Skin lesion tracking using structured graphical models.

An automatic pigmented skin lesions tracking system, which is important for early skin cancer detection, is proposed in this work. The input to the system is a pair of skin back images of the same subject captured at different times. The output is the correspondence (matching) between the detected lesions and the identification of newly appearing and disappearing ones. First, a set of anatomical landmarks are detected using a pictorial structure algorithm. The lesions that are located within the polygon defined by the landmarks are identified and their anatomical spatial contexts are encoded by the landmarks. Then, these lesions are matched by labeling an association graph using a tensor-based algorithm. A structured support vector machine is employed to learn all free parameters in the aforementioned steps. An adaptive learning approach (on-the-fly vs offline learning) is applied to set the parameters of the matching objective function using the estimated error of the detected landmarks. The effectiveness of the different steps in our framework is validated on 194 skin back images (97 pairs).