Explainable Attention-Enhanced Approach for Multimodal Breast Cancer Diagnosis Across Diverse Imaging Modalities

Abstract

Early and accurate detection of breast cancer is critical for improving survival rates. This study presents a robust deep learning framework that integrates convolutional and attention-based modules to enhance feature extraction across various imaging modalities. The proposed model is evaluated on four benchmark breast cancer datasets: BreakHis (400×), INbreast, BUSI, and CBIS-DDSM, which capture variations in histopathological, mammographic, and ultrasound images. A stratified fivefold cross-validation strategy was adopted to ensure model generalizability. The proposed approach achieves outstanding classification performance, with accuracies of 98.75% on BreakHis, 99.12% on INbreast, 98.40% on BUSI, and 99.05% on CBIS-DDSM. These results consistently surpass those of traditional CNNs and recent baseline models, such as ResNet50, DenseNet121, EfficientNet-B0, and Vision Transformers, across all datasets. A detailed ablation study confirms the effectiveness of each component in the architecture. A computational cost analysis demonstrates that the proposed model achieves superior accuracy with reduced training epochs and competitive inference times.