Skin cancer recognition for whole slide histology images with state-of-the-art Convolutional Neural Networks

Whole Slide histology image remains the gold standard. But recent research reveals a high discordance between individual pathologists. In order to reduce the misdiagnosis made by doctors, the aim of this study is to using deep learning to assist pathologists for skin cancers diagnosis. 626 whole slide images were collected (162 melanoma, 115 basal cell carcinoma, 349 nevus) from Xiangya Hospital. All the lesions in the WSIs were selected and random sampled as patches. The resulting patches of 500 WSIs were used for the training of a convolutional neural network (CNN). The resulting patches of addition 126 WSIs were used to test the performance of the CNN. We trained in ResNet50, MobileNet and InceptionV3 by transfer learning. The result shows that the InceptionV3 model achieves superior performance than other models (InceptionV3: 0.923 accuracy, MobileNet: 0.921 accuracy, ResNet50: 0.919 accuracy.) in the classification task. To locate the lesion of skin cancer, we visualized the prediction confidence map of the model in WSIs. The result showed that the model can successfully located the lesion area in three diseases, which could offer more reference to pathologists . In conclusion, CNNs indicate to be a valuable tools in diagnosing skin cancer and can provide more support for specialists beyond just the diagnosis.