Tree-structure based framework for automated skin lesion analysis

Cutaneous malignant melanoma is one of the most frequent types of cancer in the world but if a malignant lesion is detected early, it can be cured without complication. Automated skin lesion analysis attempts to accomplish early detection of malignancy using digital dermoscopic images. We address two challenging applications in automated analysis of dermoscopic skin lesion images: lesion segmentation and lesion diagnosis, both of which use a novel tree structure based framework to model the radial and the vertical growth pattern of the skin lesion. To construct the tree, the pixels are repeatedly clustered into sub-images based on color information and spatial constraints. This framework allows us to extract features by looking at the tree from a graphical aspect, or a textural/geometrical aspect on the nodes. The features are used in supervised learning algorithms on datasets containing 116 challenging images for segmentation, and 410 images for diagnosis. Our method outperforms many other published results.