Self-screening of noisy labels and hard sample color normalization in histopathology image classification with spatial-texture attention networks

Deep learning-based pathological image classification has emerged as a promising tool to aid pathologists in diagnostics. However, two critical challenges hinder its accuracy and generalization: noisy labels in datasets due to annotation errors or subjective judgments, and unique image characteristics such as multiscale textures and staining variations. This study aims to address these challenges by proposing a novel framework that combines noise-robust learning and feature-specific attention mechanisms. A noise-resistant algorithm integrating noisy sample self-screening and hard sample color normalization (NSS-HSCN) was proposed, along with a dual-stream spatial and texture attention (DSSTA) framework. In the NSS-HSCN stage, a self-screening network filtered out noisy samples, and hard samples were subjected to color normalization. The DSSTA framework utilized a multi-scale spatial attention module and a texture-enhanced attention module to extract and fuse features. Experiments were carried out on the Chaoyang and HITAFH datasets. The method outperformed other leading methods on both datasets. On the Chaoyang and HITAFH datasets, our method achieved 86.02% and 90.16% accuracy, respectively, outperforming state-of-the-art methods in all metrics. Grad-CAM visualization verified the model’s ability to focus on target areas and extract valuable features. The noise-screening network strengthened model robustness, and the dual-stream network effectively integrated features. Integration with the automated diagnostic system optimized the diagnostic process, thereby highlighting its potential for improving pathological image classification accuracy in real-world applications.