Color Detection in Dermoscopy Images Based on Scarce Annotations

Dermatologists often prefer clinically oriented Computer Aided Diagnosis (CAD) Systems. However, the development of such systems is not straightforward due to lack of detailed image annotations (medical labels and segmentation of their corresponding regions). Most of the times we only have access to medical labels that are not sufficient to learn proper models. In this study, we address this issue using the Correspondence-LDA algorithm. The algorithm is applied with success to the identification identification of relevant colors in dermoscopy images, obtaining a precision of 82.1 % and a recall of 90.4 %.