Towards prevention and early diagnosis of skin cancer: computer-aided analysis of dermoscopy images

Melanoma, the deadliest form of skin cancer, must be diagnosed early for effective treatment. Irregular pigment network and streaks are important clues for melanoma diagnosis using dermoscopy images. This thesis describes novel image processing approaches for computer-aided pigment network and streaks detection on dermoscopy images. Our methods provide meaningful visualization of these structures, and extract features for irregularity detection. Additionally we present our efforts towards prevention of melanoma, by developing a smartphone app, UV-Canada, to raise awareness of the importance of using sunscreen to prevent melanoma. To locate pigment networks, after preprocessing steps, which include segmenting the lesion from the normal skin in the dermoscopy image, we use a graph-based approach to extract the holes and meshes of the pigment network, where cyclic subgraphs correspond to skin texture structures. Each correctly extracted subgraph has a node corresponding to a hole in the pigment network, and the image is classified according to the density ratio of the graph. Our results over a set of 500 dermoscopy images show an accuracy of 94.3% on classification of the images as pigment network Present or Absent. For analyzing the irregularity of the structure, we locate the network lines and define features inspired by the clinical definition to classify the network with an accuracy of 82% discriminating between Absent, Typical or Atypical, which is important for melanoma diagnosis. To find streaks in dermoscopy images, filters are applied, and in a similar fashion to fingerprint analysis, orientation estimation and correction is performed to detect low contrast and fuzzy streak lines. A graph representation is used to analyze the geometric pattern of valid streaks, to model their distribution and coverage. We achieved an accuracy of 77% for classifying dermoscopy images into streaks Absent, Regular, or Irregular on 945 images; the largest validation dataset published to date. Our contributions will improve automated diagnosis of melanoma using dermoscopy images