Diabetic Retinopathy Detection Based On Fundus Images

Diabetic retinopathy(DR),as a complication of diabetes,is a high incidence of disease with high blinding rate.Early detection of lesion like hard exudates(HE)are beneficial in preventing visual impairment and even blindness.Traditional clinical practice places high requirements on the professional skills of ophthalmologists.In addition,such repetitive tasks can easily make them misdiagnose.Considering HEs have diverse shapes and the size of HEs also varies dramatically.What's more,they share similar intensity and contrast to other anatomical structures like vessel reflection and optic disc.Hence,it's a challenge but practically significant task to detect hard exudates automatically.The main contents of our research are as follows:1?This thesis studies the basic theory of visual detection in fundus image including mathematical morphological theory,machine learning models and some general models of image classification and semantic segmentation in deep learning.2?This thesis proposes a robust principle component analysis(RPCA)based hard exudate detection.The HE candidate regions are first obtained using RPCA.These regions are further marked by vessel mask and optic disc mask through conditional random field and SLIC super pixel segmentation respectively.Then,we extract fractal dimension feature and LBPV feature from each candidate.Finally,we use ensembled support vector machine to identify true hard exudates.Based on e-optha dataset,our algorithm achieves a sensitivity of 0.8253,a specificity of 0.8322 and an F-score of 0.8277.3?We propose a novel methodology for hard exudate detection based on deep model learned information and multi-feature joint representation.We first use optimized morphological reconstruction based-segmentation(OMRBS)to segment HE candidate regions.Then each region is characterized by combined features via ridge regression based on deep features and traditional features.Finally,a random forest is employed to classify each candidate.Based on e-optha dataset,our algorithm achieves a sensitivity of 0.8990,a specificity of 0.8868 and an F-score of 0.8929.Based on HEI-MED dataset,our algorithm achieves a sensitivity of 0.9477,a specificity of 0.9179 and an F-score of 0.9326.4?We also present a novel feature cascaded neural network(FCNet)for real-time HE segmentation.The FCNet is comprised of three critical modules:attention mechanism based module,feature fusion module and residual refine module.Based on e-optha dataset,our algorithm achieves a sensitivity of 0.8199,a specificity of 0.9669 and an F-score of 0.8874.Based on HEI-MED dataset,our algorithm achieves a sensitivity of 0.9721,a specificity of 0.9890 and an F-score of 0.9805.