HFA-Net: 3D Cardiovascular Image Segmentation with Asymmetrical Pooling and Content-Aware Fusion

Abstract

Automatic and accurate cardiovascular image segmentation is important in clinical applications. However, due to ambiguous borders and subtle structures (e.g., thin myocardium), parsing fine-grained structures in 3D cardiovascular images is very challenging. In this paper, we propose a novel deep heterogeneous feature aggregation network (HFA-Net) to fully exploit complementary information from multiple views of 3D cardiac data. First, we utilize asymmetrical 3D kernels and pooling to obtain heterogeneous features in parallel encoding paths. Thus, from a specific view, distinguishable features are extracted and indispensable contextual information is kept (rather than quickly diminished after symmetrical convolution and pooling operations). Then, we employ a content-aware multi-planar fusion module to aggregate meaningful features to boost segmentation performance. Further, to reduce the model size, we devise a new DenseVoxNet model by sparsifying residual connections, which can be trained in an end-to-end manner. We show the effectiveness of our new HFA-Net on the 2016 HVSMR and 2017 MM-WHS CT datasets, achieving state-of-the-art performance. In addition, HFA-Net obtains competitive results on the 2017 AAPM CT dataset, especially on segmenting subtle structures among multi-objects with large variations, illustrating the robustness of our new segmentation approach.