Study On The Texture Classification Of Melanoma Based On Dermoscopy Images

Melanoma is a harmful disease,which is difficult to cure under the current medical level.Therefore,it is important to treat it at the early stage of this disease.Clinicians used to diagnose melanoma by artificial judging of dermoscopy images.However,it is difficult to make an accurate judgment because the melanoma has no obvious early symptoms,it is necessary to use computer technology to help clinicians.Aiming at a variety of texture feature patterns of melanoma dermoscopy images,this thesis uses superpixel segmentation,texture feature extraction and analysis,sparse coding of texture features,support vector machine for classification and convolutional neural network to solve the task which is to classify the texture pattern of melanoma.The main contents are as follows:1.Studies the texture patterns of melanoma.Melanoma can be divided into four classes according to their texture features.Each melanoma dermoscopy image may contain one or more of these texture patterns.Currently,most of these classification algorithms only give the result of whether there is such a texture pattern in the image,but do not identify the area where the texture pattern is located in the image.To address these deficiencies,this thesis introduces the superpixel segmentation algorithm,sparse coding and convolutional neural network for the task of texture classification.2.Studies the superpixel segmentation algorithms for melanoma dermoscopy images.This thesis uses SLICO algorithm to segment superpixel images.SLICO algorithm is an improvement of the SLIC algorithm.The shape of segmented superpixel is regular,and the dimension is homogenous.Furthermore,each superpixel area preserved the effective information texture feature and the boundary information of the skin lesions,which can effectively reduce the influence of the irrelevant factors on the texture classification.3.Studies the texture analysis of superpixel.In this thesis,several texture features such as gray level co-occurrence matrix,local binary pattern and Gabor filters are extracted based on superpixel.Then the features are sparsely coded,and each superpixel is classified by using a support vector machine to obtain its own texture pattern.4.Studies the texture classification by using convolution neural network.This thesis uses a variety of convolution neural network models to automatically extractfeatures and classify textures from superpixel images.The proposed VGGNet,ResNet and GoogLeNet Xception models are studied,and the final classification results of different convolution neural networks are verified and compared.This thesis proposes two algorithms for texture classification of melanoma,respectively,one of them is based on sparse coding or group sparse coding,and another one is based on convolutional neural network.After several experiments,the results show that these two classification algorithms can achieve great accuracy.