Research Of Diabetic Retinal Image Classification Based On Deep Learning

Diabetic retinopathy has become one of the major causes of blindness in the right age population due to the variety of lesions,unobvious early symptoms and high incidence.Regular ocular examinations are the key to prevent the deterioration and blindness of the disease.Because of the large number of patients and the shortage of medical resources,it is a great significance to study an intelligent automatic diagnosis system of diabetic retinopathy which can assist the doctors to improve the diagnostic efficiency.Image quality assessment and the grade classification of diabetic retinopathy,as the two important parts of automatic diagnosis system,have significant impact on subsequent disease diagnosis.With the development of artificial intelligence algorithms,deep learning has been widely used in medical image and other fields.The application of deep learning to the image quality assessment and diagnosis of diabetic retinopathy has important significance for improving the performance of automatic diagnosis system and is also the main content of this paper.The main work and innovation of this paper are summarized as follows:Firstly,an image quality assessment method based on human visual system is proposed.The proposed algorithm imitates the working of the human visual system to combine supervised information from convolutional neural networks and unsupervised information from saliency map,and then constructs a classifier for retinal image quality classification.The saliency map selected in this paper is with full resolution image information to extract local and global features related to retinal image quality.Meanwhile,we use the fine-tuning method based on transfer learning to train the convolutional neural network for the supervised features extraction.The experiment results have the best classification performance compared with other methods.Secondly,a deep learning classification method based on attention mechanism is proposed.At first,an attention network based the fully convolutional neural network is embedded in the deep network,and uses the data with the lesion areas ground truth marked by experts to train the attention network.The attention network can generate lesion candidate regions and introduce expert knowledge.Next,the deep network we used is based on the inception architecture and residual architecture.And a feature enhancement method is used in the training of the whole network to introduce the information of lesion candidate regions into the classification of diabetic retinopathy,which is helpful for the whole network to preserve the original feature information and to enhance the feature information on the lesion candidate regions.The proposed method is a multi-task learning that can roughly locate thelesion areas of fundus image while using the expert knowledge to improve the classification performance.Experimental results show that our method is superior to other methods on classification performance,and it is medically interpretable.