Leveraging CNN and Transfer Learning for Classification of Histopathology Images

This paper aims to build a hybrid convolutional neural architecture by leveraging the power of pre-trained ResNet 50 (trained on ImageNet dataset) through transfer learning. The proposed work has achieved state-of-the-art performance metrics on the BreakHis dataset, containing microscopic histopathological images of benign and malignant breast tumours. The model incorporates global average pooling, dropout and batch normalisation layers on top of the pre-trained ResNet50 backbone. This methodical superimposing of the GAP layer in tandem with Resnet50's knowledge and training is the proposed novelty taking our model the extra mile. As a result, the binary classification problem between benign and malignant tumours is handled gracefully by our proposed architecture despite the target imbalance. We achieve an AUC of 0.946 and an accuracy of 98.7% which is better than the previously stated standard.