Study Of Algorithms For Segmenting Dermoscopy Images

Malignant melanoma is among the most rapidly increasing cancers in the world. Skin lesion segmentation, also known as skin lesion border detection, is the important step of computer automatic diagnosis based on dermoscopy images. However, existing single approach can roughly segment melanoma and lack robustness in the face of skin lesions varying in size, color, texture, and structure, and even lose effictiveness of segmentation at times. Therefore, it is limited that computer automatic diagnosis based on dermoscopy images is applied to clinical medicine.For purpose of figuring out above-mentioned technological difficulties in automatic diagnosis of dermoscopy image, the thesis conducts thorough researches and explorations on preprocessing, segmentation and post-processing. The main research content is as follows:1. A set of preprocessing algorithms are researched and implemented, including black frame removal, bubble removal, hair removal, contrast enhancement, seletion of optimal color component, and so on. Among these methods, a hair removal algorithm is proposed, which has the capabilities of locating and replacing these hair pixels, and repairing orginal skin region coverd by hair previously. Consequently, preprocessing lays a solid foundation for segmenting skin lesion accurately.2. Two fusion algorithms are researched and proposed, in order for combining multiple lesion segmentaion methods effectively. The first proposed approach is the fusion method based on region consistency, which aims at removing the results which don’t agree with the skin lesion base and forms a uniform lesion region in the end according to region area, intensity and texture consistencies. The second proposed approach is the generalized fusion method based on Markov random field, which brings color image and non-thresholing alogrithms into the fusion framework, makes full use of neighborhood and color information to fuse different segmentation results, and achieves the optimal output by means of iterated conditional mode.3. A set of post-processing algorithms are researched and implemented, including pixel labeling, subregion merging, isolated region removal, filling hole, border smoothing and expansion, and so on. Post-processing is capable to shrink the discrepancy between automatic lesion segmentation results and manual borders delineated by experienced dermatologists. Among these methods, a superpixel based greedy merging algorithm is proposed, which brings superpixel into border expansion of skin lesion and obtains the local minimum of skin color consistency through greedy method.4. Compared to manual borders delineated by three experienced dermatologists, error statistics of segmentation results are calculated at the foundation of evaluation criterion. Experimental results demonstrate that the proposed segmentation method has higher accuracy rate in lesion detection and lower error rate in overall detection than state-of-the-art methods, respectively.