SAN2N: Spatio-Angular Noise2Noise for Self-Supervised Diffusion MRI Denoising

This repository contains the official implementation of SAN2N, a self-supervised deep learning method for denoising Diffusion MRI (DMRI) data. SAN2N leverages spatio-angular redundancy in DMRI to construct paired noisy patches for training, eliminating the need for clean ground-truth data.

📌 Conference Version: This work is an extension of our 2024 ISBI conference paper, which received the **Best Poster Award Nomination**💥💥💥.

---

🧠 Overview

Problem Statement

DMRI is inherently affected by substantial noise, which reduces the precision and reliability of derived diffusion metrics. Traditional deep learning methods require noisy-clean image pairs for supervised training, which are unavailable in clinical practice.

Our Solution

SAN2N introduces a novel self-supervised framework that:

• Constructs angular neighbors in q-space based on geometric similarity
• Uses spatial sub-samplers to extract 3D patches with spatio-angular redundancy
• Trains a lightweight 3D CNN using a mixture denoising loss function

Key Features

• ✅ No clean data required for training
• ✅ Exploits both spatial and angular redundancy
• ✅ Edge-aware regularization and Rician likelihood loss
• ✅ Lightweight and efficient 3D CNN architecture

---

📊 Method Overview

Framework

Figure 1: Overall architecture of SAN2N.

🚀 Quick Start

Prerequisites

• Python 3.8+
• PyTorch 1.9+
• Nibabel
• scikit-learn
• DIPY

📁 Preprocessing

data/
├── subject1/
  ├── dwi.nii.gz   # Raw DMRI data
  ├── bvals        # b-values
  ├── bvecs        # b-vectors
  └── mask.nii.gz  # Brain mask

Generate training patches using the preprocessing script:

python get_data_final.py 

🧪 Training

python san2n.py 
  --train_sim1_dirs /path/to/train_patches1.npy # Path to the first set of noisy patches\
  --train_sim2_dirs /path/to/train_patches2.npy # Path to the second set of noisy patches\
  --mask_path /path/to/mask.nii.gz #Path to the estimated noise sigma map\
  --sigma_path /path/to/sigma.npy #Path to the estimated noise sigma map\

📈 Results

Figure 2: Qualitative denoising results on simulated DMRI data.

🙏 Acknowledgments

This work was supported by the National Natural Science Foundation of China
The conference version of this work was presented at ISBI

📧 Contact

For questions or suggestions, please contact:

Haotian Jiang: [email protected]
Geng Chen: [email protected]