Purchase this article with an account.
Alexander E. Salmon, Robert F. Cooper, Christopher S. Langlo, Ahmadreza Baghaie, Alfredo Dubra, Joseph Carroll; An Automated Reference Frame Selection (ARFS) Algorithm for Cone Imaging with Adaptive Optics Scanning Light Ophthalmoscopy. Trans. Vis. Sci. Tech. 2017;6(2):9. doi: https://doi.org/10.1167/tvst.6.2.9.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To develop an automated reference frame selection (ARFS) algorithm to replace the subjective approach of manually selecting reference frames for processing adaptive optics scanning light ophthalmoscope (AOSLO) videos of cone photoreceptors.
Relative distortion was measured within individual frames before conducting image-based motion tracking and sorting of frames into distinct spatial clusters. AOSLO images from nine healthy subjects were processed using ARFS and human-derived reference frames, then aligned to undistorted AO-flood images by nonlinear registration and the registration transformations were compared. The frequency at which humans selected reference frames that were rejected by ARFS was calculated in 35 datasets from healthy subjects, and subjects with achromatopsia, albinism, or retinitis pigmentosa. The level of distortion in this set of human-derived reference frames was assessed.
The average transformation vector magnitude required for registration of AOSLO images to AO-flood images was significantly reduced from 3.33 ± 1.61 pixels when using manual reference frame selection to 2.75 ± 1.60 pixels (mean ± SD) when using ARFS (P = 0.0016). Between 5.16% and 39.22% of human-derived frames were rejected by ARFS. Only 2.71% to 7.73% of human-derived frames were ranked in the top 5% of least distorted frames.
ARFS outperforms expert observers in selecting minimally distorted reference frames in AOSLO image sequences. The low success rate in human frame choice illustrates the difficulty in subjectively assessing image distortion.
Manual reference frame selection represented a significant barrier to a fully automated image-processing pipeline (including montaging, cone identification, and metric extraction). The approach presented here will aid in the clinical translation of AOSLO imaging.
This PDF is available to Subscribers Only