Study On The Morphological Pattern Recognition Of Melanoma Based On Dermoscopy Image

Melanoma is a clinical often encountered in a malignant skin tumor, but also one of the world’s fastest growing cancer. Dermatologists usually perform early screening and diagnosis of melanoma by visual observation and histopathological biopsy. Blind biopsy often causes economic stress and unnecessary physical trauma to the patient. Moreover,the exact range of biopsy surgery is often difficult to determine before a definite tumor is benign or malignant. Therefore, noninvasive melanoma automatic diagnosis technology has become an urgent problem to be solved in the medical profession.Pattern classification of dermoscopy images is a challenging task of differentiating between benign melanocytic lesions and melanomas. Therefore, for the multi-morphological pattern of dermoscopy images, this thesis has carried on the thorough research to the malignant melanoma morphological pattern recognition based on multi-label learning. The innovation of this thesis mainly includes the following parts:1. Studies the melanoma segmentation and feature extraction methods. In order to better extract the characteristics of the skin mirror image, a fusion algorithm based on region consistency is proposed to segment the image, and the results of multiple segmentation algorithms are combined to obtain the reference lesion area. According to the regional size, gray value, texture The principle of consistency removes the sub-regions that are contradictory to the reference area, resulting in the final segmentation result. The color characteristics,shape features and texture features of the lesion area,the lesion area and the lesion area were extracted from the segmentation results2. Studies the application of multi-label learning based on manual feature extraction in melanoma pattern recognition. The morphological patterns of melanoma are analyzed and analyzed. There are eight main types of morphological features which can be defined,which involve seven basic models and one multi-component model. The seven basic modes include: mesh pattern, spherical pattern, pebble pattern, starburst mode, parallel mode, cavity pattern, uniform pattern. The seven basic models of melanoma are used to establish a multi-label classification model to achieve the purpose of automatically identifying the pattern categories contained in the skin mirror images. The Binary Reference algorithm and the ML-kNN algorithm are used to classify the eigenvectors.The study of the multi-label classification of the two algorithms shows that the ML-kNN algorithm has a better classification effect on the morphological pattern of melanoma than the Binary Reference algorithm.4. Studies the application of multi-label learning based on feature learning in melanoma pattern recognition. On the basis of the depth learning framework, an improved method is proposed to realize the multi-label classification. The image data and the multi-tag data are used as the input layer of the network respectively. Then, by adding the Slice layer to the multi-label classification in the network structure, A multi-label classification model is obtained. In the experiments using the convolution neural network to classify the morphological patterns of melanoma, the experimental results show that our proposed multi-label classification using convolution neural network has a significant improvement over the multi-tag classification based on manual characteristics.