Interpretable CRAM‑Enhanced Lightweight Dual‑Branch CNN for Real‑Time Breast Cancer Histopathology in Internet‑of‑Medical‑Things Environments

Abstract

Breast cancer remains a primary global health concern, with histopathological image analysis serving as the diagnostic gold standard. However, manual microscopy is time-consuming and often subjective. While deep learning offers a powerful solution, existing models are typically too complex and opaque for real-time use in Internet of Medical Things (IoMT) environments. To address this, we propose an interpretable and lightweight hybrid deep learning model that combines MobileNetV2 and EfficientNet-B0, enhanced by a novel contextual recurrent attention module (CRAM). CRAM refines fused features through attention-based weighting, improving focus on diagnostically relevant regions. The model achieved 99.9% classification accuracy and an AUC of 1.00, outperforming standalone baselines while remaining efficient (∼12 M parameters) and suitable for IoMT deployment. Interpretability is ensured through integrated Grad-CAM and SHAP analyses, which visually and quantitatively explain predictions by highlighting malignant tissue features that align with pathologist judgment. This balance of accuracy, efficiency, and transparency enables real-time, trustworthy diagnostics for resource-limited and point-of-care settings. Future work includes extending to multi-class tumor subtypes and clinical validation in real-world workflows. The proposed system represents a significant step toward making AI in digital pathology more accessible and explainable.