Related Experiment Video
Updated: Aug 3, 2025
06:16
Author Spotlight: Unraveling the Pathogenesis of Age-Related Macular Degeneration and Discovering Potential Therapies
Published on: July 28, 2023
2.6K
An interpretable and interactive deep learning algorithm for a clinically applicable retinal fundus diagnosis system
Jaemin Son1, Joo Young Shin2, Seo Taek Kong1
1VUNO Inc., Seoul, Republic of Korea.
Scientific Reports
|April 12, 2023
Summary
This study introduces a novel deep learning system for diagnosing ophthalmic diseases from retinal images. The system provides interpretable diagnostic reasoning, enhancing clinical trust and accuracy in eye condition detection.
Area of Science:
- Ophthalmology
- Artificial Intelligence
- Medical Imaging
Background:
- Accurate identification of retinal abnormalities and ophthalmic diseases is crucial for managing vision-threatening conditions.
- Deep learning-based computer-aided diagnosis (CAD) systems show promise in improving reading efficiency and consistency.
- Clinical adoption of deep neural networks (DNNs) is hindered by their opaque reasoning processes.
Purpose of the Study:
- To develop a novel DNN architecture for identifying 15 abnormal retinal findings and diagnosing 8 major ophthalmic diseases.
- To introduce a method for interpreting DNN diagnostic reasoning and enabling interactive adjustments.
- To validate the model's diagnostic correlation with expert ophthalmologists.
Main Methods:
- Designed a novel DNN architecture for analyzing macula-centered fundus images.
- Developed a counterfactual attribution ratio (CAR) to elucidate diagnostic reasoning.
- Quantitatively and qualitatively evaluated the system's interpretability and diagnostic performance.
- Compared the model's CAR with expert ophthalmologists' correlation between findings and diseases.
Main Results:
- The DNN system achieved expert-level accuracy in identifying retinal abnormalities and diagnosing ophthalmic diseases.
- The CAR metric effectively illuminated the system's diagnostic reasoning process.
- Demonstrated the capability for quantitative and qualitative interpretation and interactive adjustment of CAD results.
- Confirmed that the model's reasoning aligns with ophthalmologists' understanding of finding-disease relationships.
Conclusions:
- The proposed DNN system offers a reliable and interpretable tool for ophthalmic disease diagnosis from retinal images.
- The CAR metric enhances transparency and trust in AI-driven diagnostic systems.
- This approach facilitates a deeper understanding of AI's diagnostic process, comparable to human experts.
