Related Experiment Video
Updated: Nov 19, 2025
05:56
Objectification of Tongue Diagnosis in Traditional Medicine, Data Analysis, and Study Application
Published on: April 14, 2023
2.9K
Leveraging Regular Fundus Images for Training UWF Fundus Diagnosis Models via Adversarial Learning and
IEEE Transactions on Medical Imaging
|February 3, 2021
Summary
This study uses a modified CycleGAN to generate more ultra-widefield (UWF) fundus images from regular ones, improving AI model training for eye diseases. This approach enhances diabetic retinopathy classification and lesion detection with limited UWF data.
Area of Science:
- Ophthalmology
- Medical Imaging
- Artificial Intelligence
Background:
- Ultra-widefield (UWF) 200° fundus imaging offers broader insights than traditional 30°-60° cameras.
- A significant domain gap exists between regular and UWF fundus images, hindering model performance.
- Annotating medical data for AI training is labor-intensive and time-consuming.
Purpose of the Study:
- To leverage existing regular fundus images to improve limited UWF fundus data and annotations for efficient AI model training.
- To bridge the domain gap between regular and UWF fundus imaging using generative adversarial networks.
- To enhance the performance of AI models in diagnosing and analyzing various fundus diseases.
Main Methods:
- A modified Cycle Generative Adversarial Network (CycleGAN) was employed to generate synthetic UWF fundus images from regular fundus images.
- A consistency regularization term was incorporated into the GAN's loss function to enhance generated data quality.
- Pseudo-labeling techniques were used to ensure robustness against noise and errors in the generated unlabeled data.
Main Results:
- The proposed method effectively bridges the domain gap between regular and UWF fundus images without requiring paired data or identical semantic labels.
- Generated UWF images improved AI model performance on tasks including diabetic retinopathy classification, lesion detection, and tessellated fundus segmentation.
- The method demonstrated robustness to noise and errors in generated unlabeled data, validated through pseudo-labeling.
Conclusions:
- The modified CycleGAN approach successfully generates valuable UWF fundus images from regular ones, addressing data scarcity.
- This technique significantly improves the generalizability of learned representations and enhances performance across multiple ophthalmic diagnostic tasks.
- The method offers a convenient and efficient solution for AI model training in ophthalmology, reducing the reliance on extensive manual annotation.
