Enhanced u-net for lesion segmentation in whole-slide images: Integrating attention mechanisms and multi-scale feature extraction

Abstract

Lesion segmentation in whole-slide histopathological images remains challenging due to diverse tissue patterns and varying lesion sizes, with conventional segmentation methods often struggling to maintain accuracy across different scales. This research introduces a Fine-grained Scaling Segmentation Model (FGSSM) that enhances traditional U-Net architecture through dual classifier pairs and attention mechanisms to capture multi-scale features. The model incorporates nonlinear learning dynamics through adaptive attention modules and dual classifier feedback, enabling it to respond flexibly to complex pixel-level variations. These nonlinear mechanisms support robust boundary recognition and enhance the network’s sensitivity to subtle pathological textures. The proposed model was evaluated using a comprehensive dataset of 2,500 whole-slide images from the GasHisSDB database, implementing a 10-fold cross-validation approach. FGSSM demonstrated substantial improvements over the baseline U-Net, achieving 94.3 % overall segmentation accuracy and showing particular strength in handling varying lesion sizes, with 92 % specificity for small regions (<100 pixels) and 95 % for larger areas. The model architecture yielded a 12 % increase in precision (0.91 vs. 0.79) and a 15 % improvement in F1-score (0.93 vs 0.78) compared to standard U-Net implementations. Integrating adaptive scaling factors with attention mechanisms significantly reduced false positives by 30 %, especially in challenging cases with overlapping tissue patterns. These results demonstrate that FGSSM offers a robust solution for accurate lesion segmentation across diverse histopathological contexts, making it particularly valuable for clinical applications.