Lightweight Breast Cancer Histopathological Image Classification Using a Fuzzy-Enhanced Mobilenet Architecture

Abstract

Breast cancer is one of the most common cancers in women and requires early, accurate diagnosis to improve the chances of cure. Histopathology image analysis plays a crucial role in diagnosis, but manual interpretation by pathologists is time-consuming and complex. Various deep learning approaches, particularly Convolutional Neural Networks (CNN), have been developed for automatic medical image classification. Conventional CNN models still show limitations in interpretability, probability calibration quality, and computational efficiency. This study proposes Fuzzy-Enhanced MobileNetV2 by integrating a fuzzy attention mechanism into the MobileNetV2 architecture to improve feature representation and model transparency without compromising computational efficiency. Evaluation was conducted on the BreakHis dataset at 400× magnification using accuracy, precision, recall, F1-score, ROC-AUC, and Brier Score metrics. Experimental results show that the proposed model achieves an accuracy of 95.97%, comparable to the baseline. This model also produces a higher precision of 96.09% and a lower Brier Score of 0.028 compared to 0.032 in the baseline, indicating a better quality of probability estimation. The proposed model has 4.60 million parameters, a complexity of 0.854 GFLOPs, and an inference time of 0.95 ms per sample. Attention map visualizations show improved interpretability. These findings suggest that fuzzy attention in lightweight CNNs can improve transparency, probability reliability, and computational efficiency.