Study On The Recognition Of Skin Lesion Region Based On Dermoscopy Images

With the development of computer technology,the combination of medical diagnosis and computer technology greatly improved the accuracy of human disease diagnosis.Skin cancer is the fastest growing cancer in the world.Early diagnosis and treatment of skin cancer is very important for the treatment of skin cancer.The classification of dermoscopy image with traditional machine learning methods has achieved remarkable achievements in the clinical diagnosis of melanoma.In recent years,deep learning technology has been developing continuously.Image applications based on convolution neural networks have achieved good results in many fields.Based on the in-depth study of the development of deep learning technology,this thesis presents an application of deep learning technology in classification task of skin lesions with small data,based on two major methods: supervised learning and semi supervised learning.The purpose of those methods is to classify the skin lesions of melanoma,nevus and seborrheic keratosis into three classes.Melanoma represents malignant lesion.Nevus and seborrheic keratosis represent benign lesion.Based on the purpose of the research above,three innovations are proposed as follow:1.Proposes a dermoscopy image preprocessing method based on deep learning technology.After the study of the application of deep learning technology in target recognition field,we analyze the evolution of RCNN technology,Fast RCNN,Faster RCNN and Mask RCNN model,and use the Mask RCNN model to segment the skin lesions apart from the original image and cut segmented image into a small size.Before using the images to classification,we also used some traditional CNN preprocessing methods like image resize,image normalization and data augmentation to deal with the images.2.Proposes an image classification method of skin lesions based on improved deep convolution neural network.In this section,we first study the development of convolution neural networks,and choose the perfect model with three types of network model,VGG19,ResNet50 and InceptionV3.By comparing the results of the experiment,we selected ResNet as the classification model.Secondly,by comparing ResNet Imagenet pretrained model with different layers,we find a suitable layer structure for ResNet with this data set and then optimize the Res Net model.Finally,by consideringthe sample bias between the different data set,an improved ResNet model with MMD is proposed.It can increase the classification performance of the network in different data set by reducing the bias of the sample feature distribution extracted by model.3.Proposes some semi supervised learning algorithms for the classification task of skin lesions based on GAN.This part is mainly considered to enhance classification performance by using some unlabeled dermoscopy images.It mainly includes two parts,which use GAN as feature extraction that combined with some kind of clustering algorithms to classification and the structure modified GAN model with multi-task.Among them,the clustering algorithms compared the effects of different types of similarity measure methods on the classification results,including Euclidean distance and RBF kernel function.And the GAN multi-task model with improved discriminator has added a structure of classifier which shared weighs with discriminator.