Collaborative Human-AI (CHAI): Evidence-Based Interpretable Melanoma Classification in Dermoscopic Images

Automated dermoscopic image analysis has witnessed rapid growth in diagnostic\nperformance. Yet adoption faces resistance, in part, because no evidence is\nprovided to support decisions. In this work, an approach for evidence-based\nclassification is presented. A feature embedding is learned with CNNs,\ntriplet-loss, and global average pooling, and used to classify via kNN search.\nEvidence is provided as both the discovered neighbors, as well as localized\nimage regions most relevant to measuring distance between query and neighbors.\nTo ensure that results are relevant in terms of both label accuracy and human\nvisual similarity for any skill level, a novel hierarchical triplet logic is\nimplemented to jointly learn an embedding according to disease labels and\nnon-expert similarity. Results are improved over baselines trained on disease\nlabels alone, as well as standard multiclass loss. Quantitative relevance of\nresults, according to non-expert similarity, as well as localized image\nregions, are also significantly improved.\n