GRANet-Lite: A Lightweight Ghost-Residual Attention Network for Breast Cancer Classification with Comprehensive Histopathological Preprocessing Analysis

Breast cancer classification from histopathological images is frequently challenged by staining variability and high computational requirements. This study introduces a novel lightweight convolutional neural network optimized for deployment in resource-constrained clinical environments. The architecture integrates Ghost modules, Squeeze-and-Excitation, and Coordinate Attention mechanisms with residual learning, achieving robust feature representation with only 1.72 million trainable parameters. A systematic evaluation of five preprocessing strategies including Macenko, Vahadane, Reinhard, Ruifrok normalization, and Otsu-based tissue separation was conducted across multiple magnification levels using the BreaKHis dataset. Experimental results indicate that Otsu-based tissue separation at 400x magnification consistently yields superior performance, achieving a 98.54% AUC and 95.33% accuracy. Model interpretability was validated through Grad-CAM++ and Score-CAM to ensure spatially consistent and pathology-relevant feature extraction. By balancing computational efficiency with high diagnostic precision, this framework provides a scalable and explainable solution for breast cancer screening on edge devices and low-resource platforms