An Enhanced Deep Learning Framework for Breast Cancer Histopathological Image Classification

Abstract

Accurate and timely identification of breast cancer (BC) is crucial for effective clinical decision-making and treat-ment planning. Histopathological image analysis provides valu-able information for BC diagnosis, although manual assessment is labour-intensive and subject to differences between observers. In this work, a deep learning-based framework is developed to automatically classify breast histopathological images from the BreakHis dataset. The framework leverages an EfficientNet-based transfer learning strategy and integrates focal loss with label smoothing, data augmentation, cosine learning-rate decay, and fine-tuning to improve learning robustness and classification performance. Our model is designed to optimize malignant recall with an overall high level of accuracy and stability. Experimental studies state that the proposed method achieves a test accuracy of 94.08%, AUC of 0.9727, a malignant recall of 99%, and a significant improvement over multiple baseline CNN architectures. Through detailed evaluation using confusion matrix analysis, ROC curve analysis and learning dynamics, it has been observed that the proposed framework demonstrates the effectiveness and stability for BC histopathological image classification.