Skin lesion matching using Jacobian based descriptors for topology preservation

Developing an automatic system for tracking pigmented skin lesions (PSL) is important for early skin cancer detection. PSL tracking is typically formulated as finding correspondences in an optimization framework. When lesions are close to each other, it becomes likely to mis-correspond a PSL with nearby ones. To alleviate this problem, we introduce a novel topology-preservation concept based on which we derive different energy terms for the PSL matching objective function. Given two graph-represented sets of PSLs to correspond, in addition to unary and binary energy terms, we quantify topological-preservation using a Jacobian-like matrix capturing how triplets of vertices in one graph match another triplet in the other graph. We validate the goodness of our proposed descriptors on 97 pairs of human back digital color images. The results indicate that our new descriptors improve the PSL matching accuracy by over 9%.