Robust breast cancer histopathology classification using a self-supervised multi-scale adaptive convolutional transformer

Background and objective Histopathological examination remains the gold standard for breast cancer diagnosis. However, manual review is time-intensive and subject to inter-observer variability. While deep learning models offer automated solutions, their clinical translation is often hindered by domain shifts (e.g., staining variability) and the heavy reliance on extensive annotated datasets. This study proposes a novel intelligence-based framework to address these challenges and improve classification generalizability. Methods This paper introduce MACT-CSA (Multi-Scale Adaptive Convolutional Transformer optimized by Crow Search Algorithm), a hybrid architecture combining the local feature extraction capabilities of multi-scale convolutions with the global contextual modeling of transformers. To mitigate annotation scarcity and domain shift, the model employs a contrastive self-supervised pre-training strategy. Furthermore, the Crow Search Algorithm is utilized for automated hyperparameter tuning to optimize the model's architecture for histopathological image analysis. Results The proposed MACT-CSA framework was evaluated on the BreaKHis dataset, achieving an overall accuracy of 83.46%. Notably, it demonstrated high sensitivity in detecting malignancies, with an F1-score of 88.07% and a recall of 90.12% for the malignant class. To evaluate cross-domain robustness, external validation was performed on the BACH dataset, yielding an accuracy of 81.9%, which underscores the framework's capability in handling diverse clinical data. Conclusions Directly addressing the initial objectives, the MACT-CSA framework successfully provides a label-efficient and robust solution for breast cancer histopathology classification. By integrating contrastive self-supervised pre-training with a hybrid transformer-convolutional architecture, the model effectively mitigates the targeted challenges of domain shift and annotation scarcity, achieving strong cross-domain generalization. Although moderate specificity warrants future refinement, this framework establishes a reliable baseline to assist pathologists in automated and generalizable cancer diagnosis.