Concept-based Representation for Feature Explainability in Skin Lesion Classification

Deep learning has been robustly developing and playing an important role in various fields, such as medical image analysis, machine translation, media generation, etc.However, neural networks are still black-boxes thus preventing them from high-stakes applications, especially medicine.In the study, the authors first propose a methodology to identify learned concepts in the feature space and hence it is possible to reason the network predictions based on similar concepts or samples.Secondly, a novel approach for the feature presentation, based on concept distances, is introduced so that the informative and discriminative levels of learned features can be evaluated by performance differences among the pre-trained model and baseline classifiers.Experiments are conducted in the HAM10000 dataset of skin lesion images and prove the power of reasonability of the proposed methodology.The study directly contributes to the development of explainable artificial intelligence in computer vision, especially medical image classification.