Fast and robust image registration with local motion estimation for image enhancement and activity detection in retinal imaging

Abstract

Recent retinal imaging systems, including those featuring adaptative optics (AO) correction, produce sequences of images having a very high spatial resolution, but possibly affected by a low Signal-to-Noise Ratio (SNR). A simple, fast and efficient way to improve SNR in many cases is to average images. However, image quality appears to be highly variable from one frame to another, and sequences feature large inter-frame global motions due to eye movements. We consider in this paper both global and local motion estimation to deliver enhanced images or to detect vessel activity. We tackle the problem of retinal vessels local motions due to bloodstream. If not accounted for, these motions lead to some blurring of the enhanced image, which in turn degrades image analysis, for instance the estimation of the thickness of the vessels wall or the wall-to-lumen ratio. Interframe local motions need thus to be estimated. We propose to compensate for local motions using a fast optical flow estimation called FOLKI developed at ONERA/DTIM, which combines very fast computational times with a good accuracy. The method is then applied to detect vessel activity. Improvements are demonstrated both in terms of final image quality and of local activity detection, on sequences recorded on Imagine Eyes' rtx1 system.