SRNR: Training neural networks for Super-Resolution MRI using Noisy high-resolution Reference data

Jiaxin Xiao, Zihan Li, Berkin Bilgic̦, Jon̈athan R. Polimeni, Susie Y. Huang, Qiyuan Tian

Abstract: Neural network (NN) based approaches for super-resolution MRI typically require high-SNR high-resolution reference data acquired in many subjects, which is time consuming and a barrier to feasible and accessible implementation. We propose to train NNs for Super-Resolution using Noisy Reference data (SRNR), leveraging the mechanism of the classic NN-based denoising method Noise2Noise. We systematically demonstrate that results from NNs trained using noisy and high-SNR references are similar for both simulated and empirical data. SRNR suggests a smaller number of repetitions of high-resolution reference data can be used to simplify the training data preparation for super-resolution MRI.

Introduction

In magnetic resonance imaging (MRI), there is an inherent trade-off between spatial resolution, signal-to-noise ratio (SNR), and acquisition time. Super-resolution (SR) techniques aim to recover high-resolution (HR) details from low-resolution (LR) inputs, often utilizing deep learning models. However, training these models typically requires a "ground truth" dataset of high-quality, high-SNR images. Obtaining such "clean" reference data often necessitates multiple scan repetitions and averaging, which is prohibitively slow and expensive. This study introduces SRNR (Super-Resolution using Noisy Reference), demonstrating that high-quality SR networks can be trained effectively using noisy, single-scan reference data.

Methods and Materials

The SRNR approach is inspired by the Noise2Noise principle, which posits that a network can learn to denoise images even if it only sees noisy examples, provided the noise is random and zero-mean.

Architecture

The researchers employed a Modified U-Net (MU-Net). Unlike standard U-Net architectures that use pooling and up-sampling layers, the MU-Net maintains spatial detail throughout the network by omitting these operations. The network is trained to predict the residual—the difference between a blurry, up-sampled LR input and a sharp, albeit noisy, HR target.

The MU-Net architecture (a) and examples of HCP training targets with varying noise levels (b). The network learns to predict the residual image volume to produce a high-resolution output.

Datasets

The method was validated using two datasets:

  1. Simulated Data: T1-weighted images from 30 subjects in the Human Connectome Project (HCP) were down-sampled and corrupted with varying levels of synthetic Gaussian noise to test the algorithm's limits.
  2. Empirical Data: Nine healthy subjects were scanned using a Wave-MPRAGE sequence. For each subject, one LR scan and ten HR repetitions were acquired, allowing for a direct comparison between models trained on a single noisy scan versus a 10-scan average.

Results

Quantitative results demonstrated that models trained on noisy reference data performed nearly as well as those trained on clean, averaged data.

Quantitative Performance

In simulated HCP data, the Peak Signal-to-Noise Ratio (PSNR) only decreased from 30.26 dB (clean training) to 29.43 dB even when the training target noise was 1.6 times the standard deviation of brain intensities. Structural Similarity Index (SSIM) scores remained consistently high (above 0.92), indicating that essential anatomical structures were preserved regardless of training noise levels.

Quantitative metrics (MAE, PSNR, SSIM) across different noise standard deviations in the training target. The boxplots and tables show that performance remains robust even as target noise increases.

Real-world Validation

On empirical data, the model trained on a single noisy repetition achieved a PSNR of 30.44 dB, while the model trained on the 10-scan average achieved 30.90 dB. This negligible difference suggests that the network performs implicit denoising and super-resolution simultaneously, as random noise does not map consistently across training samples.

Discussion and Conclusion

SRNR proves that "perfect" high-SNR reference data is not a prerequisite for training effective super-resolution models. By leveraging single-repetition scans, the time required for data acquisition can be reduced by a factor of 10 or more, significantly lowering the barrier to developing AI-driven imaging tools.

While the study assumes noise is largely random and zero-mean, future work should explore the impact of systematic artifacts and validate the method across other modalities like Diffusion or Functional MRI. Ultimately, SRNR provides a practical pathway for more accessible and efficient super-resolution MRI in clinical and research settings.

Summary updated

Abstract

Neural network (NN) based approaches for super-resolution MRI typically require high-SNR high-resolution reference data acquired in many subjects, which is time consuming and a barrier to feasible and accessible implementation. We propose to train NNs for Super-Resolution using Noisy Reference data (SRNR), leveraging the mechanism of the classic NN-based denoising method Noise2Noise. We systematically demonstrate that results from NNs trained using noisy and high-SNR references are similar for both simulated and empirical data. SRNR suggests a smaller number of repetitions of high-resolution reference data can be used to simplify the training data preparation for super-resolution MRI.
BibTeX
@inproceedings{xiao2024srnr,
  author = {Jiaxin Xiao and Zihan Li and Berkin Bilgic̦ and Jon̈athan R. Polimeni and Susie Y. Huang and Qiyuan Tian},
  title = {{SRNR: Training neural networks for Super-Resolution MRI using Noisy high-resolution Reference data}},
  booktitle = {Proceedings on CD-ROM - International Society for Magnetic Resonance in Medicine. Scientific Meeting and Exhibition/Proceedings of the International Society for Magnetic Resonance in Medicine, Scientific Meeting and Exhibition},
  year = {2024},
  doi = {10.58530/2023/1370},
}
Figure