Abstract: Aiming at the problems of insufficient accuracy of pathological image recognition and semantic gaps in feature fusion process caused by the diversity of lesion morphology, we proposed an improved U-Net architecture that integrated Transformer and attention mechanisms. Firstly, we designed a boundary-aware module to enhance the expression of lesion edge features in pathological images, thereby improving the model’s ability to perceive complex structures. Secondly, we introduced a regularized large-kernel attention module at the bottleneck layer to model long-range dependencies, and mitigated overfitting risk through a layer-wise regularization strategy. Finally, we further introduced a learnable visual center module to strengthen the complementarity between global and local features. Experimental results on the MoNuSeg and GlaS datasets show that the proposed method outperforms current mainstream models in terms of segmentation accuracy and boundary clarity.

Key words: boundary-aware, feature fusion, pathological image, image segmentation, convolutional neural network, Transformer architecture