Color Eye Fundus Image Micro-aneurysm Detection Method Research

Medical image processing and analysis plays an important role in digital image processing,which is challenging and has attracted significant research interests.Colored retinal imaging is the most convenient technology for directly observing human blood vessels using optical instruments.Therefore,colored retinal images are widely applied for diagnosing ophthalmic diseases.Diabetic retinopathy(DR)is one of the major eye complications of diabetes.It has become the disease that has the highest rate of leading to blindness worldwide.Microaneurysm(MA)is the earliest symptom of DR.Its early detection is of great importance for ophthalmological clinical diagnosis and precaution.However,the existing MA detection through colored retinal imaging is usually performed manually by medical practitioners,which is time-consuming and human resource-demanding.The development of intelligent,automatic,computer-aided screening systems for MA detection can facilitate the clinical diagnosis and screening of DR.Firstly,due to the influence of imaging conditions and the diverse lesion stages,retinal images suffer from low brightness,low contrastness and uneven lunimations.In addition,the local contrast between MA and background is low.Therefore,retinal image enhancement is significant for MA detection.In this work,an enhancement method based on the composite Gamma correction is proposed.The experimental results demonstrate that our method can be applied to enhance the retinal images of diabetic retinopathy with different levels of brightness and stages of lesions,overcome the influence of uneven lunimation,improve the overall brightness and the contrast between MAs and the background,while retaining the effective information.Secondly,the accuracy of MA detection was unsatisfactory with false positives(FPs)and negative positives(NPs),because of the weak edges and the complex background with noises.A MA detection method is proposed based on circular bilateral Gabor filtering(CBGF)and local gradient analysis.With this method,visible MAs with weak edges and small sizes can be detected accurately,which improves the sensitivity of our method.Meanwhile,the noises of the local background and the small vessels are removed to reduce FPs.The retinal images from ROC and Hospital are chosen to evaluate the performance of our method.For the Hospital database,our method achieves sensitivities of 100%and 93.8%at the image and lesion level,respectively.For the ROC images,our method achieves 100%sensitivity,53.9%specificity and 96%accuracy at the image level.At the lesion level,the proposed method achieves the higher sensitivity 65.4%when there are 6.9 FPs per image on average.Colored retinal image is the major evidence for clinical diagnosis in ophthalmology.Since the field of view of retina in each single image is limited,image registration and fusion are significant for retinal images mosaicking.To obtain a retinal image with a larger field of view from multiple retinal images captured from the same patient,a retinal image mosaicking method is proposed based on Scale Invariant Feature Transform(SIFT)and Otsu matching.A feature matching algorithm based on Otsu is proposed to eliminate the false feature points.Registered images are fused by a maximum algorithm.Filtering and smoothing are not performed,which helps keep the retinal information.The experimental results indicate that the proposed method can be applied to retinal images with different levels of lesions.The blood vessels are registered without dislocation for images with low-to-medium stages of lesions.The Root-Mean-Square Error(RMSE)and Mean Absolute Error(MAE)of the matching points are 0.9409 and 0.9459 respectively.The correlation coefficient of overlapped regions is 0.9797.The success rate of two-image registration is 88%.Finally,an automatic DR screening system based on colored retinal image is designed.Several public retinal image databases are used for testing.Test results demonstrate that our method can accurately distinguish the images with and without MA.In summary,our works promote the development of the techniques for detecting MA in colored retinal images and advances the research on detecting early DR lesions.Our method provides an objective basis for early diagnosis of DR in clinics and reduces the medical cost,which can be applied for large-scale screening.