Related Experiment Video
Updated: May 11, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.6K
Automated learning of glaucomatous visual fields from OCT images using a comprehensive, segmentation-free 3D
Makoto Koyama1, Yuta Ueno2,3, Yoshikazu Ito2,3
1Minamikoyasu Eye Clinic, 2-8-30 Minamikoyasu, Kimitsu-shi, 299-1162, Chiba, Japan. minamikoyasuganka@gmail.com.
Scientific Reports
|April 18, 2025
Summary
A new 3D Convolutional Neural Network (3DCNN) model accurately estimates glaucoma visual field (VF) loss using Optical Coherence Tomography (OCT) scans. A comprehensive dataset improved model performance, showing potential for enhanced glaucoma assessment.
Area of Science:
- Ophthalmology
- Artificial Intelligence
- Medical Imaging
Background:
- Glaucoma diagnosis relies on visual field (VF) testing, which can be time-consuming and subjective.
- Optical Coherence Tomography (OCT) provides detailed structural information of the retina.
- Integrating OCT imaging with AI offers a novel approach for glaucoma assessment.
Purpose of the Study:
- To develop and evaluate a segmentation-free 3D Convolutional Neural Network (3DCNN) model for estimating visual field parameters in glaucoma patients using OCT images.
- To compare the performance of models trained on different datasets: Glaucoma-Specific Training Group (GTG) versus Comprehensive Training Group (CTG).
Main Methods:
- A segmentation-free 3DCNN model was trained using OCT images from 6335 participants (12,325 eyes) at a university hospital.
- Two training groups were used: GTG (glaucoma-specific) and CTG (comprehensive, including various ocular conditions).
- Model performance was evaluated by estimating VF thresholds and Mean Deviation (MD) using Humphrey Field Analyzer (HFA) 24-2 and HFA 10-2 test patterns.
Main Results:
- The CTG model significantly outperformed the GTG model in estimating VF thresholds and MD for both HFA 24-2 and HFA 10-2 (p < 0.001).
- Strong correlations were observed between estimated and actual VF thresholds and MD in the CTG model (Pearson's r ranging from 0.878 to 0.944).
- The CTG model exhibited lower estimation errors, particularly in severe glaucoma cases, and maintained stable performance even in advanced stages.
Conclusions:
- A comprehensive, diverse training dataset without manual preselection leads to superior performance in AI-based visual field estimation from OCT images.
- The developed 3DCNN model demonstrates high accuracy and robustness, showing significant potential for improving glaucoma assessment and monitoring in clinical practice.
- Future research should focus on external validation and application in varied clinical settings to confirm the model's broad utility.
