Deep Learning And Its Applications For Diabetic Retinopathy Screening

Diabetic retinopathy(DR)is a chronic complication of diabetes,and now it has become the leading cause of vision loss and blindness in the working population.The clinical practice shows that timely diagnosis and treatment of DR helps reduce the risk of vision loss significantly.Therefore it's very important to deploy the general DR screening in the large diabetic population.For the extreme imbalance between the supply and demand of clinicians and patients,the most effective solution now is using techniques of computer vision and computer aided diagnosis(CAD)to improve ophthalmologists' diagnostic efficiency.And among these techniques,deep learning is one of the most active research fields.Based on the needs of DR diagnosis in Shanghai General Hospital,this thesis studies and analyzes multiple key issues during the development of the CAD system.And corresponding solutions are proposed in order to solve the practical problems faced by existing deep learning methods.Our main work and contributions for DR screening are listed as follows:(1)We propose a multi-stage attention model to achieve the automated assessment of fundus image quality in massive DR screening.The attention mechanism is introduced into the multi-scale feature maps of convolutional neural networks(CNNs)to improve the final classification performance.Further,it can generate focused image regions in the forward pass,which provides ophthalmologists a visual way to check the classification results and also prompts the collection personnel to eliminate the flaws in fundus imaging.(2)We propose a feature learning based method to improve the dense conditional random field(CRF)for retinal vessel segmentation in fundus images.A discriminative feature learning scheme is devised in order to learn unary features for the dense CRF.And a new thin-vessel enhancement process is also proposed for the dense CRF's pairwise potentials,further improving the vessel segmentation performance.(3)We propose a deep multiple instance learning method to achieve the joint detection of DR images and its DR lesions.This method leverages the complementary advantages of deep learning and multiple instance learning,therefore it can locate DR lesions just by using imagelevel labels.This method needs no lesion-level annotation,and it can provide ophthalmologists a visual way to verify DR detection results.Further,we propose an end-to-end multi-scale framework to better deal with the irregular DR lesions.(4)We propose a deep hybrid learning method to achieve the multi-class segmentation in fundus images.Three kinds of datasets are utilized,and each of them contains DR grades,fine-labelled vessels and weakly-labelled DR lesions respectively.To make full use of them,this method combines fully-supervised learning,weakly-supervised learning and active learning together and finally trains a CNN model to achieve the joint segmentation of retinal vessels,exudates,hemorrhages and microaneurysms.This method is quite practical as it can reduce the fine-labelling cost effectively.(5)We've developed a computer aided system for DR screening.This system integrates multiple functions like fundus image quality assessment,normal structure segmentation,DR lesion segmentation,DR grading and vessel analysis.Now it has been successfully applied to DR diagnosis and research by the ophthalmologists in Shanghai General Hospital.