Similarity Feature Fusion and Adversarial Shape Constraints for Precise Medical Image Segmentation

Accurate image segmentation can help doctors obtain important pathological information from complex medical images, and is of great significance in preoperative diagnosis, surgical planning, and postoperative evaluation of diseases. Due to the characteristics of physiological structure, the location and shape of lesions in medical image data are extremely irregular, the morphological volume in different individuals is highly variable, and the junction area of different tissues often has fuzzy boundaries. This paper proposes a segmentation method based on similarity feature fusion and adversarial shape constraint. The similarity feature fusion method designed in the feature fusion process filters out interference information and suppresses artifact activation by establishing a similarity matrix of the same spatial position in different feature maps. To address the problem of inaccurate segmentation of lesion edges, we proposed a selfsupervised shape-guided strategy based on a generative adversarial network. By training its encoder-decoder through reconstructing labels and prediction results, the prediction results of the supervised segmentation network are constantly approaching the true labels. We have achieved excellent results on both the polyp dataset and the skin lesion dataset, with mIoU values of $84.11 \%$ and $87.36 \%$, respectively.