Automatic Analysis Of Diabetic Retinopathy In Fundus Images

Diabetic retinopathy is a serious complication that may lead to retina damage.Early diagnosis and timely intervention is essential to reduce the risk of vision loss.The number of microaneurysms,hemorrhages and exudates in the standard collected fundus images can be used as a basis for diagnosis.In this paper,we focus on the lesion identification of diabetic retinopathy.We propose a local connectedness constraint and contrast normalization based framework for microaneurysm detection,and a gradual removal of vascular branches based framework for hemorrhage detection,and a dual-scale morphological segmentation based framework for exudate detection.Based on these algorithms,an assistant diagnosis system for diabetic retinopathy is realized.The problem of microaneurysm detection can be modeled as a small target detection problem in a complex background.Aiming at which,we propose an algorithm that microaneurysms are constrained by establishing connectivity rules.The difference between microaneurysms and interrupted blood vessels are enlarged with the local contrast normalization algorithm.Through comprehensive experiments on DIARETDB1 dataset and ROC dataset,we show that dots within vessels and noise points in the background can be well removed.Our method outperforms others with high sensitivity and specificity.The hemorrhages in fundus image have diferent shapes and sizes,and some even stick together with the retinal vasculature.In this paper,we adopt a dual-scale segmentation algorithm to get the dark regions.And the continuity of the vasculature are recovered by the connectedness recovery algorithm.Then by gradually removing the blood vessels from the dark regions,we get the hemorrhage candidates eventually.In the classification stage,a SVM classifier is trained to classify the candidates into hemorrhages and non-hemorrhages.The experiments on DIARETDB1 and DIARETDB0 show that our method achieves good performance on sensitivity and specificity.The exudates in fundus image usually have distinct edges,and vary with shapes and sizes.The exudate detection algorithm should be robust to different kinds of exudates.And interferences of the optic disc need to be eliminated as well.In this paper,we propose a dual-scale morphology based algorithm.The contours of the exudates are detected using the top-hat and black-hat operations.And images are resampled at the same resolution to reconstruct the final region.The experimental results show that the proposed algorithm is effective.