Related Experiment Video
Updated: May 2, 2026
08:17
Optical Coherence Tomography: Imaging Mouse Retinal Ganglion Cells In Vivo
Published on: September 22, 2017
19.2K
Identifying retinopathy in optical coherence tomography images with less labeled data via contrastive graph
Songqi Hu1, Hongying Tang2, Yuemei Luo3
1School of Information Engineering, Shanghai University of Maritime, 1550 Haigang Avenue, Shanghai 201306, China.
Biomedical Optics Express
|September 30, 2024
Summary
This study introduces a new method for detecting retinopathy using optical coherence tomography (OCT) images with minimal labeled data. The approach achieves expert-level performance, significantly reducing the need for extensive manual annotation in medical imaging analysis.
Area of Science:
- Ophthalmology
- Medical Imaging
- Computer Vision
Background:
- Diabetic retinopathy detection from optical coherence tomography (OCT) images is crucial for timely intervention.
- Traditional computer vision methods require large annotated datasets, which are costly and time-consuming to acquire.
- Developing efficient retinopathy detection models with limited labeled data is a significant challenge.
Purpose of the Study:
- To propose a novel contrastive graph regularization method for retinopathy detection using fewer labeled OCT images.
- To enhance the performance of automated retinopathy detection systems by reducing reliance on extensive manual annotation.
- To investigate the efficacy of combining class prediction probabilities and embedded image representations for improved model training.
Main Methods:
- A novel contrastive graph regularization technique was developed, integrating class prediction probabilities and embedded image representations.
- Memory smoothing constraints were employed to refine pseudo-labels by aggregating nearby samples in the embedding space.
- The method was trained and evaluated on two widely used OCT datasets, utilizing a limited number of labeled images.
Main Results:
- The proposed method achieved high classification accuracy (>0.96) and an Area Under the Curve (AUC) of 0.998 on OCT datasets.
- Expert-level performance was attained with as few as 80 labeled OCT images.
- The method surpassed the performance of most human experts with only 160 labeled images.
Conclusions:
- The novel contrastive graph regularization method effectively detects retinopathy from OCT images with significantly reduced labeled data requirements.
- This approach offers a promising solution for overcoming the bottleneck of data annotation in medical image analysis.
- The findings suggest a potential for developing more accessible and efficient AI-powered diagnostic tools in ophthalmology.
