Evaluating The Corruption Robustness Of Convolutional Neural Networks For The Task Of Classifying Skin Diseases

Skin disease classification by machine learning models is an upcoming field of research that shows great potential to assist dermatologists. These techniques are, however, still vulnerable to corruptions that can be present in the dermoscopic images. In this paper, we tested two promising methods to improve the robustness against corruptions: augmenting the training datasets with corrupted images and pre-processing the images with Contrast Limited Adaptive Histogram Equalization (CLAHE). We found that the presence of corrupted images in the training datasets can greatly improve the corruption robustness while CLAHE harms the classification accuracy of the models when faced with corrupted images. Our benchmarks can be used as a starting point to further develop AI models that can be used as reliable diagnostic tools.