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

Hao Zheng, Lin Yang, Jun Han, Yizhe Zhang, Peixian Liang, Zhuo Zhao, Chaoli Wang, Danny Z. Chen

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.
BibTeX
@inproceedings{zheng2019hfa,
  author = {Hao Zheng and Lin Yang and Jun Han and Yizhe Zhang and Peixian Liang and Zhuo Zhao and Chaoli Wang and Danny Z. Chen},
  title = {{HFA-Net: 3D Cardiovascular Image Segmentation with Asymmetrical Pooling and Content-Aware Fusion}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2019},
  doi = {10.1007/978-3-030-32245-8\_84},
}

Loading related content...

Figure