Cone Identification on Merged Non-Confocal Quadrant-Detection Adaptive Optics Scanning Light Ophthalmoscopy (AOSLO) Images

Abstract

Purpose : Cone photoreceptor inner segments in non-confocal split-detection AOSLO images appear as a bubble wrap-like mosaic in which each cell has bright and dark opposite edges, complicating algorithms for automated cell identification. Here we demonstrate that the merging of non-confocal quadrant-detection images allows cone identification in normal and pathologic eyes using algorithms with minimal complexity.

Methods : Confocal and non-confocal quadrant-detection images were collected simultaneously in 6 controls (7 eyes) and 7 patients (8 eyes) with retinal pathology (PMID: 33680539). To remove non-homogenous intensity and optimize edge contrast of photoreceptors, a merged non-confocal quadrant-detection image was generated by combining 4 emboss-filtered split-detection images (0, 45, 90, & 135°) (PMID: 36531581). Automated parafoveal cone identifications on the confocal image at retinal eccentricity 120-570µm using ImageJ TrackMate (PMID: 35654950) were treated as ground truth. Validation of cone identification accuracy of the merged quadrant-detection image processed with different kernel sizes (3×3, 5x5, and 7x7 pixels) were compared to the ground truth (Figure). Semi-automated cone identification on merged quadrant-detection images at retinal eccentricity 300-3200µm was also performed in pathologic eyes. True positive rate, false positive rate, and Dice’s coefficient were computed as described previously (Dice, 1945).

Results : Dice’s coefficient of automated parafoveal cone identification was 0.953±0.019 (range: 0.904-0.983). Statistically significant differences were observed in 3x3 vs 7x7 and 5x5 vs 7x7 kernel sizes using Friedman test (post-hoc, p<0.001 & p=0.014, respectively). No statistical difference was found in 3x3 vs 5x5 kernel sizes (post-hoc, p=0.68). Evaluation of semi-automated cone identifications on merged quadrant-detection images in pathologic eyes processed using 5x5 kernel size showed a Dice’s coefficient of 0.964±0.019 (range: 0.847-0.998).

Conclusions : We demonstrated high accuracy of cone identification on merged non-confocal quadrant-detection images at retinal eccentricity of 120-3200µm in control and pathologic eyes using ImageJ Trackmate.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.