Suppressing motion artefacts in MRI using an Inception‐ResNet network with motion simulation augmentation

Kamlesh Pawar, Zhaolin Chen, N. Jon Shah, Gary F. EganNMR 2019 · DOI

TL;DR

This paper introduces a novel deep learning approach, MoCoNet, to automatically correct motion artifacts in 3D MPRAGE MRI scans. By training a convolutional neural network on simulated motion-corrupted data, MoCoNet effectively removes artifacts without altering acquisition parameters, offering a promising post-processing solution for improved image quality.

What Problem Does This Paper Address?

Magnetic Resonance Imaging (MRI) is a powerful tool for visualizing soft tissues, but its slow acquisition speed makes it susceptible to patient motion. Even small movements during scanning can introduce significant artifacts (like blurring, ringing, and ghosting) into the final images, compromising diagnostic accuracy. Existing motion correction methods often require modifications to the MRI acquisition protocol (e.g., using navigators) or external hardware, which can be complex, time-consuming, and may negatively impact image quality metrics like SNR or contrast. This paper aims to address this challenge by developing a standalone, data-driven post-processing technique that can effectively remove motion artifacts without interfering with the standard MRI workflow.

Methods

The core of this work is a deep learning model based on an encoder-decoder Convolutional Neural Network (CNN) architecture, specifically a U-Net variant. The encoder part of the network processes the motion-corrupted MRI image, breaking it down into a series of feature maps that capture different aspects of the image. The decoder then takes these feature maps and reconstructs a motion-corrected image.

A key innovation is the creation of a motion simulation framework. This framework generates realistic motion-corrupted images by taking motion-free 3D MPRAGE datasets and artificially introducing simulated 3D rigid body random motion (translations and rotations). This allows for the creation of a large, controlled dataset for training the CNN. The network is trained to minimize the difference (mean squared error) between its output and the original motion-free images.

The trained network is designed to be a standalone post-processing tool. This means it can be applied to already acquired and reconstructed MRI data without needing any additional information from the scanner or requiring changes to the acquisition sequence.

Datasets

The study utilized two main types of datasets:

  • Simulated Motion Datasets: These were generated using 3D MPRAGE datasets acquired from 44 subjects (for training) and an additional 11 subjects (for testing). Motion was simulated by applying random 3D rigid body motion (±5 mm translation, ±5° rotation) to the k-space data. Five different random motion scenarios were simulated for each subject, resulting in a substantial number of training and testing samples.
  • Motion Corrupted Datasets (Experimental): Real-world motion-corrupted data was acquired from four subjects. Each subject underwent two scans, and they were instructed to perform small random motions two times during each scan. These datasets were used for experimental validation of the trained network.

Key Insights and Ideas

  • Motion as a Convolutional Process: The paper frames motion artifacts mathematically as a linear combination of transformed images and convolution kernels, providing a theoretical basis for using convolutional neural networks.
  • Data-Driven Kernel Learning: Instead of hand-engineering convolution kernels to remove artifacts, the deep learning approach allows the network to learn optimal kernels directly from the data, adapting to the complex nature of motion-induced artifacts.
  • Simulation is Key: The development of a robust motion simulation framework was crucial for training a deep learning model that could generalize to real-world motion.
  • U-Net Architecture Suitability: The encoder-decoder U-Net architecture proved effective for this task, enabling the network to both extract relevant features and reconstruct a clean image.
  • Post-Processing Power: The ability to perform motion correction as a post-processing step is a significant advantage, making the method easily integrable into existing clinical workflows.

Key Results

  • Simulated Data Performance: The trained network successfully suppressed motion artifacts in simulated data, achieving a mean percentage error of 2.69% with a standard deviation of 0.95% in motion-corrected images.
  • Experimental Data Performance: The network also effectively suppressed motion artifacts in real-world experimental datasets, demonstrating its generalization capability. Visual inspection showed substantially lower artifacts compared to standard reconstruction.
  • Quantitative Improvement: For experimental data, the Naturalness Image Quality Evaluator (NIQE) score consistently improved after motion correction, with an average percentage improvement of 49%.
  • Robustness: The network, trained on simulated data, was able to correct for real motion, indicating its robustness.

Limitations and Future Work

  • Sudden vs. Gradual Motion: The current implementation primarily corrects for sudden motion. The network has not been trained to correct for gradual or continuous motion that might occur during an MR acquisition.
  • Simulation Simplifications: The motion simulation did not account for spin history effects or the Bloch equations, which could influence the realism of the simulated artifacts.
  • Future Work: The authors suggest future work to improve robustness by integrating continuous motion correction, generating training images based on Bloch equations, and validating the method on other image contrasts (e.g., T2/T2*, proton density) and for non-rigid motion.

Summary updated

Abstract

The suppression of motion artefacts from MR images is a challenging task. The purpose of this paper was to develop a standalone novel technique to suppress motion artefacts in MR images using a data‐driven deep learning approach. A simulation framework was developed to generate motion‐corrupted images from motion‐free images using randomly generated motion profiles. An Inception‐ResNet deep learning network architecture was used as the encoder and was augmented with a stack of convolution and upsampling layers to form an encoder‐decoder network. The network was trained on simulated motion‐corrupted images to identify and suppress those artefacts attributable to motion. The network was validated on unseen simulated datasets and real‐world experimental motion‐corrupted in vivo brain datasets. The trained network was able to suppress the motion artefacts in the reconstructed images, and the mean structural similarity (SSIM) increased from 0.9058 to 0.9338. The network was also able to suppress the motion artefacts from the real‐world experimental dataset, and the mean SSIM increased from 0.8671 to 0.9145. The motion correction of the experimental datasets demonstrated the effectiveness of the motion simulation generation process. The proposed method successfully removed motion artefacts and outperformed an iterative entropy minimization method in terms of the SSIM index and normalized root mean squared error, which were 5–10% better for the proposed method. In conclusion, a novel, data‐driven motion correction technique has been developed that can suppress motion artefacts from motion‐corrupted MR images. The proposed technique is a standalone, post‐processing method that does not interfere with data acquisition or reconstruction parameters, thus making it suitable for routine clinical practice.
BibTeX
@inproceedings{pawar2019suppressing,
  author = {Kamlesh Pawar and Zhaolin Chen and N. Jon Shah and Gary F. Egan},
  title = {{Suppressing motion artefacts in MRI using an Inception‐ResNet network with motion simulation augmentation}},
  booktitle = {NMR},
  year = {2019},
  doi = {10.1002/nbm.4225},
}
Figure