Abstract
Purpose :
Adaptive optics (AO) scanning laser ophthalmoscopy offers high resolution in-vivo imaging and analysis of microscopic structures in the retina. A single AO image subtends a small fraction of the total retina; thus, examining a large retinal area requires the creation of a montage consisting of multiple images. These montages are often constructed manually, and can pose a bottleneck in large studies. Here, we propose an automated approach for AO montaging that extends an extant algorithm (Li et al., 2012, Opt. Eng.) by using multi-modal features from simultaneously acquired confocal, split-detection, and dark-field AO images.
Methods :
Algorithm steps: (1) The image closest to the fovea is selected as the initial reference. (2) Adjacent images that overlap the reference are aligned using scale-invariant features from each modality and a random sample consensus model to remove outliers. (3) Aligned images are used as new potential references for adjacent unaligned images. The reference used is determined by the highest number and percent of features matched. (4) Complete the montage by repeating steps (2) and (3) for all images in the dataset. We evaluated our algorithm on 11 datasets (9 different controls eyes, 1 retinitis pigmentosa, 1 central serous choiroretinopathy) containing 47-79 AO image sets, each with a confocal, split detection, and dark-field modality. For two datasets, we compared the relative translation of adjacently aligned images against the translations in a manual montage.
Results :
The median translation difference between the automatic and manual montages was 6.6 pixels (3.1 microns). Over the 11 datasets, 17 of the 659 images could not be aligned to their initial montage. In such cases, the algorithm continued with a new disjoint montage. Four of these disjoints resulted from limited overlap between adjacent images. The other 13 were correctly detected as non-overlapping. There were several cases where split detection provided features that allowed alignment where the confocal modality when used alone did not. Runtime for the algorithm was considerably faster than manual montaging and required minimal operator input.
Conclusions :
We present an accurate and automatic AO montaging algorithm that provides a faster and less operator intensive alternative to manual image alignment. In future work, the algorithm will be optimized for speed and generalized to include longitudinal data.
This is an abstract that was submitted for the 2016 ARVO Annual Meeting, held in Seattle, Wash., May 1-5, 2016.