Abstract: Cardiac multi-class segmentation is of great significance in medical imaging, which can provide accurate cardiac structure information and assist clinical diagnosis. However, in the training of multi-class semantic segmentation models with high-resolution cardiac images, the loss of deep features due to multiple downsampling operations leads to the problems oforgan discontinuity and incorrect edge segmentation in the segmented cardiac. To address this, this paper proposes a 3DCSNet based on self-attention and 3D convolution for cardiac multi-class segmentation. Specifically, our proposed network introduces the 3D feature fusion module and a 3D spatial perception module into the segmentation network. The former 3D feature fusion module integrates self-attention and 3D convolution for parallel feature extraction, which is able to efficiently allocate the attentions weights within and between channels under the same dimension of the feature map. The latter 3D spatial perception module captures the positional correlation information between different dimensions by integrating the self-attention mechanism, avoiding the loss of important information in downsampling and further retaining the deep key features. Experimental results show that the proposed 3DCSNet outperforms several existing models on a publicly available 3D computed tomography image dataset (ImageCHD).

Key words: multi-class semantic segmentation, cardiac medical images, 3D convolution, self-attention mechanism, U-Net