Multi-Task Consistency Segmentation Network with Deep Shape Prior for Cartilage Segmentation

Abstract

We introduce innovative design concepts to address hard segmentation cases in complex image backgrounds and small training data sizes. We define (1) the shape feature description, including the medial surface, mask segmentation, and contour shape, which characterize articular cartilage shape, and establish a cartilage segmentation network to achieve multi-task consistency; and (2) the shape prior description, representing the shape distribution of articular cartilages, and establish a neural network based on this description. We incorporate the shape prior network into the multi-task consistency segmentation network. This results in a deep learning framework with high accuracy and strong generalization, guided by shape feature description and constrained by shape prior description. Our framework handles difficult cases with low contrast, ambiguous boundaries, deformed portions, and touching cartilages. The effectiveness of our method is demonstrated on two public knee image datasets and one clinical hip image dataset, where our approach shows increased segmentation accuracy compared to other state-of-the-art methods. Furthermore, its generalization is demonstrated for a subset of the BTCV dataset focusing on three specific structures: the aorta, the inferior vena cava, and the portal and splenic veins.