Research On Dermatoscope Image Recognition Method Based On Deep Learning

With the popularization of dermatoscope equipment,dermatoscope images are increasing day by day.According to statistics,the gap of Chinese dermatologists has reached hundreds of thousands,and there are many types of skin pigmented lesions.Without the guidance of professional knowledge,it is difficult for patients to make a preliminary self-diagnosis.The user of dermatoscope needs to have certain operation and clinical experience in order to obtain better results.At this stage,it is time-consuming,cumbersome and subjective to rely on the shortage of dermatologists to manually screen dermatoscopic images to diagnose diseases.Therefore,the research on dermoscopic image classification methods is imminent.In order to solve the above problems,this thesis proposes a dermoscopic image recognition method based on deep learning,which includes two parts: preprocessing and model building.In the preprocessing stage of the dermoscopic image,the focus area was extracted first.In order to eliminate the ambient light noise in the dermoscopic image,the color constancy algorithm was used to eliminate it.Taking into account the insufficiency of image data,the original data is augmented by the combination of image sampling and GAN adversarial neural network.Aiming at the problem of dermoscopic image recognition,on the basis of transfer learning,this thesis selects Vgg Net and Res Net50 convolutional neural network as the basic model framework,and adds GAP,BN,Regularization and Dropout to the over-fitting problem in the model.Relevant network layer;in view of the insufficient number of dermoscopic images and the imbalance of image data among various diseases,a model optimization method with a weighted cross-entropy loss function is adopted.This method can not only optimize the over-fitting problem well,but also And it further improves the accuracy of the model.Finally,the idea of hierarchical convolutional neural network is adopted,which greatly improves the accuracy of dermoscopic image recognition.Aiming at the problem of ambient light,first use shades of grey algorithm to eliminate ambient light,and use SSIM algorithm to evaluate image quality,then build a convolutional neural network based on Res Net50 and Vgg Net,and perform overfitting optimization on the basis of migration learning,and then build a hierarchical convolutional neural network.The experimental results show that compared with the Vgg Net and Res Net50 deep learning models,on the ISIC2018 data set,the improved deep learning model is more suitable for the recognition task of dermoscopy images.The classification accuracy rate of the Res Net 50 model reached85.94%,and the classification accuracy rate of the Vgg Net model reached 83.096%.Compared with the unimproved classification models Res Net 50 and Vgg Net,the test accuracy rate increased by 5.752% and 6.249%,respectively.