Attentive Dual-Branch CNN–Transformer Fusion for Cross-Dataset Breast Histopathology Classification

Abstract

In general, breast histopathology classifiers exhibit significant performance degradation when applied to images acquired using different staining protocols, magnification rates, scanner equipment, or institutional environments. This study proposes an attentive dual-branch CNN–Transformer framework intended to learn representations that remain informative with respect to heterogeneity in histopathological data. The convolutional path captures fine cellular details, such as nuclear texture and local glandular structures, whereas the dilated convolution path captures more generalized tissue structures without extensive downscaling of spatial dimensions. Attention operations were applied separately to the two paths to diminish the impact of background areas and structures with low discriminative values. The generated feature maps were then represented in terms of spatial tokens, which were processed by a transformer encoder to represent the long-range dependencies between distant regions in the tissue structure. A learnable fusion gate regulates the contribution of the local and contextually informative parts of the input features for each image to the output. The institution-specific classification layers handle the distinct diagnostic classes in the BACH and BreaKHis datasets. The evaluation procedure included stratified testing on BACH and patient-disjoint partitioning on BreaKHis to exclude any potential data leakage at the subject level. For BACH, our model achieved an accuracy of 94.75%, a macro F1-score of 93.90%, and an area under the receiver operating characteristic curve of 96.35%, with 18.9 million trainable parameters and an average inference time of 16.1 ms per image. A component-wise ablation study, calibration evaluation, confidence interval estimation, and attention-driven visual explanations were conducted to investigate the contribution and reliability of the proposed fusion method. The resulting architecture represents a trade-off between prediction performance and computational costs but does not claim superior efficiency or applicability. Multi-institutional and prospective independent validation is necessary for further clinical application.