Detection of Melanoma in Dermoscopy Images-Using Local Binary Patterns

Abstract

This thesis is motivated by the potential gains that can be achieved by the use of computer assisted decision systems (CAD) for diagnosis of melanoma in the skin using dermoscopy. A CAD system provides quantitative and objective evaluation of the skin lesion versus the subjective clinical assessment. It automates the skin lesion analysis, and reduces the amount of repetitive and tedious tasks to be done by physicians. This research is mainly focused on the computer vision perspective to design a CAD system which will facilitate the physicians. A complete pattern recognition system that includes three vital stages to conform the analysis of skin lesions by the clinicians: segmentation, feature extraction and classification. The data-set contains images and annotations provided by physicians. Segmentation is an imperative preprocessing step for CAD system of skin lesions. Segmentation is performed using active contours with creasness features. Feature extraction of segmented skin lesions is a pivotal step for implementing accurate decision support systems. Physicians are interested in examining a specific clinically significant region in a lesion. Such a region is expected to have more information in the form of texture that can be relevant for detection. In case of detection of melanoma various local features for example pigment network and streaks usually occur in peripheral region of the lesion. This led to the extraction of peripheral part for feature extraction instead of whole lesion processing. We propose novel techniques for feature extraction on peripheral part of the lesion using joint histogram of multiresolution Local Binary Pattern along with the contrast of the patterns. Classification results obtained from the proposed feature matrix were compared with some other texture descriptors, showing the superiority of our proposed descriptor.