MSAF-UNet: A Robust Segmentation Method Integrating Traditional Image Processing and Multi-Scale Attention Fusion

Accurate segmentation of Whole Slide Image (WSI) remains a significant challenge in medical image analysis. Current methods often struggle with effective feature extraction and precise segmentation. To address these, this paper introduces a convolutional neural network, MSAF-UNet, which extracts and fuses color features and texture features. The model employs edge detection operator and Otsu algorithm to highlight texture details, then use K-means clustering and median filtering to highlight color details. The Multi-Scale Attention Fusion (MSAF) module cross-splices the feature maps and integrates the highlighted features through global and local attention branches. Additionally, K-means clustering and Generalized Hough Transform (GHT) methods reduce manual annotation effort while improving segmentation accuracy. Comprehensive comparative experiments and ablation experiments on the CAMELYON16&17 and EBHI-Seg datasets demonstrate that the performance of the proposed method achieved state-of-the-art, verifying the advantages of the proposed method demonstrate that the proposed model outperforms other representative models with 1)an average improvement of 2.27% and reached 94.5% in accuracy; 2)a sound adaptability to support varying image data.