Abstract
Purpose :
To develop a deep learning model for cone photoreceptors segmentation in Adaptive Optics Flood Illumination Ophthalmoscopy (AO-FIO), and to compare the results with manual annotations from different centres and at different eccentricities, as well as with the reference standard software (AOdetect).
Methods :
The rtx1 AO retinal camera (Imagine Eyes, France) was used to image 36 healthy subjects using the AO-VISION imaging protocol: 4° nasal(N) to 12° temporal(T), -5° inferior(I) to 5° superior(S) (21 images, 4°x4° each, 2° overlap). Each AO-FIO image was divided into patches of 128-by-128 pixels, with a 20-pixel overlap. The central location of the cones was manually annotated by graders from 3 different centers. A training set (625 patches) from 18 subjects (32 ± 12 years), was annotated by 1 center, whereas the test set (54 patches) from 18 subjects (40 ± 16 years), was annotated by all graders. The model was trained using a U-NET-based architecture with leave-one-out cross-validation. The model predictions were further processed with Otsu thresholding and peak extraction. The F1-score was used to assess the model's performance compared to the AOdetect, and to compute both intra- and inter-grader agreements.
Results :
The average intra-grader agreement was 0.85 ± 0.06 between 2°N to 2°T, followed by 0.83 ± 0.09 between 3–6°T, and 0.80 ± 0.10 between 7–10°T. The average inter-grader 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 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 and the graders (0.84 ± 0.05, 0.79 ± 0.03, and 0.68 ± 0.04, respectively). The percent error in cell density estimation resulting from the model prediction (2.5%, 0.7%, 2.9%) was lower than that of AOdetect (5.2%, 11.2%, 34.1%) when compared to the averaged values from the 3 centers at 2°N-2°T, 3°-6°T, and 7°-10°T, respectively.
Conclusions :
This study introduces a novel model for segmenting cone photoreceptors in AO-FIO images. The model demonstrated performance comparable to graders from 3 centers and yielded better results than AOdetect in healthy retinas, particularly at higher eccentricities (7°-10°T).
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.