Purchase this article with an account.
Min Chen, Robert F Cooper, Grace K. Han, James Gee, David H Brainard, Jessica Ijams Wolfing Morgan; Toward Automated Alignment of Longitudinally-Acquired Adaptive Optics Retinal Images: Constellation Features. Invest. Ophthalmol. Vis. Sci. 2018;59(9):658.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Since its initial development, adaptive optics (AO) ophthalmoscopy has shown potential for tracking photoreceptor survival over time. However, this requires cellular-scale alignment of images acquired at different times. While automated montaging algorithms have been previously developed for AO images acquired at a single time, automated alignment of longitudinal AO data remains challenging. In part this is because the appearance of the cone mosaic varies over time. Here, we propose to use cell constellation feature descriptors for automated alignment of confocal AO images acquired at different times. Our method is based on a technique used for constellation detection in astronomy.
Constellation feature (CF) construction: 1) Calculate rough estimates of photoreceptor locations for each image using regional intensity extrema. 2) Overlay a grid on each estimated location to create a descriptor (CF) of the surrounding constellation of cells. 3) Vectorize and aggregate each grid to create a set of CFs for each image. CFs are then matched across images acquired at different times. Random sample consensus modeling is used to find a set of coherent matches and to calculate an alignment transform between images. Our initial algorithm evaluation was performed on 15 extra-foveal confocal AO image pairs taken from longitudinal data acquired 6-12 months apart in 4 healthy subjects. The proposed CF algorithm was compared to an algorithm based on scale invariant feature transform (SIFT) features currently used for automated alignment of AO images acquired at a single time (Chen et al Biomed. Op. Ex., 2016).
The CF algorithm was able to find numerous coherent CF matches for 14 of 15 image pairs (mean: 187, stddev: 157), but did not find a coherent match for 1 image pair. Qualitative analysis of the alignments of the 14 image pairs with matches showed close cell-to-cell correspondence. In contrast, the SIFT based algorithm led to coherent matches for only 12 of the 15 image pairs, and many fewer matches per pair (mean: 3, stddev: 2). Qualitatively, the 12 SIFT matched image pairs showed overall inferior alignment.
We propose the use of constellation features for automated alignment of longitudinally acquired AO confocal images. Initial analysis suggests such features have the potential to enable automated longitudinal montaging at the cellular scale.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.
This PDF is available to Subscribers Only