Whole Heart Segmentation Method Based on Position Encoding Guidance
Minghui Sima, Shihao Bai, Lin Wang, Jun Wang · 2023
The heart is an important organ in the human circulatory system. In recent years, whole heart segmentation methods based on deep learning have achieved good results, but most of these algorithms fail to fully extract the global feature information of each substructure in the heart, and can not accurately segment the fuzzy boundary pixels in the heart. Aiming at the problem of whole heart segmentation, this work proposes a whole heart segmentation based on position encoding guidance. According to the particularity of the heart, this method uses the adjusted distance formula to generate the heart position information, and embeds it into the proposed network. The self-attention mechanism is introduced into the lower three layers of the network skip connection, and the internal relationships of the feature map are calculated by adding trace parameters to make full use of the features of each pixel. A multi-scale feature fusion module is introduced in the upper two layers of network skip connection to solve the problem of small receptive field of shallow feature maps. Experiments show that the proposed method has improved the segmentation efficiency of different substructure boundaries of the heart.