Multi-View Semi-Supervised Attention Network for 3D Cardiac Image Segmentation

Huaidong Li¹Delong Li¹Qing Dong²Xue Han³Suyu Dong^1*

• ¹Northeast Forestry University, Harbin, China
• ²Harbin Medical University, Harbin, Heilongjiang, China
• ³Harbin Medical University Cancer Hospital, Harbin, Heilongjiang Province, China

In recent years, semi-supervised methods have rapidly developed for three-dimensional medical image analysis. However, previous semi-supervised methods for three-dimensional medical images usually focused on single-view information and required a large number of annotated datasets. In this paper, we innovatively propose a multi-view (coronal and transverse) attention network for semi-supervised 3D cardiac image segmentation. In this way, the proposed method obtained more complementary segmentation information, which improved the segmentation performance. Simultaneously, we integrated the attention module with the VNet to enhance the focus on the segmentation regions and edge portions. We first introduced the CutMix data augmentation mechanism to enhance 3D cardiac medical image segmentation. In this way, the proposed method made full use of the mixed regions in the images and expanded the training dataset. Our method was tested on two publicly available cardiac datasets and achieved good segmentation results. Our code and models are available on https://github.com/HuaidongLi-NEFU/TPSSAN.