[Self-supervised classification method for breast pathology images fusing geometric priors and multi-scale attention].

Abstract

Breast whole slide image (WSI) serves as an essential basis for breast cancer diagnosis and subtype classification. However, its high resolution, multi-scale structural characteristics, and arbitrary orientation pose substantial challenges for automated analysis. Existing deep learning methods typically rely on large amounts of annotated data and struggle to handle the rotational variability and scale differences inherent in pathological images, which limits their generalization ability in cross-center settings. To address these issues, this study proposes a self-supervised learning model, IAM-BYOL, which incorporated geometric priors into the representation learning process. Building upon the BYOL framework, the model introduced an isotropic attention module (IAM). By employing discrete rotation group convolution, IAM enabled weight sharing across different rotated versions of the convolutional kernels, endowing the encoder with structural rotation equivariance. A subsequent group pooling operation converted the equivariant features into rotation-invariant isotropic representations. In addition, a multi-scale attention mechanism adjusted feature weights adaptively according to responses from different receptive fields, allowing the model to capture informative patterns ranging from nuclear-level details to tissue-level organization. Experimental results demonstrated that IAM-BYOL achieved classification accuracies of 98.74%, 99.04%, 99.01%, and 98.63% on the BreakHis dataset at 40×, 100×, 200×, and 400× magnifications, respectively, while attaining an accuracy of 93.02% on the cross-center private clinical dataset BCD. These findings indicate that introducing geometric inductive biases into pathological image representation learning can effectively enhance model robustness and generalization capability.