🧠 XOCT: Enhancing OCT to OCTA Translation via Cross-Dimensional Multi-Scale Feature Learning

🩺 Overview

XOCT is a deep learning framework for translating Optical Coherence Tomography (OCT) volumes into Optical Coherence Tomography Angiography (OCTA), achieving high-fidelity reconstruction of retinal microvasculature.

This model integrates:

Cross-Dimensional Supervision (CDS): segmentation-weighted en-face projections for layer-specific learning.
Multi-Scale Feature Fusion (MSFF): isotropic, anisotropic, and large-kernel 3D convolutions for fine vascular reconstruction.
Layer-aware vascular coherence: ensuring anatomically consistent vasculature across retinal depths.

Presented at MICCAI 2025, XOCT achieves state-of-the-art performance on the OCTA-500 dataset, significantly improving vessel continuity and detail preservation.

🧩 Checkpoint Structure

This repository includes pretrained weights for both 3M and 6M subsets of the OCTA-500 dataset.
Each model includes a generator (G) and multiple discriminators (D) specialized for layer-wise supervision.


checkpoints_3M/
└── xoct_gn_3M/
    ├── latest_net_D_3d.pth         # 3D volumetric discriminator
    ├── latest_net_D_ilm_opl.pth    # En-face discriminator (ILM–OPL)
    ├── latest_net_D_opl_bm.pth     # En-face discriminator (OPL–BM)
    ├── latest_net_D_other.pth      # Auxiliary 2D discriminator
    └── latest_net_G.pth            # Generator weights (XOCT model)

checkpoints_6M/
└── xoct_gn_6M/
    ├── latest_net_D_3d.pth
    ├── latest_net_D_ilm_opl.pth
    ├── latest_net_D_opl_bm.pth
    ├── latest_net_D_other.pth
    └── latest_net_G.pth

⚙️ Environment Setup

Create and activate the environment:

git clone https://github.com/uci-cbcl/XOCT.git
cd XOCT
conda create -n xoct python=3.10
conda activate xoct
pip install -r requirements.txt

🚀 Usage

🧠 Load Pretrained Generator (Python)

import torch
from models.xoct_model import XOCTModel

# Load pretrained generator (example for 3M)
gen = XOCTModel()
weights = torch.load("checkpoints_3M/xoct_gn_3M/latest_net_G.pth", map_location="cpu")
gen.load_state_dict(weights)
gen.eval()

# x: 3D OCT volume tensor (1 × D × H × W)
with torch.no_grad():
    octa_pred = gen(x)

To generate 2D en-face projections:

python create_2d_projection.py --input octa_pred.nii --output projection.png

🔍 Testing & Inference (Pretrained Checkpoints)

Pretrained checkpoints for both 3M and 6M subsets are located under:

checkpoints_3M/xoct_gn_3M/
checkpoints_6M/xoct_gn_6M/

Each includes:

Generator (latest_net_G.pth)
3D and layer-wise discriminators (latest_net_D_*.pth)

To run inference:

▶ For the OCTA-3M subset

bash scripts_3M/xoct/test.sh

▶ For the OCTA-6M subset

bash scripts_6M/xoct/test.sh

Example test3d_6M.py command (inside scripts_6M/xoct/test.sh):

python test3d_6M.py \
    --dataroot ./octa-500/OCT2OCTA6M_3D \
    --checkpoints_dir ./checkpoints_6M \
    --name xoct_gn_6M \
    --test_name xoct_gn_6M \
    --model XOCT \
    --add_scale \
    --direction AtoB \
    --base_ch_g 64 \
    --max_ch_g 512 \
    --dataset_mode alignedoct2octaall \
    --preprocess none \
    --norm group \
    --batch_size 1 \
    --input_nc 1 \
    --output_nc 1 \
    --gpu_ids 1 \
    --num_test 1000000 \
    --which_epoch latest

This script:

Loads the XOCT model architecture.
Uses pretrained weights from checkpoints_6M/xoct_gn_6M/.
Runs inference on 3D OCT volumes in ./octa-500/OCT2OCTA6M_3D/.
Saves volumetric predictions and en-face projections to:
```
./results/xoct_gn_6M/test_latest/
```

🧩 Tips:

Use --gpu_ids -1 for CPU inference.
Replace --dataroot with your own OCT dataset path.
For specific checkpoints, set --which_epoch best if available.

📊 Performance Summary

Dataset	MAE ↓	PSNR ↑	SSIM ↑	Perceptual ↓
OCTA-3M	19.22	20.21	0.608	0.573
OCTA-6M	16.65	21.46	0.568	0.541

XOCT outperforms 2D and 3D baselines, preserving vessel continuity and microvascular detail essential for retinal pathology interpretation.

📚 Citation

If you use these models or code, please cite:

@inproceedings{khosravi2025xoct,
  title={XOCT: Enhancing OCT to OCTA Translation via Cross-Dimensional Supervised Multi-scale Feature Learning},
  author={Khosravi, Pooya and Han, Kun and Wu, Anthony T and Rezvani, Arghavan and Feng, Zexin and Xie, Xiaohui},
  booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
  pages={695--705},
  year={2025},
  organization={Springer}
}

🙏 Acknowledgments

This work builds upon:

Funding: National Eye Institute, NIH F30 EY036725 (P.K.)
Institution: University of California, Irvine
Authors: Pooya Khosravi *, Kun Han *, Anthony T. Wu, Arghavan Rezvani, Zexin Feng, Xiaohui Xie
(*Equal contribution)

Downloads last month: -; Downloads are not tracked for this model. How to track