DMoC-UNet: A Dynamic Mixture-of-Convolution Network for Enhanced Pathological Image Segmentation

Pathological image segmentation is a cornerstone in medical image analysis and is crucial for tumor detection, tissue classification, and pathological diagnosis. Existing methods face challenges in addressing complex and diverse tissue structures, multi-scale features, and blurred boundaries, limiting their segmentation accuracy and generalization across datasets. This paper proposes a novel pathological image segmentation method, DMoC-UNet, which integrates Dynamic Mixture-of-Convolution (DMoC) modules, Haar wavelet downsampling, and Dual Attention Fusion (DAF) modules to enhance multi-scale feature extraction and fine-grained boundary segmentation. The DMoC modules enable dynamic routing of features to specialized expert networks, adapting effectively to the diverse tissue characteristics of pathological images. Haar wavelet downsampling preserves spatial details while improving multi-scale representation, and the DAF modules facilitate efficient fusion between shallow and deep features, ensuring semantic consistency. Extensive experiments on three publicly available datasets—EBHI, CRAG, and GlaS—demonstrate that DMoC-UNet outperforms state-of-the-art models, achieving significant improvements in Accuracy of Classification (ACC), Dice Similarity Coefficient (DSC), and Intersection over Union (IoU) metrics.Specifically, DMoC-UNet achieves ACC, DSC, and IoU values of 92.46%, 94.65%, and 90.20% on the EBHI dataset; 85.71%, 88.26%, and 82.79% on the CRAG dataset; and 85.81%, 87.97%, and 83.56% on the GlaS dataset. These results highlight the robustness and adaptability of DMoC-UNet, making it a promising approach for pathological image segmentation. Our code will be released after acceptance.