Advances in machine learning for ABCA4-related retinopathy: segmentation and phenotyping

Area of Science:

• Ophthalmology and Medical Imaging
• Artificial Intelligence in Healthcare
• Retinal Degenerative Diseases

Background:

• Stargardt disease (ABCA4-related retinopathy) is a common juvenile macular dystrophy without FDA-approved treatments.
• Current disease monitoring methods like image segmentation and phenotyping are subjective and time-consuming.
• Machine learning (ML) offers potential for automating these critical evaluation processes.

Purpose of the Study:

• To conduct a scoping review of ML applications in ABCA4-related retinopathy (ABCA4R).
• To focus on ML's role in segmentation and phenotyping for disease progression monitoring and patient subgroup categorization.
• To identify challenges and advanced ML techniques in the field.

Main Methods:

• A systematic scoping review following PRISMA guidelines.
• Selection of 15 relevant articles from 264 screened.
• Analysis of studies focusing on segmentation (lesions, flecks, layers, en-face imaging) and phenotyping (ERG, visual acuity, microperimetry).

Main Results:

• ML models demonstrated high performance, with segmentation DICE scores of 0.99 and ERG phenotyping accuracies exceeding 90%.
• Effective ML approaches included ensemble modeling, self-attention, soft-labeling, and dynamic frameworks.
• Challenges such as small datasets and variable disease presentation were noted, with Monte Carlo dropout and active learning showing promise for improvement.

Conclusions:

• ML techniques are highly effective in automating key ABCA4R evaluation steps.
• These advancements can accelerate the development of new therapies for Stargardt disease.
• ML has the potential to significantly deepen the understanding of ABCA4R.