Abstract
Purpose :
Split detection imaging in adaptive optics (AO) ophthalmoscopy allows for visualization of cone inner segments, even when retinal pathology disrupts photoreceptor waveguiding. Due to the small field of view of individual AO images, adjacent images are typically montaged. Overlapping images in each montage then must be sorted so that the image with the highest quality is visible and used for further analyses of the cone mosaic. This sorting is currently a laborious manual task. Here, we propose an edge-based measure that enables automated ranking of image quality in AO split detection images.
Methods :
Our measure is based on the Canny edge detection algorithm, as implemented in Matlab (2020a). For each image, we run the edge detection and evaluate the proportion of the image occupied by detected edge pixels. Our method has a single threshold parameter which can be used statically to set the detection level, or adaptively chosen to produce a specified percentage of edge pixels. We compared rankings of image quality generated using the proposed approach against those of two human graders, and against two published image quality assessment methods (BRISQUE and PIQE). Using AO split detection montages from 5 choroideremia subjects, we identified 29 image locations, each with a minimum of 4 overlapping images. Each grader ranked every possible pair of images within each location (469 comparisons). The pairwise rankings were then aggregated into an overall ranking for each location using the Bradley and Terry model.
Results :
For the pairwise comparisons, the inter-grader agreement was 0.75. The agreement of the automated approaches {Proposed-Threshold, Proposed-Percent, BRISQUE, PIQE} were {0.67, 0.67, 0.52, 0.59} relative to Grader 1 and {0.68, 0.71, 0.43, 0.51} relative to Grader 2. The Spearman correlation of the aggregated rankings across all 29 locations was 0.75 between Grader 1 and 2. The correlation of the aggregated rankings for the automated approaches were {0.60, 0.61, 0.20, 0.34} relative to Grader 1 and {0.59, 0.61, 0.08, 0.21} relative to Grader 2.
Conclusions :
Our proposed edge-based image quality assessment generates quality rankings that are comparable to those of trained manual graders, and substantially outperforms two existing methods. This method has potential to reduce manual input to AO montages, thus enabling higher volumes of cone mosaic analyses.
This is a 2021 ARVO Annual Meeting abstract.