Consensus Clustering From Experts' Partitions for Patients' nevi: Model the Ugly Duckling

Ugly duckling (UD) concept is widely admitted as major criteria for malignant melanoma (MM) detection. It demonstrates that nevi in the same individual tend to resemble one another and human brain can identify a few similarity clusters, which include all nevi sharing the same morphological features. UD is the nevus that does not fit into any of the clusters of the individual, likely to be suspicious (i.e. MM). Our research program was to modeling the perceived similarity clusters (PSC) as a basis for a computer-aided diagnosis system. In the present part of the study, nine observers (dermatologists) participated to do the clustering of nevi and to identify the UDs using dermoscopic images of nevi of 80 individuals (2089 nevi with 7 histologically confirmed melanomas). Observers identified all confirmed melanomas as UDs (100% sensitivity and 98% specificity). The observers tend to be concordant about the identification of PSCs. We combined the multiple clusterings of experts to find the consensus clustering, which yields a stable and robust final clustering that is in agreement with the multiple clusterings of experts. We demonstrated the limited variability of nevi patterns per individual (2.45 PSCs in average), whatever the number of nevi, which human brain has a natural intuitive ability to perceive.