Location-Guided Coarse-To-Fine Network For Whole Heart Segmentation

The human heart is a hollow muscular organ with very large variations in shape and size, automatic and accurate segmentation of the whole heart is challenging for medical image analysis. To tackle this challenge, a simple and effective fully automatic coarse-to-fine network is proposed to enable whole heart segmentation. It comprehensively captures the anatomical dependencies of the heart and its substructures as explicit guidance to assist the segmentation of these challenging substructures. Based on the anatomical location of the heart in CT slices, we design a Slice Channel Attention (SCA) module to aggregate specific categories of response channels to highlight feature representation of different slices, thereby enhancing the features of the same substructure and avoiding confusion between different substructures. Furthermore, we propose a Location-aware Object Generation (LOG) module, which can generate the heart mask for guiding heart substructures segmentation by processing the CNN feature map. Experimental results show that our method achieves state-of-the-art on MICCAI 2017 MM-WHS dataset. Ablation studies and visualizations are provided to understand our method.