Microvasculature Segmentation of Co-Registered Retinal Angiogram Sequences

In this paper, we extend our published work [Cao et al., 2011] on methods to support the analysis of both vessel bed segmentation and microvascular centreline extraction in co-registered retinal angiogram sequences. These tools should allow researchers and clinicians to estimate and assess the (micro-)vessel diameter, capillary blood volume and microvascular topology for early stage disease detection, monitoring and treatment. We first discuss the problem of 2D+t intraand inter-sequential registration of retinal angiograms. On the spatio-temporally registered sequential angiograms, we then take into account both the global and local context to enable vessel bed segmentation and microvascular centreline extraction. Global vessel bed segmentation is achieved by combining phase-invariant orientation fields with neighbourhood pixel intensities in a patchbased feature vector for supervised learning. This approach is evaluated against other benchmarks on the DRIVE database [Niemeijer et al., 2004]. Local microvascular centrelines within Regions-of-Interest (ROIs) are segmented by linking the phase-invariant orientation measures with phase-selective local structure features. For each individual frame in the sequence, the extracted microvascular centrelines can be compared either against an objective ground truth to assess the quality and accuracy of registration, or with other frames intra-sequentially and inter-sequentially for non-invasive clinical monitoring of the micro-circulation. The combination of registration and segmentation has the potential to detect the presence of both microemboli and pathological structural alterations in a longitudinal study.