Related Experiment Video
Updated: Jul 15, 2025
04:36
Author Spotlight: Understanding Retinal Vessel Resilience and Disease Progression
Published on: January 12, 2024
1.1K
Artificial intelligence, explainability, and the scientific method: A proof-of-concept study on novel retinal
Parsa Delavari1,2, Gulcenur Ozturan1, Lei Yuan1
1Ophthalmology and Visual Sciences, University of British Columbia, Vancouver, V5Z 0A6 BC, Canada.
PNAS Nexus
|September 25, 2023
Summary
We combined artificial intelligence (AI) explainability with the scientific method to discover novel retinal biomarkers for sex classification. This approach identified new diagnostic features in retinal images, enhancing clinical capabilities.
Area of Science:
- Ophthalmology
- Artificial Intelligence
- Biomedical Discovery
Background:
- Sex-based differences in retinal images are not well-recognized by diagnosticians.
- Artificial intelligence (AI), specifically convolutional neural networks (CNNs), can classify patient sex from retinal images.
- Explainability methods are crucial for understanding AI decisions and facilitating scientific discovery.
Purpose of the Study:
- To develop a structured approach combining AI explainability with the scientific method for novel biomarker discovery.
- To identify and verify new sex-based biomarkers in retinal images using a CNN model.
- To assess the potential of these biomarkers for enhancing clinical diagnostic capabilities.
Main Methods:
- Developed a Visual Geometry Group (VGG) CNN model to classify sex from retinal images.
- Utilized post hoc interpretability tools to generate hypotheses on retinal sex differences.
- Tested 14 hypotheses on independent datasets, with nine showing significant differences.
- Verified five key hypotheses, identifying specific vascular and peripapillary region differences in male eyes.
Main Results:
- Identified five novel sex-based biomarkers in retinal vasculature and peripapillary regions.
- Demonstrated significantly greater length, nodes, and branches of retinal vasculature in male eyes.
- Observed greater retinal area covered by vessels in the superior temporal quadrant and a darker peripapillary region in male eyes.
- Trained ophthalmologists to recognize these novel features, significantly improving their classification performance.
Conclusions:
- The integration of AI explainability and the scientific method is a powerful tool for biomarker discovery in ophthalmology.
- Novel sex-based biomarkers in retinal images can be identified and validated using this approach.
- These findings have the potential to augment diagnostic capabilities and improve patient care through enhanced clinical toolkits.
