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

Overview

XOCT is a deep learning framework for OCT → OCTA translation that integrates Cross-Dimensional Supervision (CDS) and a Multi-Scale Feature Fusion (MSFF) architecture to improve retinal vascular reconstruction.
It introduces layer-aware guidance and multi-scale contextual fusion, enabling sharper vessel delineation and better preservation of fine microvascular structures across heterogeneous retinal layers.

Key Highlights

• Cross-Dimensional Supervision (CDS) — leverages segmentation-weighted en-face projections to enforce layer-specific learning and structural coherence.
• Multi-Scale Feature Fusion (MSFF) — captures vessel details across scales using isotropic, anisotropic, and large-kernel convolutions with adaptive channel reweighting.
• Demonstrated state-of-the-art performance on the OCTA-500 dataset, outperforming 2D, projection-based, and volumetric baselines.

📘 Reference

Paper:
Khosravi P., Han K., Wu A.T., Rezvani A., Feng Z., Xie X.
XOCT: Enhancing OCT to OCTA Translation via Cross-Dimensional Supervised Multi-Scale Feature Learning.
In MICCAI 2025, pp. 695–705. Springer.
[PDF] | [arXiv] | 📄 Springer Proceedings

---

🧠 Repository Structure

├── data/                # Data preparation scripts
├── models/              # Model architectures (3D UNet, CDS, MSFF)
├── options/             # Training and evaluation configurations
├── scripts_3M/          # Training/evaluation scripts for OCTA-3M
│   └── xoct/
│       ├── train.sh
│       ├── test.sh
│       └── eval.sh
├── scripts_6M/          # Scripts for OCTA-6M dataset
├── unet/                # Base 2D/3D UNet modules
├── util/                # Utility functions (data loaders, metrics, plotting)
├── create_2d_projection.py
├── eval2d.py / eval3d.py
├── train2d.py / train3d.py
├── test2d_3M.py / test2d_6M.py / test3d_3M.py / test3d_6M.py
└── README.md

---

⚙️ Environment Setup

Create and activate the environment:

conda create -n xoct python=3.11
conda activate xoct
pip install -r requirements.txt

---

🚀 Training & Evaluation

1. Training

To train XOCT on the OCTA-3M subset:

bash scripts_3M/xoct/train.sh

For the OCTA-6M subset:

bash scripts_6M/xoct/train.sh

2. Testing and Evaluation

bash scripts_3M/xoct/test.sh
bash scripts_3M/xoct/eval.sh

---

📦 Model Checkpoints

Pretrained XOCT model weights are available on Hugging Face:

➡️ https://huggingface.co/pooyakhosravi/xoct

---

📊 Dataset

The model is trained and evaluated on the OCTA-500 dataset, which includes paired OCT/OCTA volumes and retinal layer segmentation annotations (3 mm × 3 mm and 6 mm × 6 mm subsets).

---

🙏 Acknowledgments

This implementation builds upon and extends prior work:

• TransPro (Li et al., 2024)
• Pix2Pix (Isola et al., CVPR 2017)

If you find this work useful, please consider citing:

@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}
}