A Deep Learning-Based Method for Metal Artifact-Resistant Syn-MP-RAGE Contrast Synthesis

Ziyi Zeng, Yuhao Wang, Dianlin Hu, T. Michael O’Shea, Rebecca C. Fry, Jing Cai, Lei ZhangarXiv 2024 · DOI · arXiv

Abstract

In certain brain volumetric studies, synthetic T1-weighted magnetization-prepared rapid gradient-echo (MP-RAGE) contrast, derived from quantitative T1 MRI (T1-qMRI), proves highly valuable due to its clear white/gray matter boundaries for brain segmentation. However, generating synthetic MP-RAGE (syn-MP-RAGE) typically requires pairs of high-quality, artifact-free, multi-modality inputs, which can be challenging in retrospective studies, where missing or corrupted data is common. To overcome this limitation, our research explores the feasibility of employing a deep learning-based approach to synthesize syn-MP-RAGE contrast directly from a single channel turbo spin-echo (TSE) input, renowned for its resistance to metal artifacts. We evaluated this deep learning-based synthetic MP-RAGE (DL-Syn-MPR) on 31 non-artifact and 11 metal-artifact subjects. The segmentation results, measured by the Dice Similarity Coefficient (DSC), consistently achieved high agreement (DSC values above 0.83), indicating a strong correlation with reference segmentations, with lower input requirements. Also, no significant difference in segmentation performance was observed between the artifact and non-artifact groups.

Introduction

In neuroimaging, T1-weighted magnetization-prepared rapid gradient-echo (MP-RAGE) sequences are essential for brain volumetric analysis due to their superior contrast between gray and white matter. However, these sequences are highly susceptible to magnetic field inhomogeneities caused by metallic objects, such as dental braces or implants. These artifacts often manifest as significant signal loss or "black holes" in critical areas like the prefrontal cortex, rendering automated segmentation tools like FreeSurfer ineffective.

The DL-Syn-MPR Approach

Traditional methods for generating synthetic MP-RAGE (syn-MP-RAGE) rely on mathematical Bloch simulations using quantitative MRI (qMRI) data. While effective, these methods require multiple perfectly aligned, high-quality inputs (PD-w, T2-w, and T1-w scans). In clinical practice, such multi-modality datasets are frequently incomplete or corrupted by motion.

To address these challenges, the authors developed DL-Syn-MPR, a deep learning-based framework designed to synthesize high-contrast MP-RAGE images from a single-channel turbo spin-echo (TSE) input. TSE sequences are inherently more resistant to metal artifacts, making them ideal candidates for synthesis in patients with dental hardware.

Overview of the synthesis workflow comparing traditional mathematical mapping with the proposed deep learning method.

Network Architecture: KAN-UNET

The core of the DL-Syn-MPR method is a conditional Generative Adversarial Network (cGAN) that utilizes a novel KAN-UNET generator. Unlike standard convolutional neural networks that use fixed activation functions, Kolmogorov-Arnold Networks (KANs) employ learnable B-spline functions on the network edges. This allows for more flexible and accurate approximation of the complex non-linear mappings required for medical image contrast synthesis.

The architecture integrates "Tok-KAN" modules—tokenized KAN blocks—into the bottleneck of a UNET structure. This hybrid approach combines the spatial feature extraction capabilities of convolutions with the high-order representation power of KANs.

Detailed architecture of the KAN-UNET generator, discriminator, and the overall cGAN training workflow.

Experimental Results

The model was trained on 51 artifact-free subjects and tested on 44 subjects, including 11 with significant metal artifacts.

  • Performance Metrics: The KAN-UNET achieved a Peak Signal-to-Noise Ratio (PSNR) of up to 30.39 dB and a Structural Similarity Index (SSIM) of 0.976, significantly outperforming standard UNET and CycleGAN baselines.
  • Segmentation Fidelity: Brain segmentation using the synthesized images reached a Dice Similarity Coefficient (DSC) above 0.83 across seven key brain regions.
  • Artifact Resilience: Most importantly, the segmentation performance showed no significant difference between the artifact and non-artifact groups. The DL-Syn-MPR effectively "filled in" the signal loss areas caused by metal, restoring anatomical detail necessary for volumetric analysis.

Conclusion

DL-Syn-MPR provides a robust solution for brain volumetric studies in the presence of metal artifacts. By leveraging the representational power of Kolmogorov-Arnold Networks and the inherent artifact resistance of TSE sequences, this method enables high-quality brain segmentation even when traditional MP-RAGE scans are unusable. Future work will focus on extending the model to 3D architectures to further improve volumetric consistency.

Summary updated

Abstract

In certain brain volumetric studies, synthetic T1-weighted magnetization-prepared rapid gradient-echo (MP-RAGE) contrast, derived from quantitative T1 MRI (T1-qMRI), proves highly valuable due to its clear white/gray matter boundaries for brain segmentation. However, generating synthetic MP-RAGE (syn-MP-RAGE) typically requires pairs of high-quality, artifact-free, multi-modality inputs, which can be challenging in retrospective studies, where missing or corrupted data is common. To overcome this limitation, our research explores the feasibility of employing a deep learning-based approach to synthesize syn-MP-RAGE contrast directly from a single channel turbo spin-echo (TSE) input, renowned for its resistance to metal artifacts. We evaluated this deep learning-based synthetic MP-RAGE (DL-Syn-MPR) on 31 non-artifact and 11 metal-artifact subjects. The segmentation results, measured by the Dice Similarity Coefficient (DSC), consistently achieved high agreement (DSC values above 0.83), indicating a strong correlation with reference segmentations, with lower input requirements. Also, no significant difference in segmentation performance was observed between the artifact and non-artifact groups.
BibTeX
@misc{zeng2024deep,
  author = {Ziyi Zeng and Yuhao Wang and Dianlin Hu and T. Michael O’Shea and Rebecca C. Fry and Jing Cai and Lei Zhang},
  title = {{A Deep Learning-Based Method for Metal Artifact-Resistant Syn-MP-RAGE Contrast Synthesis}},
  year = {2024},
  eprint = {2410.16817},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2410.16817},
}
Figure