A Comparative Deep Learning Framework for Automated Breast Cancer Detection Using Histopathological Images

Breast cancer is the leading cause of death from all causes of cancer in women worldwide. Thus, accurate and reliable computer-aided diagnosis are needed for early detection. Histopathological image analysis is the gold standard for diagnosis of breast cancer, but it is laborious, time consuming, and variable among the observers. We present a comparison of deep learning techniques for automatic breast cancer detection from histopathological images. Five convolutional neural networks, VGG16, Custom CNN, ResNet50, DenseNet121, and EfficientNetB0, were evaluated against the publicly available Breast Histopathology Images (IDC) dataset that contains 277,524 image patches from 162 whole-slide images at 40 magnifications. Image preprocessing techniques such as normalization, resizing, and data enhancement were employed before the test. Accuracy, Precision, Recall, F1-score, Specificity, Receiver Operating Characteristics (ROC), Area Under the Curve (AUC), and Confusion Matrix showed superior to traditional convolutional neural network models. The model with the highest accuracy was EfficientNetB0 with an AUC of (0.981) and also had the highest specificity. EfficientNetB0 performed particularly well with an accuracy of 92.74%, precision of (87.38%), recall of (87.38%), F1-score of (88.26%), specificity of (94.71%, and AUC of (0.981), showing superior performance in distinguishing between malign and benign breast tissues. DenseNet121 performed well in a diagnostic performance, while ResNet50 achieved the highest specificity. The results confirm that EfficientNetB0 features learned well, better classification accuracy, and reliable diagnostic performance, making it suitable for computer-aided breast cancer diagnosis, and a promising model for future intelligent healthcare.