Diabetes Fundus Retinal Image Normalization And Optic Location Method

Diabetes is one of the highly frequent illnesses in modern society. An average50%of diabetic patients suffer from diabetic retinopathy(DR),which is an eye disease and acommon complication of diabetes that can cause blindness and vision loss if leftundiagnosed at an early stage.Therefore an early detection of DR is very important todelay the progress of the disease and to postpone blindness. In western developedcountries, diabetes patients are currently promoted to have regular fundus retinal imagescreening. Since diagnostic procedures require attention of an ophthalmologist, as well asregular monitoring of the disease, the workload and shortage of personnel will eventuallyexceed the current screening capabilities. To cope with these challenges, digital imagingof the eye fundus, and automatic or semi-automatic image analysis algorithms based onimage processing and computer vision techniques provide a great potential.Preprocessing stage is important in automatic image screening system, which is alsobenefit to the subsequent lesions detection. Preprocessing technical commonly refer toprocessing of filtering, non-uniformly luminosity and contrast normalization, imageenhancement, physiological structure segmentation (including localization andsegmentation problem of blood vessel, optic disc and macular). Correlation processingplay an important role in subsequent lesions detection, elimination of fake lesions andinterference suppression of other physiological structure to detection result. This paperresearches mainly on two points: the normalization of non-uniformly luminosity andcontrast and the location of optic disc in fundus image.The fundus images usually exist non-uniformly luminosity and contrast phenomenonin intra-image and inter-image resulting from influence of various factors in imageacquisition process. This makes certain lesion areas difficult to be observed and seriouslyaffects the diagnosis process and outcome. To copy with this problem, the paper proposestwo novel normalization methods. They are normalization enhancing method based onbackground estimation and homomorphic filtering and normalization enhancing methodbased on singular value decomposition low-pass-filter respectively. On the one hand, theproposed two methods implement the normalization of non-uniformly luminosity andcontrast, on the other hand, they enhance the contrast of lesion areas without changing thecharacteristic of original physiological structure and lesion areas. Moreover, subjectivevisual observation and objective evaluation test indicate that these two methods are significantly better than many other similar methods.As accurately locating optic disc is favorable for the diagnosis of eye disease,researchers have always been interested in automatic localization of OD. For example, alarge number of OD segmentation algorithms need initial seed point. In addition, the ODlocation can serve as a landmark for localizing and segmenting other anatomicalstructures such as the fovea (where the distance between the OD center and the center ofthe fovea is roughly constant). Also, since the OD can be easily confounded with largebright lesions, the detection of its location is important to remove it from a set ofcandidate lesions. Considering the vessel distribution and appearance characteristics ofoptic disc (OD) comprehensively, a novel OD localization method based onone-dimensional projection is proposed. Horizontal location is determined by vascularscatter degree, an evaluation index of vessel distribution. And vertical location is foundby brightness and edge gradient around OD. Four publicly-available databases and aself-selection database are used to evaluate the proposed method. The OD wassuccessfully located in357images out of380images (94%) with an averagecomputation time of13.2s. The proposed method shows good robustness on both normaland disease images.