Related Experiment Video
Updated: May 1, 2026
11:45
Generation and 3-Dimensional Quantitation of Arterial Lesions in Mice Using Optical Projection Tomography
Published on: May 26, 2015
10.3K
Unsupervised quality assessment with generative adversarial networks for 3D OCTA microvascular imaging
Edmund Sumpena1,2, Andrew Cornelio2, Ana Collazo2
1Department of Computer Science, Johns Hopkins University, Baltimore, MD 21218, USA.
Biomedical Optics Express
|January 14, 2026
Summary
A new deep learning model, OCTA-GAN, automatically assesses optical coherence tomography angiography (OCTA) scan quality. This unsupervised 3D generative adversarial network effectively distinguishes high-quality from suboptimal scans, improving diagnostic accuracy.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Ophthalmology
Background:
- Optical coherence tomography angiography (OCTA) volumes can suffer from artifacts (e.g., eye movements, opacities) leading to suboptimal image quality.
- Existing quality assessment methods often use supervised 2D classifiers, which have limitations in generalization, data requirements, and loss of 3D information.
- There is a need for automated, quantitative, and objective methods for OCTA image quality assessment.
Purpose of the Study:
- To develop an automated deep learning model for distinguishing excellent-quality from suboptimal-quality OCTA volumes.
- To overcome limitations of existing 2D supervised methods by leveraging 3D information and unsupervised learning.
- To provide an objective and quantitative measure of OCTA scan quality.
Main Methods:
- Proposed OCTA-GAN, an efficient 3D generative adversarial network incorporating multi-scale processing layers.
- The unsupervised model learns quality patterns from excellent-quality OCTA volumes.
- Evaluated the model's ability to fuse fine vasculature details with anatomical context for quality assessment.
Main Results:
- OCTA-GAN's discriminator achieved an AUC of 0.92, sensitivity of 95.7%, and specificity of 76.6% in distinguishing scan quality.
- Outperformed a baseline 3D architecture (AUC=0.55) and demonstrated superior generalization across diverse artifacts compared to 2D supervised classifiers.
- The model's performance is attributed to the synergy between its generator and discriminator, capturing intricate vasculature features effectively.
Conclusions:
- OCTA-GAN provides an effective unsupervised approach for automated OCTA volume quality assessment.
- The 3D generative adversarial network architecture successfully leverages multi-scale features for robust quality evaluation.
- The model offers interpretable output scores reflecting the severity of artifacts, surpassing previous methods in generalization and data efficiency.
