Domain Shifts and Interpretability in AI-based Skin Cancer Diagnosis

Abstract

Explainability is crucial in order to build trust in AI applications and increase their acceptance, particularly in safety-critical environments. One such application is skin cancer classification, where dermatologists may use AI models as digital assistants for diagnostic purposes. In this work, the publicly available domain-separated ISIC-Archive dataset, comprised of melanoma and nevus image data, is examined. Initially, an inter-domain dataset is used to train a binary ResNet18 classifier and the Grad-CAM output is interpreted. Subsequently, the performance and Grad-CAM output of a binary classifier is investigated to understand how neural network activations change during domain shifts within the ISIC-Dataset. Thereafter, the established domain shift mitigation approaches - i.e. augmentation and DANN - are investigated regarding their influence on performance and neural network activations. Their effects on Grad-CAMElementwise output are furthermore quantitatively compared. The findings presented in this work, provide insights into the underlying reasons for unequal performance degradation during domain shift, shortcomings of Grad-CAM and influential factors of unsupervised domain adaptation. The results indicate a limited potential of Grad-CAM based explainability methods for building trust among dermatologists in domain-adapted ResNet models. Nevertheless, Grad-CAM based explainability methods have demonstrated their ability to identify failure modes of neural networks.