Medical Image Segmentation Based on Clustering Feature Information Enhancement

Medical image segmentation provides reliable basis for clinical diagnosis and pathology research, and has important medical application value. However, it remains difficult to obtain a large number of accurate annotation data in medical image segmentation. Recently, some progress has been made in a study based on metric learning, which adopts self-supervised learning on prototype network. In this work, an adaptive local prototype pooling method was plugged into prototype network to solve the common challenge of foreground–background imbalance in medical image. However, this study ignored the influence of foreground–background imbalance and intra-class imbalance on query image, which would lead to intra-class differences in query features extracted from the query branch. Therefore, the differences of intra-class similarity score would be generated when calculating the similarity between the support prototypes and the query features. In order to solve the problem, we proposed a clustering feature information enhancement method that aims to fuse prior information obtained through clustering into the segmentation network. We optimized the similarity score through an information interaction module, and introduced Gaussian filter to smooth the clustering noises. We conducted experiments on the ABD-MRI dataset, Which is from ISBI 2019 Combined Healthy Abdominal Organ Segmentation Challenge. Our results demonstrate the effectiveness of fusing the clustering feature information, and the mean Dice coefficient has been improved by more than 5%.