Uncertainty Quantification to Improve the Classification of Melanoma and Basal Skin Cancer Using ResNet Model

Abstract

Traditional skin cancer screening necessitates a time-consuming physical examination by a dermatologist. One of the most difficult tasks in image analysis is automated medical picture recognition and categorization. Making a skin-based analysis to predict the abnormalities present in a particular area is always a competitive issue. In the case of automatic region identification, deep learning (DL) methodologies play a major role. Most DL approaches are unable to offer uncertainty quantification (UQ) for the output, resulting in overconfidence and excessive sample deviation. The dataset was contributed by the International Skin Image Collaboration (ISIC) Archive. To deal with uncertainty during skin cancer image classification and disappearing gradients, we used Monte-Carlo (MC) dropout with the ResNet 50 DL model. This assists in overcoming the problem of vanishing gradients. After the preprocessing, the training was completed, yielding a 94% accuracy rate. The MC dropout method identifies the out-of-domain data, reduces the training size which results in an accuracy of 89.9%. As a result, the ResNet50 model for classifying skin cancer achieved 89% of accuracy. In addition, our model was in general robust to noise and performed well with previously unseen data.