AI-Assisted Early Detection of Breast Cancer Using Histopathological Image Analysis

Abstract

Breast cancer is one of the most common cancers among women worldwide with more than 2.3 million new breast cancer cases diagnosed each year [1], where early detection has a significantly positive effect on patient survival rates. Examination of histopathological features remains the gold standard [2] in the diagnosis of breast cancer, however, clinical histopathological assessment is labor-intensive, time consuming and can potentially have inter-observer variability in the determination of the final clinical diagnosis. The current paper presents an AI-assisted deep learning framework for automatic breast cancer detection using histopathological images. Multiple transfer learning models, including VGG16, ResNet50, DenseNet121, EfficientNetB0, and EfficientNetB5, were evaluated. The proposed framework incorporates data augmentation, mixed precision training, and Binary Focal Loss to improve performance and computational efficiency. Experimental results on a dataset of over 14,000 histopathological images demonstrate that DenseNet121 achieves the best performance, with a test accuracy of 81%, sensitivity of 80.8%, specificity of 76.7%, and ROC-AUC of 0.87. VGG16 also shows stable performance, while ResNet50 suffers from low specificity due to over classification of malignant classes. EfficientNet models exhibit poor generalization and strong class bias. The result highlights the importance of evaluating multiple performance metrics beyond accuracy in medical image classification. The proposed frameworks can assist pathologists by improving diagnostic consistency and reducing workload, making it a practical tool for clinical applications.