Abstract
Purpose :
Human color vision is based on the mixing of neural signals from three types of cone photoreceptors that are sensitive to short (S cone), medium (M cone), and long (L cone) wavelengths of light. Although S, M, and L cones differ primarily in differences of photopigment, other structural differences might be detectable by imaging and potentially useful for distinguishing cone spectral types. To test, we extracted morphological features of thousands of cones from AO-OCT images to classify spectral types and validate against functionally classified results.
Methods :
Adaptive optics optical coherence tomography (AO-OCT) volume images of 1°×0.8° at 3.8° temporal retina were acquired in three color normal subjects. Cones were first classified using the cone phase optoretinogram [1]. Morphological features of each cone cell were then extracted from the AO-OCT images: inner segment length (ISL), outer segment length (OSL), and reflectance diameters of the inner-segment-outer-segment junction (ISOS) and cone outer segment tip. We tested for statistical differences in morphological features between cone types and assessed whether these differences could identify spectral cone types using machine learning classifiers (K-nearest neighbors, support vector machine (SVM), and multilayer perceptron). Five-fold cross-validation was performed to evaluate classification performance in terms of precision, recall, and F1 scores.
Results :
In the three subjects, 1,986 cones were evaluated (1,328 L cones, 542 M cones, and 116 S cones). The only notable difference we found between M and L cones in all three subjects was the longer (0.46μm) OSL of L cones, which was significant (p<0.05) in two of the subjects. Compared to M and L cones, S cones had a longer (3.6μm) ISL, shorter (3.8μm) OSL, and larger (0.71μm) ISOS diameter. These differences were significant for all subjects and consistent with histology [2]. While differences were too small to classify M cones from L cones with any classifier, S cones could be classified effectively using all three classifiers with SVM generating the best performance of 81.2±7.2% precision, 71.4±6.3% recall, and an F1 score of 75.7±5.1%.
Conclusions :
Differences in S, M, and L cone structures are revealed in AO-OCT images enabling individual S cones to be distinguished from M and L cones.
[1] Zhang, et al. PNAS 116.16 (2019): 7951-6
[2] Curcio, et al. JCN 312.4 (1991): 610-24
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.