Research On Image Segmentation And Recognition Of Chronic Wound Image Based On Deep Neural Network

Accurate medical care is an important application of deep learning in the medical field.The lesions are processed by a deep learning method to obtain accurate processing results.These processed results can assist the doctor in the diagnosis and treatment of the patient's condition.Chronic wounds are typical chronic diseases that not only take up a lot of medical resources,but are also hard to be cured in the short term.This causes tremendous physical and mental damage to the patient and increases the workload of the doctors.If accurate medical treatment of chronic wounds can be achieved,the pain of patients can be alleviated,the waste of medical resources can be reduced,and the workload of doctors can be reduced.In order to achieve this goal,accurate segmentation and identification of chronic wound areas is required.In order to achieve accurate segmentation and identification of chronic wound areas,a systematic methods for segmentation and identification of chronic wound images based on deep neural network was proposed.The specific work can be summarized as the following three aspects:(1)The chronic wound image generator based on GANs was proposed.Insufficient number of chronic wound image datasets will lead to low segmentation and identification accuracy.To solve this problem,this paper proposes the chronic wound image generator based on GANs,which contains two deep neural networks: a generator network D that generates chronic wound images and a discriminator network G that discriminates whether it is a generated chronic wound image.Two network adversarial training,when the balance is reached,the new image of the chronic wound image is generated.These results can be used to augment the training set of subsequent segmentation and identification networks.Experimental results show that this method can achieve the best results when combined with traditional image data augment methods.(2)Chronic wound image segmentation and identification network U-net-Connected was proposed based on U-net.To solve the problem of image information loss caused by U-net downsampling,U-net-Connected connects the upper layer network information with the lower layer network information and integrating the multi-level feature map at the output.At the same time,with the advantage of the atrous convolution operation,thereby the accuracy of chronic wound image segmentation and identification was greatly improved.The experimental results show that the accuracy of U-net-Connected in the chronic wound image segmentation and identification task is indeed higher than other image segmentation and identification networks.(3)A post-processing method of chronic wound image segmentation and identification based on fully connected CRF was proposed.This post-processing method combines multi-scale score map with fully-connected CRF algorithm,overcomes the difficulty of insufficient information when single score map serves as input for fully connected CRF algorithm,solves the problem that some results of segmentation and identification of chronic wound image by U-net-Connected are in large error.The experimental results show that the proposed post-processing method of segmentation and identification of chronic wound images can solve the problem of insufficient segmentation accuracy or segmentation errors in partial results of chronic wound images segmentation and identification by U-net-Connected.