Evaluating neural network robustness for melanoma classification using mutual information

Abstract

Deep Neural Networks (DNN) have made rapid progress in medical image analysis, and it is now common to see algorithms proposed in the literature that claim to meet or exceed clinician performance. However, DNNs typically operate like black box systems. Deployment of these algorithms for safety-critical tasks, such as medical applications, is therefore challenging without methods to characterize and identify the robustness of DNNs in a generalized setting. Furthermore, DNNs are known to be sensitive to small changes in input data, and may rely on spurious correlations seen in the training samples which precludes them from generalizing to data with a different distribution. We previously proposed an attribute-ranking algorithm that ranks discrete data attributes by how informative they are about the DNN performance using Mutual Information. In this study we leverage this algorithm in a novel way to determine if the data attributes that impact the DNN performance are clinically relevant and thereby whether the DNN predictions are robust. We demonstrate the applicability of this method on melanoma classification with a DNN trained on the publicly available HAM 10,000 dataset and achieve 0.855 AUC on the held-out HAM test set. Our analysis identifies that image saturation, which is not a clinically relevant feature, is highly indicative of both whether an image is melanomic in the HAM training data and whether the DNN prediction is correct. Further testing reveals that when the classifier is tested on the SIIM-ISIC Melanoma dataset, where the correlation between image saturation and melanoma is not present, the classifier achieves only an AUC of 0.591, confirming that the DNN is not robust.