Research On Analysis And Localization Algorithm Of Diabetic Retinal Images Based On Deep Learning

Diabetic retinopathy,which is one of the complications caused by diabetes,is a major blindness-causing disease.However,the mismatch between the number of ophthalmologists and the number of diabetic patients has resulted in a large number of patients with diabetic retinal disease not being diagnosed and treated in a timely manner,leading to an exacerbation of their condition.In this paper,in order to solve the problems of diabetic retinal image lesion class classification,lesion optic disc detection and lesion point segmentation and localization,automatic control algorithm model is designed to identify fundus images autonomously,which can reduce the workload of ophthalmologists.The following is the main research content:Aiming at the problem of diabetic retinal image classification,an improved VGG16 network model based on deep migration learning is designed.The classification of fundus image lesion classes is mainly a two-class classification experiment for the presence or absence of lesions and a five-class classification experiment for lesion classes on the preprocessed dataset.The VGG16 algorithm model is improved by optimizing the structure of the full connection,reducing the overall parameters of the network,and designing the best combination of parameters for the full connection layer to reduce the resource consumption and improve the classification accuracy.The experimental results showed that the improved network can effectively improve the classification accuracy and has good classification effect.For the object detection problem of diabetic retinopathy images,the object detection experiment based on the improved Faster-RCNN network for diabetic retinal images was designed.Firstly,an improved VGG16 model and Res Net50 model are selected for the backbone feature extraction network,and the Res Net50 model is improved to reduce the number of network parameters,and finally the Faster-RCNN network accuracy is improved by using the cascade detector.Compare the differences in accuracy,F1 values and mean average precision of the algorithm Faster-RCNN under the feature extraction network of improved VGG16 and improved Res Net50.The experimental results showed that the cascaded Faster-RCNN network model with the improved Res Net50 as the feature extraction network detected more effectively.Aiming at the problem of lesion point segmentation and localization of diabetic retinal proliferation lesion images,an improved Mask-RCNN algorithm-based lesion fundus image segmentation and localization model is designed.An improved non-maximal value suppression algorithm is used to effectively improve the error detection and omission phenomenon of Mask-RCNN algorithm model for small target object segmentation and localization,comparing the differences in accuracy,recall and F1 value between the Mask-RCNN algorithm model before and after the improvement.The experimental results showed that the improved Mask-RCNN algorithm model performed effectively on the segmentation and localization results of lesion images lesions.Finally,the GUI graphical user interface is designed to make the algorithm classification and segmentation results visualized.