[Paper Review] Deep artifact learning for compressed sensing and parallel MRI

Dongwook Lee, Jae Jun Yoo+1 more|arXiv (Cornell University)|Mar 3, 2017

Advanced MRI Techniques and Applications22 references19 citations

TL;DR

This paper proposes a deep learning-based method for fast and accurate MRI reconstruction from highly undersampled k-space data by directly learning and removing coherent aliasing artifacts. Using a multi-scale U-Net with a large receptive field, the network estimates artifacts from magnitude and phase images reconstructed from subsampled data, achieving state-of-the-art image quality and reconstruction speed—up to an order of magnitude faster than traditional compressed sensing methods.

ABSTRACT

Purpose: Compressed sensing MRI (CS-MRI) from single and parallel coils is one of the powerful ways to reduce the scan time of MR imaging with performance guarantee. However, the computational costs are usually expensive. This paper aims to propose a computationally fast and accurate deep learning algorithm for the reconstruction of MR images from highly down-sampled k-space data. Theory: Based on the topological analysis, we show that the data manifold of the aliasing artifact is easier to learn from a uniform subsampling pattern with additional low-frequency k-space data. Thus, we develop deep aliasing artifact learning networks for the magnitude and phase images to estimate and remove the aliasing artifacts from highly accelerated MR acquisition. Methods: The aliasing artifacts are directly estimated from the distorted magnitude and phase images reconstructed from subsampled k-space data so that we can get an aliasing-free images by subtracting the estimated aliasing artifact from corrupted inputs. Moreover, to deal with the globally distributed aliasing artifact, we develop a multi-scale deep neural network with a large receptive field. Results: The experimental results confirm that the proposed deep artifact learning network effectively estimates and removes the aliasing artifacts. Compared to existing CS methods from single and multi-coli data, the proposed network shows minimal errors by removing the coherent aliasing artifacts. Furthermore, the computational time is by order of magnitude faster. Conclusion: As the proposed deep artifact learning network immediately generates accurate reconstruction, it has great potential for clinical applications.

Motivation & Objective

• Address the high computational cost of traditional compressed sensing MRI (CS-MRI) and parallel imaging (pMRI) reconstruction methods.
• Overcome the limitations of existing CS-MRI approaches in removing coherent aliasing artifacts, especially under uniform subsampling patterns.
• Develop a deep learning framework that learns the artifact patterns directly rather than the artifact-free image, leveraging topological simplicity of artifact manifolds.
• Enable real-time clinical MRI reconstruction by drastically reducing reconstruction time compared to iterative CS methods.
• Improve performance on both single-coil and multi-coil MRI data, particularly in phase image reconstruction where artifacts are globally distributed.

Proposed method

• Propose a deep artifact learning framework that estimates aliasing artifacts directly from magnitude and phase images reconstructed from undersampled k-space data.
• Use a multi-scale U-Net architecture with dilated convolutions and pooling layers to achieve a large, fully receptive field covering the entire 256×256 input image.
• Train the network end-to-end to predict the aliasing artifact from the corrupted input, then subtract it to recover the artifact-free image.
• Incorporate additional low-frequency ACS (autocalibration signals) lines to simplify the data manifold of aliasing artifacts, improving learnability.
• Apply the network separately to magnitude and phase components to handle their distinct artifact distributions—local for magnitude, global for phase.
• Leverage topological data analysis (persistent homology) to justify that uniform subsampling with ACS lines produces simpler, more regular artifact manifolds, facilitating deep learning.

Experimental results

Research questions

• RQ1 Can deep learning effectively estimate and remove coherent aliasing artifacts in highly accelerated MRI, even when traditional CS methods fail?
• RQ2 Does the topological simplicity of aliasing artifacts under uniform subsampling with ACS lines improve the learnability of deep neural networks?
• RQ3 How does a multi-scale network architecture with a large receptive field compare to single-scale networks in reconstructing MRI with globally distributed aliasing artifacts?
• RQ4 Can the proposed method achieve both high reconstruction accuracy and significantly faster computation than conventional iterative CS-MRI methods?
• RQ5 Is the deep artifact learning approach effective for both single-coil and multi-coil MRI data, especially in phase reconstruction where artifacts are spatially coherent?

Key findings

• The proposed deep artifact learning network achieved minimal reconstruction errors by effectively removing coherent aliasing artifacts, outperforming existing CS and parallel imaging methods.
• The multi-scale U-Net with a large receptive field fully covering the input image showed superior performance, especially in phase reconstruction where artifacts are globally distributed.
• The single-scale network failed to capture global artifact patterns, resulting in higher errors—particularly in phase images—while the multi-scale network achieved significantly lower error rates.
• Reconstruction time was reduced by an order of magnitude compared to traditional iterative CS-MRI methods, enabling near real-time clinical application.
• The inclusion of ACS lines in the sampling pattern simplified the data manifold of aliasing artifacts, making them easier to learn and improving overall network performance.
• The method demonstrated strong generalization across both single-coil and multi-coil MRI data, maintaining high accuracy even under strong acceleration factors.