN) is trained to maximize the overlap between the deformed liver label template and the ground truth segmentation. The proposed method is validated with 51 T2-weighted liver image volumes and achieves an average Dice coefficient of 95.2% with a mean Hausdorff distance of 20.0 mm. Compared to the results obtained with a standard FCN-based method, a three-fold improvement of the Hausdorff distance is observed, indicating the substantial gains achieved by incorporating the shape prior.