Research On Skin Disease Image Segmentation Algorithm Based On Deep Learnin

Medical image segmentation is an important process of clinical treatment,the accuracy of segmentation results has an important impact on the subsequent tissue structure decomposition,disease analysis and operation.In recent years,medical image segmentation technology based on deep learning has become the main method of medical image segmentation,and the use of U-Net network for medical image segmentation has made remarkable progress.However,it is still very difficult to obtain high-quality segmentation mask for the effects of category imbalance and noise in dermatological images.In addition,the blurred boundary of skin lesions and low contrast with skin tissue are a problem worth studying.Based on the above problems,this paper focused on some key problems in skin disease images and combined with U-Net framework to carry out a series of studies.Firstly,in order to solve the difficulty of skin disease image segmentation caused by complex features,a skin disease image segmentation algorithm combining probabilistic graph model and deep learning model is proposed.The network adaptively reassigns weight to the features to improve the learning ability of the model.In addition,conditional random field module is introduced for segmentation post processing and segmentation results are refined,so as to improve the accuracy of segmentation results.By comparing the method with the existing high level segmentation network on different data sets,the best comprehensive segmentation effect is obtained.The experimental results show that the segmentation network proposed in this paper can effectively solve the problem of complex features in skin disease images.The highest segmentation accuracy is 0.924,and the Jaccard similarity is 6.9% higher than before the improvement.Secondly,aiming at the problems of noise interference and inherent induction bias of convolutional neural network in skin disease images,an improved conditional random field based skin disease image segmentation model is proposed.By adding variables that can measure the intensity of image pixel transformation,the problem of noise interference in skin disease images is solved.In the segmentation task with low contrast of target region and fuzzy boundary,an improved conditional random field module is proposed.This module can model the relationship between the target pixel and all remaining pixels,solve the inherent induction bias of convolutional neural network,and improve the segmentation accuracy of the model through global modeling.The highest segmentation accuracy is 0.957,which is 7.5% higher than before improvement.Finally,a series of segmentation experiments were carried out on ISIC2017 and ISIC2018 data sets,and the experimental results show that the proposed method can effectively complete the dermatological image segmentation task.In addition,compared with some advanced segmentation models,all of them are higher than the current advanced network,which proves the effectiveness of the proposed algorithm.