DualPath-CRC: Adaptive Swin Transformer blocks and Attention-Augmented InceptionNeXt for classification of colorectal cancer histopathology

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Colorectal cancer (CRC) is the third most common cancer and the second leading cause of cancer-related death worldwide, with about 1.85 million cases and 850,000 deaths annually. While a definitive diagnosis depends on the histopathological examination of hematoxylin and eosin-stained tissue slides, considerable variability remains in slide interpretation among pathologists, and manual examination is time-consuming. These limitations have led to the development of computer-aided methods for analyzing colorectal histopathology slides. Recent hybrid Transformer models often lack adaptive channel re-weighting, use fixed window sizes that may restrict the multi-scale receptive field, or rely on traditional feature-fusion techniques that assume simple element-wise addition. To overcome these shortcomings, DualPath-CRC proposes three architectural innovations: (i) Attention-Augmented InceptionNeXt blocks, which apply sigmoid-based channel attention to multi-branch depthwise convolutions; (ii) Adaptive Window Swin Transformer blocks, in which a Window-Size Prediction Network adaptively controls the receptive field of self-attention; and (iii) a Gated Cross-Pathway Fusion module, which fuses the local and global feature streams through learned sigmoid gating. For the classification of histological tissue patches, DualPath-CRC achieves an accuracy of 99.98% on NCT-CRC-HE-100K and 99.34% on Kather-5K. The regions highlighted by Grad-CAM, Grad-CAM++, and LIME analyses provide qualitative information about which parts of the image are important for the model’s predictions and typically correspond to morphologically relevant tissue areas.