Bridging the Gap Between Accuracy and Efficiency in AI-Based Breast Cancer Diagnosis from Histopathological Data.

Background/Objectives: Breast cancer diagnosis using histopathological images remains a critical yet challenging task in computational pathology due to overlapping visual features between benign and malignant tissues, inconsistencies in staining, and variations in magnification. The objective of this study was to design a lightweight yet high-performing deep learning model that bridges the gap between diagnostic accuracy and computational efficiency. Methods: We propose CellSage, a novel convolutional neural network (CNN) architecture enhanced with attention mechanisms. It integrates three core modules: a multi-scale feature extraction unit designed to capture both global and local tissue patterns; a depthwise separable convolution block for reducing computational load; and a Convolutional Block Attention Module (CBAM) to dynamically focus on diagnostically relevant regions. The model was trained and evaluated on the BreakHis dataset, using stain normalization (via contrastive augmentation modeling, CAM) and extensive data augmentation techniques. A patient-wise cross-validation strategy was employed to ensure robust generalization. Results: CellSage achieved 94.8% accuracy, a 0.93 F1 score, and a 0.96 AUC, while remaining compact at only 3.8 million parameters. It outperformed deeper and larger models such as ResNet-50, DenseNet-121, and Vision Transformers in terms of both predictive performance and computational efficiency. Ablation studies confirmed that multi-scale feature extraction and attention refinement were critical components. Conclusions: CellSage is an interpretable, reliable, and computationally lightweight system for breast cancer diagnosis using histopathological data. Its efficiency and low computational footprint render it an ideal candidate for real-time deployment on digital pathology platforms, particularly in environments with limited computational infrastructure.