A Study On Fast Localization And Registering Of Retinal Fundus Images

Retinal fundus image processing has become one of the most activescience areas which combine clinical medicine with computer science.Retinal fundus images are used by clinicians to check for any abnormalitiesor pathological change in the retina. A typical fundus image is captured byusing special devices called ophthalmoscopes with its main features includeoptic disc (OD), fovea and blood vessels.In this paper, we mainly study on the solutions to OD detection fromboth color and IR fundus images and fundus image registration. We focus onresearching fundus image processing technology in systematic way, as wellas developing the fundus image processing system with high efficiency andaccuracy which makes our work practical. Specific work of this paper isdescribed as follows.Firstly, a three-phase classifier based on a line operator is designed inthis paper to capture the circular brightness structure associated with the OD,and a unified circle convolution mask within the polar coordinate system is used to improve diversity. We then use down-sampling for acceleration, anovel equalization of histogram for contrast enhancement and an improvedconnected component extraction for post-processing. Tested over Topcondataset, the proposed method outperforms the earlier methods by achievingthe accuracy of93%.Secondly, a registration algorithm based on mutual information isproposed, which gains best effect on retinal fundus image registration. Weapply the OD detection mentioned before to extract ROI. In this way, themultimodality diversity is ignored and the big data is shrunk.Thirdly, the fundus image processing is accelerated by parallelcomputing on both multi-core CPU and GPU. We apply OpenCL as theheterogeneous parallel scheme. The improved system is able to locate the ODfrom fundus image as fast as0.05second, faster than the older one by2636times which turns to be real-time processing. Simultaneously, the speed ofimage registration archives1.86second which actually makes a big stepforward comparing with the minutes-count registration previously.Lastly, a retinal fundus image processing system has been developedto detect OD and register fundus images automatically, accurately and rapidly.We design and apply a multi-level parallelism framework to make sure thesoftware can run validly on any kind of computing machines. The software does a good job on portability and usability.