Deep residual neural networks for automated Basal Cell Carcinoma detection

摘要

Despite being the most common cancer around the world, five year survival rates of skin cancer are above 95%, as long as they are detected and treated before they have spread. The key to prolonged survival is early detection, making an automated analysis tool indispensable. Current machine learning analyses of Basal Cell Carcinoma (BCC) dermoscopy images have failed to create a model viable for use in clinical applications. In this paper, we demonstrate a sensitivity and specificity that could make neural networks a realistic tool for dermatologists. Our algorithm follows a three step process: first, the original image is preprocessed and fed into the segmentation model; second, a black and white lesion map is produced to extract the minimum area of the image; third, the classification model is introduced for classifying whether an input image is BCC or not. By building upon melanoma research performed by He, et al., we reached an overall weighted sensitivity and specificity of 96% and 89%, respectively. We demonstrated that deep residual neural networks (> 100 layers), carefully optimized, can surpass the limitations of depth one sees with more common convolutional neural networks.