Research On The Key Techniques For Analysis And Diagnosis Of Fundus Lesion Images

Diabetic retinopathy（DR）and diabetic macular edema（DME）are eye diseases that harm the vision of diabetic patients for a long time.Timely detection and intervention in the early stage of the disease can effectively prevent the risk of blindness.Therefore,efficient and accurate automatic screening of fundus diseases is particularly important.In recent years,the DR or DME grading methods based on deep learning have achieved good classification accuracy,but most of them only focus on the grading of specific diseases and lack of generality.In addition,too many model parameters are not conducive to real-world applications.More importantly,most methods ignore the problem of lack of annotated medical data.In order to solve these problems,we have explored the problem of the lack of annotated data in small sample medical datasets from two perspectives:how to make full use of annotated data and how to generate additional data.To make the model more conducive to real-world applications,we propose an adaptive attention block（AAB）,which can be adjusted adaptively according to the grading tasks.From the perspective of making full use of annotated data to solve the lack of labeled data,we use the improved Mean Teacher semi-supervised learning model to train the proposed AABNet and achieve good results.We evaluated the network on Messidor dataset and DDR dataset,and achieved the best results in multiple tasks on Messidor dataset,surpassing other methods on DDR dataset.Our model has only 3.3M parameters,which is very beneficial for practical applications.From the perspective of data augmentation,we propose a fundus images generation model based on GAN.Specifically,we employ the least square Gan（LSGAN）as the basic architecture of the model,improve the network structure of its generator and discriminator,add the perceptual loss to the final loss,and use spectral normalization as the normalization method of the discriminator.Our model has two training stages.In the first stage,we train a network Ladder Net for vascular tree segmentation of fundus image.In the second stage,we concatenate the vascular tree mask,AOV mask,hard exudation mask,soft exudation mask,microvascular mask and hemorrhage mask as the input of the generator.The vascular tree mask and the real image are concatenated as positive samples,and the vascular tree image and the generated image are concatenated as negative samples to train the model.Our model can generate fundus images with pathological features by controlling the morphology of soft exudation,hard exudation,microaneurysms and hemorrhage,and fundus images containing these features can be created.Experimental results show that the fundus images generated by our method are clear and realistic and close to the data distribution of real images.