Feature Explainability and Enhancement for Skin Lesion Image Analysis

Deep learning based methodologies, especially convolutional neural networks, recently contributed to the field of medical image analysis. Diseases are recognized by passing digital images into end-to-end networks. However, identifying visual patterns making significant effects on prediction results is recently a challenge for scientists. A remarkable number of researches, in the field of Explainable AI, are conducted, including data explainability, model explainability, and post-hoc explainability. In the research, the authors utilize a deconvolutional neural network to reconstruct original skin lesion images from learned feature vectors to emphasize visual patterns mostly contributing to recognition results. Additionally, a feature enhancement technique based on blending is proposed to verify the effect of learned features. Experiments conducted in the HAM10000 data set show that the proposed methodology is promising for data explainability.