| Diabetic retinopathy is one of the four major causes of blindness in the world.Long-term hyperglycemia can lead to diabetic retinopathy.Early diabetic patients may have no obvious symptoms,but with the development of the disease,patients may gradually experience decreased vision,and in severe cases,they may even be permanently blinded,so early screening of diabetic patients is particularly important.However,traditional color fundus images have many problems in ophthalmologists diagnosis,they are gradually disappearing in clinical.The ultra-wide-angle fundus images have the advantage of wide field of view,simple and fast operation,non-mydriatic,non-invasive,high-resolution,etc.So in this thesis,a diabetic retinopathy staging algorithm model based on lesion detection is designed and implemented,in which,the ultra-wide-angle fundus images of diabetic retinopathy,combined with deep learning,image processing and computer vision technologies are used.Now,this model is used in the diabetic retinopathy intelligent diagnosis system of a hospital in Chengdu for practical use,which can improve the efficiency and accuracy of ophthalmologists’ diagnosis of diabetic retinopathy patients.The main contents of this thesis are as follows:(1)On the basis of the characteristics of the fundus images of stage 4 and non-stage 4 of diabetic retinopathy,this thesis used the improved Res Net50 for staging,replaced the size of the convolution kernel of Res Net50,and introduces the CBAM attention mechanism.The improved model improves the precision of four-stage images of diabetic retinopathy.(2)Aiming at the problems of diabetic retinopathy stage 0 to 3 ultra-wide-angle fundus images,the target is small and the image resolution is too large,this thesis proposed a method based on improved Faster RCNN and image segmentation for ultra-wide-angle image lesions of diabetic retinopathy(hemorrhages,microaneurysms,cotton wool spots,hard exudation)were detected.In the feature extraction stage,the residual network combined with the feature pyramid is used to extract the lesion information,the size of the candidate frame area is reduced in the RPN network,and the soft non-maximum value suppression method is used for optimization to solve the problem that the features of small lesions are difficult to extract,the candidate box is too large and uncomfortable and the problem of small lesion detection.The lesion area of the image is segmented during model training,and the images are segmented and submitted for inspection during detection and then synthesized to solve the problem of distortion and loss of lesion information caused by image scaling.(3)Aiming at the problem that microaneurysms and hemorrhages are easy to be misdetected in ultra-wide-angle fundus images of diabetic retinopathy,this thesis first used Unet based on residual network to segment retinal blood vessels,and then combined the bleeding points based on blood vessel continuity proposed in this thesis.The lesion detection results of the improved Faster RCNN are modified with the microaneurysms detection algorithm,which improved the detection accuracy of microaneurysms and hemorrhages.After establishing the parameters and model of the final algorithm,the accuracy of the proposed method on the ultra-wide-angle fundus images of diabetic retinopathy in a hospital in Chengdu reached 83.8% for stage 0,86.5%for stage 1,92.7% for stage 2,88.1% for stage 3,and 91%for stage 4.(4)Finally,after determining the algorithm model based on ultra-wide-angle fundus images,the intelligent diagnosis system for diabetic retinopathy was improved and perfected,and the algorithm model was integrated into a hospital system in Chengdu for practical application. |
PDF Full Text Request