In order to handle the high diversity of cancer lesions, many computer vision approaches have been developed ad-hoc for a specific lesion type (e.g. low contrast, heterogeneous). However, due to the substantial diversity of the lesions’ characteristics, those methods do not properly represent even a whole single dataset. I introduced an adaptive generalizable framework of multi-lesion segmentation, applying exactly the same technique on different lesion datasets with excellent results (Figure below). The method automatically estimates the weighting parameters of the level set cost function and the adaptive local window size surrounding each contour point. Those are re-estimated over iterations of the segmentation process and for every lesion separately. Currently, I am working on developing a new energy functional that is based on adaptive selection of texture features and as a result, the cost function will fit much better to each tested data.