Research On Diabetic Retinopathy Screening Based On Deep Learning

Diabetic retinopathy（DR）is a common diabetic chronic complication caused by diabetes,which is also the main cause of vision loss and blindness in adults.Early diagnosis and timely treatment are of great significance to delay the deterioration of the disease and reduce the loss of vision.The automatic screening of diabetic retinopathy is of great significance for timely diagnosis,shortening the time consuming and reducing the dependence on manual screening.DR grading and segmentation of lesions are two main tasks in automatic screening.This thesis mainly adopts the deep learning model,and combines the methods of image processing and machine learning to explore the above two tasks:（1）In the segmentation of diabetes retinopathy,there are many problems,such as the inability to detect small lesion areas and the difficulty to distinguish similar disease areas.During the down sampling process of deep neural network,Low dimensional information extracting will discard information at the mean while.It is easier for the model to ignore the small lesion area,especially the small plaque like bleeding in the early stage of the disease,which only accounts for about 0.1% of the relative background in the image.The loss of information will also make it difficult for the model to distinguish similar diseases,resulting in the degradation of segmentation performance.To solve this problem,this thesis expands the U-Net and adds the decoder to output the disease segmentation and the corresponding edge segmentation results respectively.An effective end-to-end image segmentation framework with two branches is proposed.In order to optimize the detection effect of subtle diseases,dice loss is used as the loss function and the results of the sub network on the two branches are integrated to improve the segmentation performance of the network.The experimental results show that the introduction of auxiliary supervision information and the proposed network structure helps to improve the segmentation performance of the network for fine regions and achieve certain advantages in IDRiD,DDR and E-Ophthy datasets.（2）In the classification of diabetic retinopathy,it is intuitive to grade the level of DR according to the lesion on the image.Benefited from the sufficient data and the clear criteria,the classification of diabetic retinopathy has achieved considerable process.However,how to combine segmentation and classification features to optimize the performance of the model and provide a more convincing process for model decision-making are still a problem.Multi-task learning method is utilized to perform the above tasks,and the lesion segmentation features assist in DR grading.But the difference in the scale of classification and segmentation data sets caused by the difficulty of labeling is inevitable,which will have an unpredictable impact on the performance of the model.Therefore,Semi-supervised learning is also introduced in the training process.After training the model with lesion segmentation pixel-level annotation data,more segmentation predictions are generated as pseudo labels.In the end,the multi-task model is trained in an end-to-end manner to improve the performance of disease classification and lesion segmentation.