To automatically detect hard exudates from fundus images

and to develop an automatic diabetic retinopathy screening system

an automatically detecting approach based on RBF neural network and threshold segmentation was established and studied. First

the green channel of the original fundus image is coarsely segmented by an improved Otsu thresholding based on minimum inner-cluster variance

and candidate regions are obtained. Second

several features of candidate regions are extracted and selected by means of logistic regression. Finally

the RBF neural network is built with the optimal subset of features and judgments of these candidate regions. Furthermore

post-processing is carried out to improve the detection accuracy. The approach is tested on a new set which contained 50 fundus images with variable color and brightness. With an image-based criterion

sensitivity of 100%

specificity of 90.9%

and accuracy of 96.0% are achieved. Average sensitivity of 93.9% and average positive predict value of 95.5% are also achieved with a lesion-based criterion. Furthermore

the average time cost in processing an image is 13.6 s. Results suggest that the approach is stable and reliable

and can fast and effectively detect hard exudates from fundus images.