Breast Cancer Detection in Histopathology Images using Attention-Based Deep Learning

Breast cancer is the major cause of deaths among women globally and the early detection of the disease is important in enhancing the survival of the victims. The gold standard in the confirmation of breast cancer is considered histopathology examination, which is labor intensive, subjective and likely to give inter-observer variability when manual interpretation of tissue slide is used. In this research We apply the IDC breast histopathology dataset and come up with a balanced preprocessing plan to minimize the bias of both classes of cancerous and non-cancerous patches. There are four transfer-learning architectures, namely, EfficientNet-B0, ResNet50, DenseNet121, and VGG16, that are trained and tested to classify malignant and benign tissue regions. The suggested scheme includes attention-based scoring of confidence, which allows the model to identify salient regions in the image and justify more understandable predictions. The experimental findings indicate that among all the trained models DenseNet121 performs the best with an accuracy if 93%, followed by good F1 score, recall and precision.