Contribution of a classifier of skin lesions to the dermatologist's decision

Abstract

In this paper, we investigated to what extent the melanoma diagnosis can be impacted by an automatic system using dermoscopic images of pigmented skin lesions. Nine dermatologists were asked to give their diagnosis about 1097 dermoscopic images of skin lesions, including 88 histopathologically confirmed melanomas. The automatic diagnosis of black tumors was based on Local Binary Patterns (LBP) without segmentation of the dermoscopic images. The classification was performed using a simple linear support vector machines (SVM). The classifier showed a comparable performance with respect to dermatologists (AUC: 0.85). It appeared that a fusion of dermatologist's diagnosis with the automatic diagnosis improves the overall performances. We proposed a simple fusion strategy (highest-risk approach) with the automatic diagnosis, which improves the dermatologists' daily practice performance.