Deep learning model restores 3D retinal blood vessels with 51% fidelity boost
A new algorithm recovers hidden capillary structures from single OCTA scans using AI...
A new deep learning algorithm restores three-dimensional retinal microvasculature from a single OCT angiography volume. Using an EfficientNet-B5 encoder with concurrent spatial and channel squeeze-and-excitation modules, the model takes three adjacent B-frames as input to predict the restored middle B-frame. Compared to ground truth from averaged multiple scans, the model significantly improved image quality: PSNR rose from 22.23 ± 0.78 to 26.16 ± 1.26, and SSIM from 0.72 ± 0.03 to 0.91 ± 0.02 (both p < 0.001). Microvascular fidelity, measured by Dice coefficient overlap with ground truth, improved by at least 3.8% in 2D and 51.2% in 3D across multiple vascular slabs.
- Model uses EfficientNet-B5 encoder with squeeze-and-excitation attention; processes three adjacent B-frames to restore 3D microvasculature from single OCTA volumes.
- PSNR improved from 22.23 to 26.16 (+17.7%); SSIM from 0.72 to 0.91 (+26.4%); both p < 0.001.
- Dice coefficient for microvascular fidelity increased 3.8% in 2D and 51.2% in 3D, enabling accurate detection of nonperfusion and capillary dropout.
Why It Matters
Enables high-fidelity 3D vessel maps from single OCTA scans, improving early diagnosis of blinding eye diseases.