Automated Cone Photoreceptor Detection in Adaptive Optics Flood Illumination Ophthalmoscopy

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Cone photoreceptor cells are crucial for high-acuity vision.
• Adaptive optics flood illumination ophthalmoscopy (AO-FIO) enables in vivo imaging of retinal cells.
• Accurate detection of cone cells is vital for diagnosing and monitoring retinal diseases.

Purpose of the Study:

• To develop and validate a deep learning (DL) model for automated cone photoreceptor cell detection using AO-FIO.
• To compare the DL model's performance against human graders and manufacturer software.

Main Methods:

• A U-Net based DL model was trained on 625 AO-FIO image patches from 18 healthy participants.
• The model underwent parameter optimization using the tree-structured Parzen estimator.
• Performance was evaluated using the F1 score, comparing the model to manual annotations from 3 institutions and manufacturer software (AOdetect Mosaic).

Main Results:

• The developed DL model achieved an average F1 score of 0.89 ± 0.04.
• Model performance (F1 scores: 0.87–0.81) was comparable to intergrader agreement (0.84–0.76) across different retinal eccentricities.
• The DL model outperformed manufacturer software (AOdetect Mosaic) auto-detection, especially at higher eccentricities (7°–10°T).

Conclusions:

• The AO-FIO ConeDetect model demonstrates high accuracy in detecting cone cells from AO-FIO images.
• The model's performance rivals expert graders and exceeds current automated software, reducing the need for manual corrections.
• This technology has the potential to significantly accelerate cone mosaic analysis in clinical and research settings.