Joint spatio-temporal registration and microvasculature segmentation of retinal angiogram sequences

We discuss the problem of 2D+t intra- and inter-sequential registration of retinal angiograms. A joint spatio-temporal registration algorithm is presented based on a RANSAC (RANdom SAmple Consensus) approach incorporating a quadratic model to describe "pairwise" image homography. This is incorporated into a local-to-global hierarchical joint registration framework. After registration, vessel centrelines are segmented to subpixel accuracy by applying multi-scale steerable complex wavelet filters. Frame-by-frame microvascular centrelines in Regions-of-Interest (ROIs) are evaluated against segmented centrelines of the temporal average of the registered sequences. The microvascular centrelines in registered sequences can be compared intra-sequentially and inter-sequentially, allowing non-invasive clinical monitoring of micro-circulation. This has the potential to detect the presence of microemboli and pathological structural alterations.