Automatic Detection Methods Of Exudates On Diabetic Retinal Images

Diabetic retinopathy is a serious complication of diabetes, one of the main reasonsleading to adult vision loss and even blindness. Hard exudates is one of the early fea-tures of diabetic retinopathy, therefore early census of hard exudates will be efective inpreventing visual impairment. At present,diabetic retinopathy are manually checked byan ophthalmologist, with a long time and heavy working pressure. Auto-detection tech-niques with computer-based can efectively improve the retinal image detection efciencyand to bring the Gospel to patients and physicians.Many researchers and research institutions have proposed a variety of hard exudateautomatic detection technology. These methods can be roughly divided into four difer-ent categories: thresholding-based, region growing-based, morphology-based and clas-sifcation-based. These methods have their advantages and disadvantages, which haveachieved a certain efect on the detection of hard exudates, but they ignored the presenceof various confounding factors in the eye fundus image. As a result,based on the hardexudates characteristics, an automatic detection method of hard exudates is proposed inthis paper,which combines a histogram segmentation and classification.Image Enhancement in the preprocessing stage focus on brightness equalization andcontrast enhancement,which solve the problem of uneven brightness and low contrast inretinal fundus images. In the histogram segmentation stage, according to the brightnessof the characteristics of hard exudates,we use a Gaussian mixture model to fit the graydistribution in the histogram with a density estimation,and dynamic threshold for the ini-tial segmentation is obtained according to the probability density distribution function ofthe model. In order to make the boundary of the candidate area of hard exudates clear-er, we use the morphological reconstruction. By histogram segmentation stage, we get aset of candidate regions of hard exudates. In the classification stage, in order to separatehard exudates and non-hard exudates,44features are extracted for each candidate regionbased on the characteristics of hard exudates, and then support vector machine is used forclassification to obtain the final results.We assess the diagnostic accuracy of our approach in terms of two diferent criteria:on the DIABETIC1database our method can achieve superior performance comparedto existing techniques with sensitivity, positive predictive value of94.7%and90.0%in lesion-based criterion; sensitivity, specificity and accuracy of100%,81.3%and93.18%respectively in image-based criterion. Comparison with other experimental results, ourmethod has certain advantages and can provide an automatic detection method for diabeticretinopathy screening, which has the potential to be applied in clinical diagnosis.