VAMOS-OCTA: Vessel-Aware Multi-Axis Orthogonal Supervision for Inpainting Motion-Corrupted OCT Angiography Volumes

Handheld OCT Angiography (OCTA) enables retinal imaging in uncooperative patients but suffers from severe motion artifacts, including fully corrupted B-scans. VAMOS-OCTA is a deep learning framework for inpainting motion-corrupted B-scans in 3D OCT Angiography (OCTA) volumes, enabling volumetric reconstruction in challenging handheld imaging scenarios. Our method restores missing slices while preserving vessel structures across both cross-sectional and projection views.

Key Features:

• 2.5D U-Net inpainting model conditioned on surrounding slices
• VAMOS loss combining vessel-weighted MSE with orthogonal projection supervision (axial + lateral)
• Dynamic synthetic corruption pipeline simulates realistic B-scan dropouts during training
• Evaluation across multiple metrics targeting both pixel-level accuracy and perceptual quality

Installation

git clone https://github.com/MedICL-VU/VAMOS-OCTA.git
cd VAMOS-OCTA
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Usage

Prepare data:

OCTA volumes must be 3D .tif stacks of shape (D, H, W) and type uint16. Place all volumes in a folder such as:

data
├── volume1_corrupted.tif
├── volume1_gt.tif
├── volume1_mask.tif
├── volume2_corrupted.tif
├── .

Train and test a model:

python main.py --data_dir data/train --epochs 100 --stride 1 --kfold

A complete list of flags is available in main.py.

Evaluate a previously-trained model:

python main.py --skip_train

Evaluation Metrics

VAMOS-OCTA uses both pixel-wise accuracy and perceptual quality metrics:

B-scan Metrics

• Gradient L1
• LPIPS
• Laplacian Blur Diff
• Edge Preservation Ratio
• PSNR

Projection Metrics (MIPs)

• L1
• MIE (Mean Intensity Error)
• SSIM
• NCC
• PSNR

Contact

For questions or feedback, please contact Nick DiSanto

License

This repository is licensed under the MIT License.