xQSM: quantitative susceptibility mapping with octave convolutional and noise‐regularized neural networks

Yang Gao, Xuanyu Zhu, Bradford A. Moffat, Rebecca Glarin, Alan H. Wilman, G. Bruce Pike, ‪Stuart Crozier‬, Feng Liu, Hongfu SunNMR 2020 · DOI

TL;DR

This paper introduces xQSM, a novel deep learning framework for Quantitative Susceptibility Mapping (QSM) that leverages octave convolutions and noise regularization within a U-net architecture. xQSM significantly improves reconstruction accuracy, particularly in deep gray matter regions, and offers remarkable robustness across various resolutions and susceptibility ranges, all while drastically reducing computation time compared to traditional methods.

What Problem Does This Paper Address?

Quantitative Susceptibility Mapping (QSM) is a powerful MRI technique for visualizing and quantifying magnetic susceptibility in biological tissues, offering valuable insights into neurological disorders. However, reconstructing QSM images is notoriously challenging due to the ill-posed nature of the dipole inversion problem. This means that small errors or noise in the input data can lead to significant artifacts and inaccuracies in the final susceptibility maps. Existing methods, both conventional iterative approaches and prior deep learning models, struggle with issues like streaking artifacts, noise amplification, susceptibility underestimation in deep gray matter (DGM), and limited generalization across different image resolutions and susceptibility ranges.

Methods

The core innovation of this work is the xQSM framework, built upon a U-net architecture. Key methodological advancements include:

  • Modified Octave Convolutional Layers (OctConv): Instead of standard convolutional layers, xQSM employs OctConv layers. These layers explicitly factorize feature maps into high and low spatial resolution groups, allowing for efficient multi-scale representation learning and better communication between different resolution levels. This is particularly beneficial for handling the varying spatial scales present in MRI data and improving robustness to resolution mismatches.
  • Noise Adding Layer: A custom noise adding layer is integrated into the training process. This layer introduces controlled levels of Gaussian noise to the training inputs, making the network more robust to noise and measurement deviations encountered in real-world in vivo data. This helps suppress streaking artifacts and improve the signal-to-noise ratio (SNR) of the reconstructed QSM maps.
  • U-net Backbone with Enhancements: The U-net architecture provides a strong foundation for image-to-image translation tasks. xQSM retains the contracting and expanding paths, skip connections for information flow, and residual blocks for faster convergence.

Datasets

The study utilized a combination of datasets for training and evaluation:

  • In Vivo Datasets: QSM data from 90 healthy subjects acquired at 1 mm isotropic resolution (3T) were used. These were processed using the LN-QSM method.
  • Synthetic Datasets:
    • Digital phantoms with basic geometric shapes (spheres, squares, rectangles) were generated, consistent with previous work (DeepQSM).
  • A 3D Shepp-Logan phantom was constructed with various ellipsoids to test generalization capabilities.
  • Data from the 2019 QSM Challenge 2.0 were also used, including simulated COSMOS maps and local field maps with varying contrast and noise levels.

The authors trained two main versions of their network: xQSMinvivo (trained on in vivo data) and xQSMsynthetic (trained on synthetic data).

Key Insights and Ideas

  • Octave Convolutions for Multi-Scale Robustness: The explicit factorization of feature maps into high and low-resolution groups within OctConv layers is crucial for handling variations in image resolution and improving generalization. This multi-scale approach allows the network to better capture both fine details and broader contextual information.
  • Noise Regularization is Key for In Vivo Data: The noise adding layer significantly improves the performance of the network on real-world data by making it more resilient to noise and artifacts, which are often problematic in QSM reconstruction.
  • In Vivo Training Data Yields Superior Results: While synthetic data is useful, training with actual in vivo data (xQSMinvivo) leads to demonstrably better reconstructions, especially in complex biological tissues like deep gray matter.
  • Addressing DGM Underestimation: The combination of OctConv and in vivo training effectively mitigates the susceptibility underestimation in deep gray matter regions, a persistent issue in previous deep learning QSM methods.

Key Results

  • Accuracy Improvements:
    • xQSMinvivo achieved the best reconstructions among all deep learning methods tested.
  • Compared to DeepQSM, QSMnet⁺, and iLSQR, xQSMinvivo showed an average improvement of 32.3%, 25.4%, and 11.7% in globus pallidus susceptibility estimation for digital simulations, respectively.
  • For in vivo acquisitions, improvements were even more substantial: 39.3% over DeepQSM, 21.8% over QSMnet⁺, and 6.3% over iLSQR.
  • On the 2019 QSM Challenge datasets, xQSMinvivo achieved the best PSNR and SSIM, with noticeable underestimation only in the DGM region (average 8.0%), significantly less than other methods.
  • Robustness:
    • xQSMinvivo demonstrated the most robust generalization capability across various spatial resolutions, outperforming other deep learning methods.
  • It maintained comparable image contrast to COSMOS at 0.6 mm resolution without significant susceptibility contrast loss.
  • The method showed high linearity against different susceptibility intensity scales.
  • Speed:
    • xQSM significantly reduced reconstruction time, from minutes for traditional methods like MEDI (over 1000 seconds) to just a few seconds (e.g., 4.2 seconds for a 224x304x224 image).

Limitations and Future Work

  • Noise Distribution Mismatch: The noise adding layer currently uses normally distributed noise with limited predefined levels. This might not perfectly replicate real-world noise distributions, potentially leading to slight underperformance in certain scenarios (e.g., compared to MEDI on 2019 QSM Challenge data due to noise mismatch).
  • Future Work: The authors suggest exploring more realistic and comprehensive noise distributions for the noise adding layer and developing more advanced training dataset designs to further enhance the framework's robustness across a wider range of noise levels and achieve even more accurate QSM dipole inversion results.

Summary updated

Abstract

Quantitative susceptibility mapping (QSM) provides a valuable MRI contrast mechanism that has demonstrated broad clinical applications. However, the image reconstruction of QSM is challenging due to its ill-posed dipole inversion process. In this study, a new deep learning method for QSM reconstruction, namely xQSM, was designed by introducing noise regularization and modified octave convolutional layers into a U-net backbone and trained with synthetic and in vivo datasets, respectively. The xQSM method was compared with two recent deep learning (QSMnet<sup>+</sup> and DeepQSM) and two conventional dipole inversion (MEDI and iLSQR) methods, using both digital simulations and in vivo experiments. Reconstruction error metrics, including peak signal-to-noise ratio, structural similarity, normalized root mean squared error and deep gray matter susceptibility measurements, were evaluated for comparison of the different methods. The results showed that the proposed xQSM network trained with in vivo datasets achieved the best reconstructions of all the deep learning methods. In particular, it led to, on average, 32.3%, 25.4% and 11.7% improvement in the accuracy of globus pallidus susceptibility estimation for digital simulations and 39.3%, 21.8% and 6.3% improvements for in vivo acquisitions compared with DeepQSM, QSMnet<sup>+</sup> and iLSQR, respectively. It also exhibited the highest linearity against different susceptibility intensity scales and demonstrated the most robust generalization capability to various spatial resolutions of all the deep learning methods. In addition, the xQSM method also substantially shortened the reconstruction time from minutes using MEDI to only a few seconds.
BibTeX
@inproceedings{gao2020xqsm,
  author = {Yang Gao and Xuanyu Zhu and Bradford A. Moffat and Rebecca Glarin and Alan H. Wilman and G. Bruce Pike and ‪Stuart Crozier‬ and Feng Liu and Hongfu Sun},
  title = {{xQSM: quantitative susceptibility mapping with octave convolutional and noise‐regularized neural networks}},
  booktitle = {NMR},
  year = {2020},
  doi = {10.1002/nbm.4461},
}

Loading related content...

Figure