Automated detection of tumor-associated collagen signatures in breast cancer histopathology using deep learning
Yihui Zheng, Deyong Kang, Lianhuang Li, Liqin Zheng, Jianhua Chen, Wenhui Guo, Fangmeng Fu, Chuan Wang, Jianxin Chen, Xiahui Han · 2025
Breast cancer persists as a major public health concern for women worldwide. Within the tumor microenvironment (TME), the architectural organization of collagen fibers, characterized by tumor-associated collagen signatures (TACS) at the invasive front, holds significant prognostic significance. However, conventional identification of these signatures relies on costly and technically complex multiphoton microscopy (MPM). To address this limitation, we developed a novel deep learning methodology for the automated classification of TACS4–6 directly from standard hematoxylin and eosin (H&E) stained histopathological whole-slide images. Our approach leverages MPM imaging as the reference standard to generate annotations of TACS features. These annotations were precisely mapped to corresponding regions on paired H&E-stained tissue sections to construct a dedicated, high-fidelity dataset. Following comprehensive preprocessing and data augmentation, a ResNet-based convolutional neural network (CNN) was implemented and trained for robust feature extraction and multi-class classification. This study establishes a cost-effective, scalable, and clinically accessible deep learning framework for the accurate classification of prognostically relevant TACS patterns within routine diagnostic H&E-stained images. The proposed methodology represents a promising new tool to enhance breast cancer prognostication and inform personalized therapeutic strategies.