Clinically inspired analysis of dermoscopy images using a generative model

Dermatologists often prefer clinically oriented Computer Aided Diagnosis (CAD) Systems that provide medical justifications for the estimated diagnosis. The development of such systems is hampered by the lack of detailed image annotations (medical labels and segmentations of the associated regions). In most cases, we only have access to weakly annotated images (text labels) that are not sufficient to learn proper models. In this work we address this issue and propose a CAD System that uses medically inspired color information to diagnose skin lesions. We deal with the weakly annotated dermoscopy images using the Correspondence-LDA algorithm to learn a probabilistic model. The algorithm is applied with success to the identification of relevant colors in dermoscopy images, obtaining an average Precision of 83.8% and a Recall of 89.8%. The proposed color representation is then used to classify skin lesions, resulting in a Sensitivity of 77.6% and Specificity of 73.0% using Random Forests, and a Sensitivity of 75.1% and Specificity of 77.5% using SVM. These results comparable favorably with related works.