Purchase this article with an account.
Alexander E Salmon, Robert F. Cooper, Christopher S Langlo, Benjamin S Sajdak, Moataz M Razeen, Alfredo Dubra, Joseph Carroll; Automated Reference Frame Selection (ARFS) for Registration of Scanning Ophthalmoscope Image Sequences. Invest. Ophthalmol. Vis. Sci. 2016;57(12):5971.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Due to involuntary eye motion, images from scanning ophthalmoscopes are subject to inter- and intra-frame distortion. When averaging is used to increase the signal-to-noise ratio, a reference frame subjectively determined to be minimally-distorted is selected and additional frames are registered to it before averaging. Manual frame selection is time consuming and subjective. We present an automated algorithm for organizing frames into spatial clusters and objectively selecting representative frames with minimal distortion.
The algorithm consists of a weighted sum of image metrics including intensity, Fourier-based line detection, normalized cross-correlation, as well as motion tracking and clustering to choose spatially representative frames. Validation was achieved by using strip-based registration (Dubra & Harvey, 2010) to align undistorted AO-flood illuminated images to either ARFS- or manually-derived AOSLO images of photoreceptors (PR) in 7 normal subjects, and comparing the magnitude of the pixel shift vector (PSV) required for registration in each case. We assessed overlap in frame choice by ARFS with that of one expert observer (9 normal subjects, 3 retinal locations, 10 frames/location), and between ARFS and several human observers (15 subjects with albinism, achromatopsia, or retinitis pigmentosa; 10 videos/subject, 1 frame/cluster).
Automated reference frame selection resulted in mean ± SD PSV of 1.77±1.1 pixels versus 2.55±0.7 pixels for subjective selection (p=0.22, paired t-test). For the normal datasets, there was a 9.3% overlap in frame choice between ARFS and the expert observer and a 7.0% overlap in the pathological datasets. In all cases, the images generated using ARFS- and manually-selected reference frames were subjectively comparable in quality.
ARFS identifies minimally-distorted reference frames in AOSLO image sequences, even in subjects with nystagmus (a particularly challenging population for manual reference frame selection). The low overlap in frame choice between ARFS and human observers illustrates the difficulty in subjectively assessing image distortion. While ARFS has been shown to operate successfully on AOSLO PR image sequences, further work is needed to evaluate performance on other types of images.
This is an abstract that was submitted for the 2016 ARVO Annual Meeting, held in Seattle, Wash., May 1-5, 2016.
This PDF is available to Subscribers Only