Research On Key Technology Of Assistant Diagnosis Of Diabetic Retinopathy Based On Fundus Images

Diabetic Retinopathy(DR)has become a major cause of vision loss and blindness in adults.It is one of the major syndromes of diabetes.In view of the fundus image of diabetic retinopathy,this dissertation has carried out the related research on lesion detection and lesion evaluation,and has made some progress.It mainly includes the following aspects:1.Aiming at various imaging quality problems such as inconsistent brightness,uneven illumination,low contrast and reflection in the process of fundus image acquisition,as well as adverse factors for subsequent detection,research on fundus image enhancement algorithms is carried out.At the same time,combined with the fundus structural characteristics of the retina itself,the segmentation algorithms under different structures and their relationship with lesion detection are analyzed,including the vessels segmentation method conducive to the detection of microaneurysm,the optic disc extraction method as the prerequisite for the detection of hard and soft exudation,and so on.This part of the research work on DR image preprocessing method has important significance for follow-up lesion detection and grading evaluation algorithm.2.According to the characteristics of small size and unified structure of microaneurysm in DR,this dissertation studies and designs a detection algorithm based on convolutional neural networks(CNN).Compared with various classical algorithms,this method can detect microaneurysm more accurately.At the same time,by reducing the number of candidate regions,the time overhead of the overall algorithm is further reduced and the operation efficiency is improved.At the same time,this dissertation proposes a method for detecting candidate areas of microaneurysm based on morphological reconstruction,which can also be used to extract candidate areas of other retinal lesions.3.In view of the complexity of the distribution of hard exudation structure of DR and the difficulty of accurate description and description of traditional machine vision features,this dissertation studies and implements the integrated detection method integrating depth features and traditional manual features,which solves the problem of hard exudation detection of fundus image and greatly improves the detection performance.The experimental results show that the fusion depth feature makes the model have stronger feature expression ability and robustness,and the F-Score value on dataset eoptha increases from 0.8097 to 0.8928,and the F-score value on dataset HEI-MED increases from 0.5461 to 0.9318.4.Aiming at the problem of soft exudation detection in DR,this dissertation studies and implements an integrated detection method of soft exudation based on multi feature cascade.The performance of the model is improved by fusing depth features.At the same time,the neural disc segmentation method based on super-pixel segmentation is used in the preprocessing process,which eliminates a large number of non soft regions,greatly reduces the time overhead of the detection algorithm and improves the detection efficiency.5.Finally,aiming at the difficulty of accurately extracting key descriptors and semantic information of retinal lesion images,a new automatic classification method based on deep learning is presented.At the same time,image preprocessing,label coding and transfer learning are adopted to greatly improve the performance of the algorithm,which lays a foundation for the follow-up clinical application and technical promotion of the research results.The research results obtained in this dissertation have been tested in Sichuan Provincial People's Hospital,and achieved preliminary application results on the actual data set.