Hybrid CNN Architecture for Reliable Histopathological Breast Cancer Classification with Attention-Based Feature Fusion and Cross-Domain Generalization

• Mradul Kumar Jain

  Author

• Dr. Veerendra Yadav

  Author

• Dr. Anu Chaudhary

  Author

The gold standard of breast cancer diagnosis from hematoxylin and eosin (H&E)-stained histopathological images is limited by the presence of inter-observer variability, complexity of high-resolution images, and limited inter-class morphological differences. In this paper, a hybrid convolutional neural network (CNN) architecture that incorporates a dual-encoder framework and an attention-driven feature fusion mechanism is proposed to accomplish excellent and generalizable performance for the classification of breast cancer. The global encoder captures the macro-architectural tissue context, whereas the local encoder captures the microscopic cellular morphology. The two modalities are then fused into a discriminative diagnostic embedding using an adaptive attention module, followed by a calibrated probabilistic classifier. The proposed architecture was validated on the BreakHis dataset using stratified cross-validation and then on the PatchCamelyon dataset to validate the cross-domain generalization performance. The proposed architecture achieved 96.32 % and 89.27 % classification accuracy, respectively, and demonstrated excellent F1-scores and Matthews Correlation Coefficients, as well as excellent calibration performance with a reduced Brier score and expected calibration error. The incorporation of patient-level aggregation also enhances clinical interpretability. The results demonstrate that the proposed architecture can accomplish excellent discrimination, domain robustness, and probabilistic trustworthiness, and thus provides a scalable and deployment-ready platform for computational histopathology.

• PDF

Copyright (c) 2026 Package Printing

This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.

All articles published are made available under the Creative Commons Attribution–NonCommercial 4.0 International (CC BY-NC 4.0) licence.