Uncertainty-based feature learning for skin lesion matching using a high order MRF optimization framework.

We formulate the pigmented-skin-lesion (PSL) matching problem as a relaxed labeling of an association graph. In this graph labeling problem, each node represents a mapping between a PSL from one image to a PSL in the second image and the optimal labels are those optimizing a high order Markov Random Field energy (MRF). The energy is made up of unary, binary, and ternary energy terms capturing the likelihood of matching between the points, edges, and cliques of two graphs representing the spatial distribution of the two PSL sets. Following an exploration of various MRF energy terms, we propose a novel entropy energy term encouraging solutions with low uncertainty. By interpreting the relaxed labeling as a measure of confidence, we further leverage the high confidence matching to sequentially constrain the learnt objective function defined on the association graph. We evaluate our method on a large set of synthetic data as well as 56 pairs of real dermatological images. Our proposed method compares favorably with the state-of-the-art.