Benchmarking Deep Learning Approaches for Single-Excitation Wideband MRI: Establishing the First SE-Wideband Brain Dataset

Tzu‐Yi Wang, Yu Tsao, Tsung-Ren Huang, Jyh‐Horng ChenResearch Square 2025 · DOI

Abstract

Single-Excitation Wideband MRI enables efficient acquisition, but its rapid-switching design to mitigate intrinsic blur leads to incompatibility with GRAPPA and other fast sequences, restricting its clinical adoption. Regarding this challenge, we established the first dedicated SE-Wideband brain dataset, comprising 76 subjects and 8000 paired slices, and developed a systematic framework for benchmarking deep learning–based restoration methods. Nine networks were evaluated across multiple loss configurations, with performance quantified using conventional metrics and human perceptual ratings. Results demonstrate consistent improvements over raw SE-Wideband inputs, improving NRMSE from ~ 8% to ~ 5% and SSIM from 0.62 to 0.80. Human ratings verified the effectiveness of AI-based denoising, and correlation analyses showed that metric-optimized configurations generally aligned better with subjective scores than perceptual-guided ones, suggesting that conventional metrics remain a reliable reference for loss selection. In contrast, SE-Wideband-GRAPPA restoration proved challenging, yielding limited gains across networks. These findings highlight both the promise and limitations of AI-based SE-Wideband image restoration: substantial improvements are achievable for SE-Wideband, but clinical translation will require new strategies tailored to compounded artifacts in SE-GRAPPA. The curated dataset and benchmarking framework presented here provide a valuable reference for future methodological development and the broader integration of Wideband MRI into neuroimaging practice.

Introduction

Magnetic Resonance Imaging (MRI) is a vital diagnostic tool, yet its clinical utility is often constrained by long scan times. To address this, researchers have developed "Wideband MRI," a technique that accelerates acquisition by exciting multiple brain slices simultaneously. While early versions suffered from geometric distortions known as "voxel shearing," the newer Single-Excitation (SE) Wideband MRI variant corrected these distortions at the cost of introducing significant image blurring and "ringing" artifacts.

Furthermore, SE-Wideband has remained largely incompatible with standard parallel imaging acceleration methods like GRAPPA. This study seeks to overcome these barriers by establishing the first dedicated SE-Wideband brain dataset and benchmarking deep learning architectures to restore image quality, potentially enabling faster scans without sacrificing diagnostic detail.

Methodology and Framework

The researchers approached the restoration of SE-Wideband images as an image-to-image translation task. They evaluated nine different deep learning architectures, ranging from established convolutional neural networks (CNNs) like U-Net and ResDnCNN to modern, lightweight Transformer-based models such as NAFNet-Lite and Uformer-Lite.

Overview of the SE-Wideband MRI acquisition and deep learning restoration framework.

A central focus of the study was Loss Configuration Optimization. Beyond standard pixel-wise error (L1 loss), the team tested combinations of Structural Similarity (SSIM), Gradient Loss (to target high-frequency edges), and Perceptual Loss (to mimic human visual judgment). The final results were validated through both quantitative metrics and a human perceptual evaluation conducted by nine MRI experts.

The First SE-Wideband Brain Dataset

A primary contribution of this research is the creation of a standardized dataset to facilitate future AI development in this field:

  • Scope: 76 healthy adult subjects.
  • Volume: Approximately 8,000 paired image slices.
  • Data Tiers: Each subject provided three distinct scan types: high-quality standard MPRAGE (Ground Truth), 2x accelerated SE-Wideband (blurred), and ~3.4x accelerated SE-Wideband-GRAPPA (compounded artifacts).

Key Insights and Results

  • Loss Functions Drive Quality: The choice of loss function proved more influential than the specific neural network architecture. Adding Gradient Loss was particularly effective, acting as a "silver bullet" to recover crisp anatomical edges and suppress Gibbs artifacts.
  • Quantitative Performance: Deep learning models successfully reduced the Normalized Root Mean Square Error (NRMSE) from 8% to 5% and improved the Structural Similarity Index (SSIM) from 0.62 to 0.80. Models such as ResDnCNN, DRUNet, and IRCNN were identified as the most robust performers.
  • The "Compounded Artifact" Challenge: While AI significantly improved standard Wideband images, it struggled with the complex noise patterns found in SE-Wideband-GRAPPA. These compounded artifacts remain a significant hurdle for image-domain models.
  • Metric Reliability: The study found that conventional metrics like SSIM and MS-SSIM correlated well with human expert ratings, confirming their utility as reliable benchmarks for MRI restoration tasks.

Conclusion and Future Directions

While deep learning provides a powerful tool for cleaning up SE-Wideband images, clinical translation will require overcoming the limitations of current image-domain models. Future work will likely focus on "hybrid-domain" strategies—incorporating raw k-space data and MRI physics—to address the compounded artifacts seen in highly accelerated sequences. The release of this dataset provides a critical foundation for the broader integration of Wideband MRI into neuroimaging practice.

Reference: Wang, T.Y., et al. "Benchmarking Deep Learning Approaches for Single-Excitation Wideband MRI: Establishing the First SE-Wideband Brain Dataset." Research Square Preprint, 2025.

Summary updated

Abstract

<title>Abstract</title> Single-Excitation Wideband MRI enables efficient acquisition, but its rapid-switching design to mitigate intrinsic blur leads to incompatibility with GRAPPA and other fast sequences, restricting its clinical adoption. Regarding this challenge, we established the first dedicated SE-Wideband brain dataset, comprising 76 subjects and 8000 paired slices, and developed a systematic framework for benchmarking deep learning–based restoration methods. Nine networks were evaluated across multiple loss configurations, with performance quantified using conventional metrics and human perceptual ratings. Results demonstrate consistent improvements over raw SE-Wideband inputs, improving NRMSE from ~ 8% to ~ 5% and SSIM from 0.62 to 0.80. Human ratings verified the effectiveness of AI-based denoising, and correlation analyses showed that metric-optimized configurations generally aligned better with subjective scores than perceptual-guided ones, suggesting that conventional metrics remain a reliable reference for loss selection. In contrast, SE-Wideband-GRAPPA restoration proved challenging, yielding limited gains across networks. These findings highlight both the promise and limitations of AI-based SE-Wideband image restoration: substantial improvements are achievable for SE-Wideband, but clinical translation will require new strategies tailored to compounded artifacts in SE-GRAPPA. The curated dataset and benchmarking framework presented here provide a valuable reference for future methodological development and the broader integration of Wideband MRI into neuroimaging practice.
BibTeX
@inproceedings{wang2025benchmarking,
  author = {Tzu‐Yi Wang and Yu Tsao and Tsung-Ren Huang and Jyh‐Horng Chen},
  title = {{Benchmarking Deep Learning Approaches for Single-Excitation Wideband MRI: Establishing the First SE-Wideband Brain Dataset}},
  booktitle = {Research Square},
  year = {2025},
  doi = {10.21203/rs.3.rs-8024132/v1},
}
Figure