Automatic montaging of adaptive optics SLO retinal images based on graph theory

We present a fully automatic montage pipeline for adaptive optics SLO retinal images. It contains a flexible module to estimate the translation between pairwise images. The user can change modules to accommodate the alignment of the dataset using the most appropriate alignment technique, provided that it estimates the translation between image pairs and provides a quantitative confidence metric for the match between 0 and 1. We use these pairwise comparisons and associated metrics to construct a graph where nodes represent frames and edges represent the overlap relations. We use a small diameter spanning tree to determine the best pairwise alignment for each image based on the entire set of image relations. The final stage of the pipeline is a blending module that uses dynamic programming to improve the smoothness of the transition between frames. Data sets ranging from 26 to 119 images were obtained from individuals aged 24 to 81 years with a mix of visually normal control eyes and eyes with glaucoma or diabetes. The resulting automatically generated montages were qualitatively and quantitatively compared to results from semi-automated alignment. Data sets were specifically chosen to include both high quality and medium quality data. The results obtained from the automatic method are comparable or better than results obtained by an experienced operator performing semi-automated montaging. For the plug-in pairwise alignment module, we tested a technique that utilizes SIFT + RANSAC, Normalized cross-correlation (NCC) and a combination of the two. This pipeline produces consistent results not only on outer retinal layers, but also on inner retinal layers such as a nerve fiber layer or images of the vascular complexes, even when images are not of excellent quality.