COINS: Counting Cones Using Inpainting Based Self-supervised Learning.

Area of Science:

• Ophthalmology
• Computational Imaging
• Machine Learning

Background:

• Accurate cone counting in retinal images is crucial for diagnosing and monitoring eye diseases.
• Traditional methods struggle with low-resolution and wide field-of-view adaptive optics (AO) images.
• Existing algorithms lack robustness in diverse imaging conditions.

Purpose of the Study:

• To introduce a novel self-supervised learning (SSL) approach for counting photoreceptor cones in AO images.
• To evaluate the performance of the proposed COINS (COunting cones using IN-painting based Self-supervised learning) method against established algorithms.
• To demonstrate the efficacy of COINS in handling wide field-of-view, low-resolution AO datasets.

Main Methods:

• Developed a COINS model utilizing an inpainting pretext task for representation learning.
• Fine-tuned the SSL model with a limited set of expert-annotated images.
• Applied the COINS approach to a 4°×4° dataset of AO images acquired with an AO rtx1 device.

Main Results:

• The COINS method significantly outperformed the baseline Delaunay triangulation Voronoi algorithm for cone counting.
• COINS demonstrated accurate cone counting capabilities in regions as small as 80×80 pixels.
• The model showed robust performance across various locations within the wide field-of-view images.

Conclusions:

• COINS offers a powerful and accurate solution for cone counting in challenging AO retinal images.
• Self-supervised learning, combined with inpainting, provides an effective strategy for analyzing low-resolution ophthalmic data.
• The proposed method has the potential to advance diagnostic capabilities for retinal diseases.