Quantifying Uncertainty of Deep Neural Networks in Skin Lesion Classification

Deep neural networks are becoming the new standard for automated image classification and segmentation. Recently, such models are also gaining traction in the context of medical diagnosis. However, when using a neural network as a decision support tool, it is important to also quantify the (un)certainty regarding the outputs of the system. Current Bayesian techniques approximate the true predictive output distribution via sampling, and quantify the uncertainty based on the variance of the output samples. In this paper, we highlight the limitations of a variance based metric, and propose a novel uncertainty metric based on the overlap of the output distributions. We show that this yields promising results on the HAM10000 dataset for skin lesion classification.