Microaneurysms Detection In Diabetic Retinopathy Fundus Image

As the development of social and the improvement of living conditions, theincidence of diabetes has shown an upward trend in recent years. DR(DiabeticRetinopathy) is one of the complications associated with diabetes, and it is the mainreason results in blindness and visual impairment. If DR can be diagnosed and treatedin the earliest stage, patient’s condition can be controlled effectively.Fundus image isthe basis for the diagnosis of ophthalmology disease, and periodical inspection hasbecome an important way for DR screening. MA(Microaneurysms) is the first signand the smallest pathological changes can be observed in retinal image. ThereforeMAs detection and localization is particularly important for the diagnose of earlylesions.Due to the disturbance of texture and noise, optic disc, macular, and vasculature,the non-uniform illumination and contrast appear in most of retinal image, and thevariation of the size MAs, which result in error to location and segmentation thelesions. Therefore MAs detection is still a challenging problem in practice. Previousliterature shown low accuracy for subtle MAs detection and it is easily influenced byvessel extraction result. It’s been observed that inaccurate and imperfect vascularstructures are responsible for30%false negatives and90%false positive results. So,researchers have long been involved in improving positive ratio at the same timereducing false positive ratio. Based on digital color fundus image processing andrecognition technology，this paper made an exhaustive study on MAs detection andaims at improving robustness and accuracy of MAs detection. The main work issummarized as follows:A candidate MAs points are found based on background modeling andMahalanobis distance measure. Because blood vessel is the main cause of largely falsepositive, in order to exclude vascular pixels，we must extract vasculature as accuratelyas possible. Then a vessel structures extraction method based on Gabor filter isproposed, and non vessel inhibition operator, multi scale and multi hysteresisthreshold technique are combined to achieve blood detection. In the post processing,the relatively simple shape analysis and double rings filter is used to further removefalse lesions points. Experimental results show that the performance of our method isbetter than or approximate to other similar approaches, and it is greatly improve theaccuracy to detect the MAs close to blood vessels. A combination of multi-scale Gaussian filtering and integrated classification ofmicro-aneurysms detection method is researched. Based on the2-D Gaussiandistribution characteristics of MAs and variation in size, candidate microaneurysmslesions are picked out as seed points by multi-scale matched filtering. Then a regiongrowing technology is applied to segment the lesion areas, and the features of thelesions areas are analyzed. Finally, Adaboost neural network ensemble classifier isdesigned to distinguish the real MAs from all of the candidate lesions. The proposedmethod is tested on ROC database. Experiment show that the performance of ourmethod is better than that of previous double-ring filtering and morphologicalmethods.