SwinCNN+OE: A Swin Transformer and CNN Architecture for Breast Histopathology Classification With OOD and Grad-CAM Integration

This study aimed to design and evaluate a fusion deep learning architecture (SwinCNN + OE) for robust and interpretable breast cancer classification using histopathological images. The proposed model combines the local feature extraction strengths of EfficientNet-EO with the global representation power of the Swin Transformer, enhanced by Outlier Exposure (OE), and energy-based OOD scoring to improve safety and real-world adaptability. The SwinCNN+OE model exhibited consistently high performance across all primary classification metrics, achieving an overall accuracy of 97.83% and precision, recall, and F1-score of 97.88%, 97.64%, and 97.68%, respectively. Extensive experiments on a real-world histopathology dataset demonstrated the model’s superior diagnostic performance, achieving an overall accuracy of 96.36%, a macro-averaged F1-score of 96.35%, and a mean AUC score of 99.49%. The model also achieved perfect detection (AUC = 1.00) of unknown samples, illustrating its ability to identify out of distribution cases using energy-based scoring. Furthermore, Grad-CAM heatmaps confirmed that the model consistently focused on biologically meaningful regions such as densely packed nuclei and irregular tissue structures. Integrating Grad-CAM for interpretability and energy-based OOD detection significantly contributes to the trustworthiness and safety of AI applications in clinical pathology.