Aligning fundus images via the automatic extraction of retinal surface vessels

Extraction of retinal surface vessels allows for fast registration of human fundus images. Two unsupervised, automatic blood vessel extraction algorithms were implemented. The first approach enhanced the surface vessels by applying Gabor filters (Oloumi et al, 2007 Conference Proceedings of IEEE Engineering in Medicine and Biology Society 6451–6454). The second approach combines multiple preprocessing steps to reduce the influence of noise within the images (such as the central light reflex) (Marín et al, 2011 IEEE Trans Medical Imaging 30 146–158). Both approaches resulted in similar accuracy (∼94%) when applied to the DRIVE database of fundus images (Staal et al, 2004 IEEE Trans Medical Imaging 23 501–509). Spectroscopic fundus images were then aligned using the outputs of each approach. In particular, the automatically generated vessel masks for each fundus image were shifted both vertically and horizontally relative to a reference vessel mask. For each possible shift, the absolute difference between the two masks was computed. The translation that generates the minimal difference between the two masks was used to align the raw images. Thus, translations, resulting from minor eye tremors and movements, were corrected. Spectroscopic fundus images can then be used to calculate the relative oxygenation of different retinal compartments.