Study Of Malignant Melanoma Detection Based On Dermoscopy Images

Malignant melanoma spreads through metastasis and is one of the most rapidly increasing cancers in the world. Statistical data has pointed out that the majority of deaths resulting from skin cancer are as a result of melanoma. The survival rates in patients depend on the stage of the infection and early detection or clinical intervention make the chances of curing the disease higher. Diagnosis and treatment of melanoma is a challenging task since the processes are prone to misdiagnosis and inaccuracies due to doctors' subjectivity. Therefore, this thesis if focus on solving the classification problem of dermoscopy image by research the detection methods.In consideration of the situation that melanomas are asymmetrical and have irregular borders, notched edges, and color variations, this thesis mainly focus on image denoising, skin lesion segmentation, feature extraction and its classification. The main contents of the thesis are as follows:1. For the black frame, hair and bubbles happened during the acquisition process of dermoscopy image. We study a set of algorithms to removal black frame, hair and bubbles and achieve promising results by this framework of algorithms. Consequently, it lays a solid foundation for subsequent processing.2. Analysis of existing skin lesion segmentation algorithms, a segmentation method based the classification of super pixels is proposed to separate skin lesion from the background. Then extract the feature, such as colors, shapes and texture of the super pixels.3. In the step of lesion feature extraction, feature design and feature learning are considered. For feature design approach, this thesis design a set features of color, shape and texture according to the clinical diagnosis methods. For feature learning, this thesis focus on bag of word model and sparse coding model. In the bags of phrase model, BOW model is studied and in order to overcome the loss of space information and polysemy, a visual phrase model based on the spatial pyramid matching model is proposed. For sparse coding, dermoscopy image is code by the sparse coding algorithm and group sparse coding algorithm.4. A support vector machine classifier is used to handle the low-level feature, bags of phrase feature and group sparse coding feature. In order to take advantage of the complementarity of low-level feature, bags of phrase feature and group sparse coding feature, this thesis proposed a classifier fusion method based on multi-view mechanism. Experimental results show that this fusion method has better classification performance.