Purchase this article with an account.
Brea D Brennan, Heather Heitkotter, Joseph Carroll, Robert F Cooper; AOSLO Image Quality: Quantifying the Good, the Bad and the Ugly. Invest. Ophthalmol. Vis. Sci. 2022;63(7):220 – F0067.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Adaptive optics scanning light ophthalmoscopes (AOSLO) are a valuable tool for assessing retinal health. AOSLO image quality can degrade from factors such as poor media quality and eye motion, resulting in lost subject and analysis time. Here we assess the performance of a novel image quality metric against human graders.
We retrospectively obtained confocal and split-detection images of normative and pathological retinas (>8 diseases) from two AOSLOs with similar designs (Denoted as AOSLO 1 and 2). The dataset was comprised of 125 images from each modality and AOSLO (500 images total) at 1, 2, 4, and 8° from the fovea with field of views between 1 and 3°. For each image, we first determined its quality using a custom algorithm: First, the radial average of the log power spectrum of each image was obtained and differentiated. Next, the differentiated radial average was divided into “signal” and “noise” ranges by selecting a cutoff based on the plausible frequency range of photoreceptors. Finally, a signal to noise ratio (SNR) was determined as the ratio of the summed “signal” and “noise” ranges in decibels. For comparison to a gold standard, three individuals graded the same images (randomized, masked) for cell identifiability on a scale from 0-5, where 0 was considered unanalyzable and 5 easily identifiable. We examined the correlation between log-transformed SNR and grader scores and assessed grader agreement using intra-class correlation (ICC; R package: irr).
On average (±std dev), confocal images had SNR values of 56.2±3.1 and 38.2±6.9 and split-detection images had SNR values of 72.2±1.3 and 30.2±6.4 for AOSLO 1 and 2, respectively. On average, grader scores for confocal images were 3.5±1.3 and 2.4±1.3, and grader scores for split-detection images were 2.5±1.3 and 1.5±1.2 for AOSLO 1 and 2, respectively. These data are summarized in Table 1. All SNR and grader scores exhibited weak (r2<0.35), but significant correlations (p<0.01). Grader ICC was 0.71, implying moderate agreement amongst the graders.
The algorithm presented here successfully quantifies image quality in both confocal and split-detection AOSLO images. On average, graders showed agreement with our algorithm, and moderate agreement with each other.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.
SNR vs Grade for each grader (color coded), modality, and AOSLO. In general, the range of SNR in AOSLO 1 was small compared to AOSLO 2.
This PDF is available to Subscribers Only