Analysis Of Diabetic Retinal Fundus Image Based On Deep Learning

Diabetic retinopathy is one of the most serious and common complications of diabetes,manifested as microvascular lesions in the eye,which is the main cause of blindness.At present,the process of expert diagnosis relies on analyzing the retinal fundus image,finding out the related lesion and giving the diagnosis and corresponding treatment measures.However,due to the differences in medical conditions and low artificial efficiency,many patients miss the best diagnosis and treatment time,which leads to visual impairment.Therefore,it is of great significance to provide automatic retinal fundus image analysis algorithm and build a model for assistant medical diagnosis.In this paper,based on the deep learning method,the research and analysis of retinal fundus images mainly includes the following four aspects:(1)For deep learning,a large amount of labeled data is needed.But the acquisition of pathological data in reality is limited.Different data preprocessing methods are adopted for image classification and segmentation tasks,including image denoising,normalization,data augmentation and post-processing.Several steps are processed to complete the data cleaning process while effectively increasing the amount of available data.(2)The ConciseNet model for classifying the severity of retinal fundus image is proposed.The model is based on the component architecture of Inception and ResNet.The preprocessed images are directly input into the network.After feature extraction by the main stem,residual module and decrement module,the preliminary classification is completed by the auxiliary classification module,and the final output is obtained by fusing the classification results.(3)The DRUnet model for lesion segmentation of retinal fundus images is proposed.Input is overlapping sampled image blocks and artificial segmentation results.First,semantic information is obtained through the downsampling path composed of convolution layer and pooling layer.Then,the deconvolution layer in the upsampling path is restored to the original input size continuously.At the same time,it combines with the corresponding feature map in the downsampling path to complete the detection of lesion area by supplementing detailed information.(4)According to the physiological structure characteristics of the ocular blood vessels,the semantic segmentation network FC-DenseNet in the general field is modified,and a blood vessel segmentation network suitable for retinal fundus images is proposed.The network makes full use of image features to reduce pixel loss through dense connection and dilated convolution,and enlarges the image by transposed convolution and fuses the high-resolution image from the skipping structure to achieve precise segmentation of blood vessels.