Related Experiment Video
Updated: Jul 7, 2026
10:14
Author Spotlight: Ex Vivo OCT-Based Multimodal Imaging of Human Donor Eyes for Research into Age-Related Macular Degeneration
Published on: May 26, 2023
4.1K
Towards Accountable AI in Eye Disease Diagnosis: Workflow, External Validation, and Development
Qingyu Chen1,2, Tiarnan D L Keenan3, Elvira Agron3
1National Library of Medicine, National Institutes of Health, Maryland, USA.
Arxiv
|September 19, 2025
Summary
Artificial intelligence (AI) significantly improved accuracy and efficiency in diagnosing age-related macular degeneration (AMD). Further AI development is crucial for generalizability and real-world clinical adoption.
Area of Science:
- Ophthalmology
- Medical Artificial Intelligence
- Clinical Diagnostics
Background:
- Timely disease diagnosis is hindered by limited clinical resources and increasing patient loads.
- Artificial intelligence (AI) demonstrates expert-level diagnostic accuracy but faces adoption barriers due to a lack of downstream accountability.
- Gaps in workflow integration, external validation, and AI model development impede real-world implementation of medical AI.
Purpose of the Study:
- To address the downstream accountability challenges of medical AI.
- To evaluate an AI-assisted diagnostic and classification workflow for age-related macular degeneration (AMD).
- To enhance AI generalizability through further model development and testing on diverse datasets.
Main Methods:
- Developed and assessed an AI-assisted diagnostic workflow for AMD, involving 24 clinicians over four randomized assessment rounds comparing manual vs. AI-assisted diagnosis.
- Evaluated 2,880 AMD risk features across 960 images from 240 patient samples.
- Enhanced the DeepSeeNet AI model to DeepSeeNet+ using additional US population data and validated it on three datasets, including an external cohort from Singapore.
Main Results:
- AI assistance improved diagnostic accuracy for 23 of 24 clinicians, increasing the average F1-score by 20% (37.71 to 45.52).
- AI reduced diagnostic time per patient by 10.3 seconds, with sustained efficiency gains in later rounds.
- The enhanced DeepSeeNet+ model showed improved performance, achieving a 13% higher F1-score in the Singapore cohort, highlighting the importance of further development for generalizability.
Conclusions:
- AI-assisted diagnosis significantly enhances both accuracy and efficiency for AMD detection.
- Further AI development and validation are essential for ensuring generalizability across diverse patient populations.
- This study underscores the critical need for downstream accountability in the clinical evaluation of medical AI, with all code and models made publicly available.
