Blurry Boundary Segmentation with Semantic-Aware Feature Learning

Encoder-decoder architectures are widely adopted for medical image segmentation tasks. These models utilize lateral skip connections to capture and fuse both semantic and resolution information in deep layers, enhancing segmentation accuracy. However, in many applications, such as images with blurry boundaries, these models often struggle to precisely locate complex boundaries and segment tiny isolated parts due to the fuzzy information passed through the skip connections from the encoder layers. To solve this challenging problem, we first analyze why simple skip connections are insufficient for accurately locating indistinct boundaries. Based on this analysis, we propose a semantic-guided encoder feature learning strategy. This strategy aims to learn high-resolution semantic encoder features, enabling more accurate localization of blurry boundaries and enhancing the network's ability to selectively learn discriminative features. Additionally, we further propose a soft contour constraint mechanism to model the blurry boundary detection. Experimental results on real clinical datasets demonstrate that our proposed method achieves state-of-the-art segmentation accuracy, particularly in regions with blurry boundaries. Further analysis confirms that our proposed network components significantly contribute to performance improvements. Experiments on additional datasets validate the generalization ability of our proposed method.