From Retinal Images To Diabetic Retinopathy Diagnosis

The diabetic retinopathy is the leading cause of blind,and timely diagnosis and treatment can redeem the lose of vision.Diagnosis of the diabetic retinopathy is to infer the degree of diabetic retinopathy according to the provided digital retinal images.The proposed approach in this paper is able to accurately segment the retinal vessels,and accomplish the diagnosis with the features provided by the segmentation.The main contribution of this paper are as follows:A patch-wise segmentation approach is proposed to solve the deficiency of samples.To maintain the small vessels,we design the size-invariant features,avoiding the considerable lose of detailed information.And to solve the deficiency of samples,a patchwise sampling based segmentation approach is proposed.Moreover,to solve the bar problem appearing at the boundrary of each patch,an overlapped patching approach is proposed.We design a fully convolutional network which segments the retinal images patch-wisely,and can accomplish computation on scalable computers.On international datasets,the proposed approach performs better than traditional approaches.A full image segmentation approach is proposed,which can segment the retinal image rapidly.We propose a weight evolving approach to solve the problem that the ratio of vessel pixels is much lower than background pixels,and accelerate the convergence of the network.To expand the receptive field,we introduce of atrous convolution,which can easily control the receptive field.The SegVessel network is proposed with three losses to cover the complex segmentation problem.And the experimental results demonstrate that the approach achieves 96.00% of accuracy and 77.46% of recall rate on the DRIVE dataset,and 95.17% of accuracy and 77.56% of recall rate on the CHASE_DB1 dataset,outperforming the previous state-of-art approaches.To solve the problem that the extraction of features demanded by diabetic retinopathy diagnosis is difficult,a deep convolutional network based on residual network is proposed,which takes the advantage of features extracted by segmentation task.And the experimental results demonstrate that the network indeed has learned the features needed by the diagnosis,such as vessels,lesion areas,from the visualized features.On the international public dataset,the proposed approach reaches 93.09% of accuracy,and 94.61% of recall rate,and there are some improvement.