ATLASS: An AnaTomicaLly-Aware Self-Supervised Learning Framework for Generalizable Retinal Disease Detection.

Area of Science:

• Ophthalmology and Medical Imaging
• Artificial Intelligence in Healthcare
• Computational Biology

Background:

• Medical imaging, especially retinal fundus photography, is vital for early ocular disease detection.
• Deep learning diagnostics are hindered by the lack of expertly annotated retinal data, which is costly and time-consuming to acquire.
• Current self-supervised learning (SSL) models lack integration of crucial retinal anatomical knowledge, limiting clinical relevance.

Purpose of the Study:

• To develop an anatomically aware SSL framework for retinal imaging.
• To address the challenge of limited labeled data in medical AI by integrating domain knowledge.
• To enhance the clinical relevance and diagnostic capabilities of deep learning models for retinal diseases.

Main Methods:

• Introduced a novel SSL framework utilizing specialized masking of retinal structures (vessels, optic disc) during pretraining.
• Employed vessel and optic disc segmentation maps to guide the SSL process.
• Combined a Vision Transformer with dual-masking strategies and anatomically informed loss functions.

Main Results:

• The anatomically aware SSL framework demonstrated competitive performance in classifying various retinal diseases.
• Successfully developed clinically relevant feature representations without requiring extensive labeled data.
• Validated effectiveness across multiple datasets for diabetic retinopathy, glaucoma, and age-related macular degeneration detection.

Conclusions:

• Anatomically aware SSL effectively advances automated retinal disease diagnosis.
• The proposed framework overcomes the critical challenge of limited labeled medical data in ophthalmology.
• This approach enhances the utility of deep learning for ocular disorder detection and treatment planning.