Automated Cone Photoreceptor Detection in Adaptive Optics Flood Illumination Ophthalmoscopy

Abstract

Purpose

To develop and validate a deep learning–based model for detecting cone photoreceptor cells in adaptive optics flood illumination ophthalmoscopy (AO-FIO).

Design

Healthy volunteer study.

Participants

A total of 36 healthy participants were included.

Methods

The imaging protocol consisted of 21 AO-FIO images per eye acquired with the rtx1 adaptive optics retinal camera (Imagine Eyes), 4° × 4° each with 2° overlap, imaging a retinal patch 4° nasal (N) to 12° temporal (T) and −5° inferior to 5° superior relative to the fovea. Each image was divided into patches of 128 × 128 pixels, with a 20-pixel overlap. A training set (625 patches) from 18 subjects (32 ± 12 years, 6 males and 12 females) was annotated by a single center, whereas the test set (54 patches) from 18 subjects (40 ± 16 years, 11 males and 7 females) was annotated by graders from 3 different institutions. The deep learning model, based on the U-Net architecture, underwent a parameter search using the tree-structured Parzen estimator.

Main Outcome Measures

The F1 score was used to determine both intragrader and intergrader agreements and to evaluate the model’s performance compared with the automated detection by the manufacturer’s software (AOdetect Mosaic).

Results

The average intragrader agreement was 0.85 ± 0.06 between 2°N and 2°T, followed by 0.83 ± 0.09 between 3 and 6°T, and 0.80 ± 0.10 between 7 and 10°T. The average intergrader agreement for the 3 centers was 0.84 ± 0.05, 0.79 ± 0.05, and 0.76 ± 0.06 at 2°N–2°T, 3–6°T, and 7–10°T, respectively. The best combination of hyperparameters based on the tree-structured Parzen estimator algorithm achieved an F1 score of 0.89 ± 0.04. The average agreement between the model and the graders was 0.87 ± 0.04, 0.85 ± 0.03, and 0.81 ± 0.03 at 2°N–2°T, 3°–6°T, and 7°–10°T, respectively. These values were higher than those between AOdetect’s auto detection without manual correction and the graders (0.84 ± 0.05, 0.79 ± 0.03, and 0.68 ± 0.04, respectively). A reduction in cone density was noted at greater eccentricities, in line with previous research findings, and the model indicated variations in estimating cell density for individuals aged 18 to 30 compared with those aged ≥50 years.

Conclusions

The performance of the developed deep learning–based model, AO-FIO ConeDetect, was comparable to that of graders from 3 medical centers. It outperformed the manufacturers’ software auto-detection, particularly at higher eccentricities (7°–10°T). Hence, the model could reduce the need for manual correction and enable faster cone mosaic analyses.

Financial Disclosure(s)

Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.