Diabetic Retinopathy Assessment Based On Multi-feature Fusion

Diabetic retinopathy is a kind of fundus disease with specific changes.It is mainly caused by diabetic damage to the retinal vascular wall,resulting in retinal microaneurysms,bleeding points and other lesions.It is one of the serious complications caused by diabetes.Diabetic retinopathy is the main cause of visual loss in diabetic patients.At present,with the change of life style,the number of diabetic patients increases rapidly,resulting in more and more diabetic retinopathy patients and a large number of workers blinded.Diabetic retinopathy is generally divided into two main stages:non-proliferative retinopathy and proliferative retinopathy.If the patient can be diagnosed and treated without delay during the period of non-proliferative lesions,the probability of blindness can be greatly reduced.At present,ophthalmologists mainly use color fundus camera to get fundus images and make diagnosis by observing abnormal lesions in fundus images.However,the number of professional ophthalmologists is far from enough to meet the diagnosis needs of patients because of the variety and form of diseases that may appear in the fundus.Therefore,it is very important to employ some new techniques to improve the diagnostic efficiency of glycosycoid lesions.Based on the theory and technology of machine learning and image analysis,this paper aims to study the automatic diagnosis of diabetic retinopathy.The main contributions are listed as follows:(1)Select and improve the existing deep network framework so as to make it more suitable for the specific diabetic retinopathy assessment.In addition,we combine the image datasets from other domains and the diabetic retinal image dataset to train the classification model so as to alleviate the problem of lack of labelled diabetic retinal images.(2)Propose two new multi-feature fusion methods for disease assessment of fundus images.According to the characteristics of fundus images and the distribution of data sets,vwe cooperatively employ the complementary information from both eyes to improve the diagnostic accuracy of the patients.In addition,an individual transformation strategy is further conducted to improve the generalization ability of the model and the classification accuracy of a single image.The experimental results show that the multi-fusion methods significantly improve the correctness of fundus image diagnosis.(3)Proposed an attention mechanism to evaluate the condition of fundus images.By adding different weights to different regions,the features which are helpful to classification are strengthened in the network,while the features that are not helpful to classification are weakened.As a result,the network classification performance can be improved,and the obvious lesion area can be visualized.The experimental results show that adding the attention mechanism to the basic classification network can improve the classification accuracy,and the location of the lesion can be determined with the help of the visual image of the attention mechanism.