Research On Automatic Detection Method Of Diabetic Retinopathy Based On Deep Learning

At least 2.2 billion people worldwide suffer from various degrees of eye diseases or vision impairment,and 50% of them can be identified through screening for early intervention and prevention.As one of the most common methods for eye health examination,fundus photography can be used for the preliminary diagnosis of various eye diseases such as diabetic retinopathy(DR),age-related macular degeneration,and glaucoma.Among these retinal diseases,DR is a common complication of diabetes with a high incidence rate and is the leading cause of blindness.Patients often miss the best treatment opportunity in the early stage of DR due to the lack of obvious symptoms,resulting in a rapid increase in the risk of irreversible visual impairment or even blindness.Therefore,the large-scale standardized application of fundus photography-based DR screening can effectively reduce the health risks associated with DR.However,the implementation of large-scale screening is hindered in China.Thus,there is an urgent need to develop automatic diagnosis methods to solve this problem.Data-driven machine learning methods,especially deep learning models in recent years,have provided a technique for automatically diagnosing eye diseases from fundus photographs.Deep learning-based eye disease diagnosis generally requires the support of high-quality,diverse,and evenly distributed datasets.However,most existing datasets focus on specific types of eye diseases.Some datasets have problems such as label noise or low image quality,which hinder the study of deep learning models for automatic screening and diagnosis.With the support of high-quality datasets,deep learning contributes to the automatic detection of DR,which greatly assists ophthalmologists in early screening and diagnosis of DR.Studies have shown that convolutional neural networks can detect or segment DR-related lesions such as microaneurysms(MA),hemorrhages(HE),hard exudates(EX),and soft exudates(SE),which benefits DR automatic screening and diagnosis.The key to fine-grained DR lesion detection tasks lies in(1)extracting discriminative features that are sensitive to small lesion areas and are robust to DR-unrelated interference,and(2)learning lesion features from highly imbalanced image datasets.Therefore,constructing high-quality fundus photograph datasets and developing high-performance DR automatic detection methods are of great significance for promoting large-scale eye health screening.This thesis conducts comprehensive research on DR automatic detection methods from the aspects of the dataset and deep learning models.The specific research contents are as follows:(1)In order to expand and enrich the existing multi-disease fundus datasets,we construct a new high-quality dataset EDDF containing 28,877 fundus photographs,which may support research on deep learning-based diagnostic,under the guidance of ophthalmologists from First Hospital of Nanchang.This dataset contains 15,000 healthy samples and 8 eye diseases,including DR,age-related macular degeneration,glaucoma,pathological myopia,hypertensive retinopathy,and retinal vein occlusion.(2)In order to improve the performance of automatic diagnosis of DR,we propose a novel deep learning network for DR lesion segmentation,named Lesion-Aware Network(LANet).It utilizes the attention mechanism to better capture DR lesion features from imbalanced data.The LANet is constructed based on the encoder-decoder structure with Lesion-Aware Module(LAM)and Feature Fusion Module(FPM)embedded.LAM is designed to capture lesion regions from high-level features,while FPM is utilized to assist the fusion of low-level and high-level features.A segmentation map of DR lesions can be output via LANet.By performing experiments on three fundus datasets with pixel-level annotation,the results of DR lesion segmentation on IDRi D-Seg,DDR-Seg and FGADR-Seg datasets showed that the mAP scores obtained by LANet increased by 18.0%,4.6% and 1.9% compared with the second-best method.In addition,the effectiveness of the submodules proposed in this thesis is verified by ablation experiments.(3)In order to achieve the screening diagnosis prediction results while obtaining the DR lesion segmentation map,this thesis takes LANet as the main network,in which the DR lesionrelated features are involved,and further extends it to a DR Lesion-Aware Screening Network(LASNet)by simply adding a classification block.Finally,in the DR screening experiment on DDR-Scr and the proposed EDDF,LASNet achieved state-of-the-art performance with AUC scores of 0.963 and 0.987,respectively.