PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Learning with Domain Adaptation for Accelerated Projection-Reconstruction MR

Yo Seob Han, Jaejun Yoo, Jong Chul Ye

arXiv:1703.01135v2cs.CV

TL;DR

Under-sampled radial k-space data create streaking artifacts, while compressed-sensing reconstruction is computationally complex. The paper uses domain adaptation from CT or synthetic radial MR data, fine-tuned with few MR datasets, and reports better image quality and faster computation than existing compressed-sensing methods.

  • Problem

    Radial MR acquisition faces a trade-off between the many lines needed for high-resolution reconstruction and the streaking artifacts produced by insufficient sampling, while compressed-sensing reconstruction has high computational complexity.

  • Method

    A deep network removes streaking artifacts and is adapted from pre-training on CT or synthetic radial MR data using only a few radial MR datasets for fine-tuning.

  • Results

    The proposed method outperforms total variation and PR-FOCUSS in image quality and computation time.

  • Takeaways & Limitations

    Pre-training with similar organ structures is more important than pre-training with the same modality but different organs, supporting domain adaptation when MR data are limited.

  • Takeaways & Limitations

    Compressed-sensing methods are limited by increased computational complexity caused by iterative reconstruction.

Abstract

from arXiv · show

Purpose: The radial k-space trajectory is a well-established sampling trajectory used in conjunction with magnetic resonance imaging. However, the radial k-space trajectory requires a large number of radial lines for high-resolution reconstruction. Increasing the number of radial lines causes longer acquisition time, making it more difficult for routine clinical use. On the other hand, if we reduce the number of radial lines, streaking artifact patterns are unavoidable. To solve this problem, we propose a novel deep learning approach with domain adaptation to restore high-resolution MR images from under-sampled k-space data. Methods: The proposed deep network removes the streaking artifacts from the artifact corrupted images. To address the situation given the limited available data, we propose a domain adaptation scheme that employs a pre-trained network using a large number of x-ray computed tomography (CT) or synthesized radial MR datasets, which is then fine-tuned with only a few radial MR datasets. Results: The proposed method outperforms existing compressed sensing algorithms, such as the total variation and PR-FOCUSS methods. In addition, the calculation time is several orders of magnitude faster than the total variation and PR-FOCUSS methods.Moreover, we found that pre-training using CT or MR data from similar organ data is more important than pre-training using data from the same modality for different organ. Conclusion: We demonstrate the possibility of a domain-adaptation when only a limited amount of MR data is available. The proposed method surpasses the existing compressed sensing algorithms in terms of the image quality and computation time.

2 Department of Radiology, Research Institute of Radiology,

The paper is associated with Magnetic Resonance in Medicine and KAIST’s Department of Bio and Brain Engineering.

  • The paper is published in Magnetic Resonance in Medicine.
  • Jong Chul Ye, Ph.D., is listed as an author.
  • The listed affiliation is the Department of Bio and Brain Engineering at KAIST.
  • The paper’s keywords include deep learning, convolutional neural networks, domain adaptation, projection reconstruction MRI, and compressed sensing.

Introduction

The introduction motivates accelerated radial MR reconstruction by highlighting streaking artifacts, compressed-sensing complexity, and limited radial-MR training data. It proposes domain adaptation from CT or synthetic MR data to make deep-learning restoration practical with few MR datasets.

  • Compressed-sensing reconstruction reduces scan time but incurs increased computational complexity from iterative reconstruction.
  • Radial MR data resemble CT sinograms because the projection slice theorem converts radial k-space data into sinogram data.
  • The paper proposes domain adaptation by transferring a deep network trained on CT data to MR reconstruction.
  • Only a few radial MR datasets are needed for fine-tuning, reducing training time and expanding deep learning’s applicability to MR imaging.
  • Synthetic radial MR data from public MR images can support domain adaptation when underlying organ structures are similar.
  • The trained network is more accurately characterized as an artifact-removal restoration method than as a reconstruction method.

Theory

The paper models radial MR reconstruction through its Fourier-domain relationship to projection data and addresses sparse-view streaking with domain adaptation. The framework minimizes target-domain risk by combining source-domain training with discrepancy reduction using limited MR data.

  • Radial reconstruction: Radial k-space data at angle θ are represented by the one-dimensional Fourier transform of projection data, with inverse reconstruction using a ramp filter.The formulation is associated with filtered back-projection for parallel-beam geometry.
  • Sparse-view artifacts: Insufficient radial scan lines Nθ produce streaking artifacts in reconstructed MR and CT images.The paper notes that similar artifact patterns across modalities motivate transferring CT-trained networks to MR restoration.
  • Domain adaptation: Domain adaptation uses CT or synthesized radial MR data as a source domain and in vivo radial MR data as the target domain.The practical setting assumes abundant source-domain projection data but relatively few labeled target-domain MR datasets.
  • Domain adaptation: The target-domain objective is bounded by source empirical risk, inter-domain discrepancy, and a network-complexity penalty.The method therefore minimizes source risk and domain discrepancy through pre-training and fine-tuning.
  • Network design: The proposed network uses a multi-scale U-net with residual paths and a large receptive field to capture globally distributed streaking artifacts.Pooling, unpooling, and skip connections combine information across scales.

Materials and Methods

The study pre-trains a streak-removal network on CT or synthetic radial MR data, then fine-tunes it with a small number of in vivo radial MR slices. Brain and abdominal datasets support retrospective and in vivo evaluations across projection counts.

  • Data preparation: The domain-adaptation pipeline combines source-domain pre-training with target-domain fine-tuning using only a few in vivo radial MR datasets.Fine-tuning was performed separately for brain and abdomen data using CT- and HCP-derived pre-trained networks.
  • Preprocessing: Reconstructed 512×512 real-valued images were converted to magnitude images, while parallel-imaging data were combined using square root of sum of squares.Training used artifact-contaminated images generated across multiple downsampling factors.
  • Pre-training datasets: Synthetic radial MR training data were generated from 100 HCP subjects and 3,600 slices using Radon projections and 36, 45, 60, and 90 views.Artifact-corrupted inputs were regenerated with the inverse Radon operator.

Results

The proposed domain-adapted networks removed streaking artifacts while preserving detailed MR structure and outperformed TV and PR-FOCUSS. They also generalized to in vivo data and substantially reduced reconstruction time.

  • Domain adaptation: HCP pre-training performed better than CT pre-training for brain reconstruction, preserving detailed MR structures that CT pre-training sometimes removed.Fine-tuning with one MR dataset improved the CT-pre-trained result, while additional fine-tuning produced more realistic MR images.
  • Quantitative comparison: The two proposed networks exceeded TV and PR-FOCUSS in average NMSE across retrospective evaluations using 90, 60, 45, and 36 projection views.Average NMSE was computed over twenty restored slices.
  • Image quality: The proposed method eliminated streaking artifacts while preserving detailed brain structure across projection views in qualitative comparisons.This included retrospective 45-view data and in vivo accelerations.
  • In vivo validation: The proposed network generalized to in vivo accelerations with significantly improved image quality, including 90-view acquisitions.In vivo NMSE was calculated against separately acquired 180-view data.
  • Diagnostic evaluation: Radiologists found 15-slice fine-tuning comparable to fully sampled reconstruction, while 1-slice fine-tuning still provided very good diagnostic quality.TV and PR-FOCUSS produced artifacts that affected diagnosis.
  • Computation time: 0.05 seconds was required for restoration, versus approximately 24 to 38 seconds for TV and nearly 29 to 60 seconds for PR-FOCUSS.The reported timings show a large computational advantage for the proposed method.
  • Organ and modality effects: CT pre-training was quantitatively better than HCP pre-training for abdominal data, suggesting that similar-organ data mattered more than matching imaging modality.This conclusion contrasted with the brain comparison.

Discussions

Discussion analyses show that domain adaptation improves stability with limited MR data, while performance increases with more fine-tuning slices. The method remains scoped to static magnitude reconstruction and has stated acceleration boundaries.

  • Domain adaptation versus MR-only learning: Domain adaptation outperformed an MR-only residual network because few MR datasets were insufficient to learn strong global artifacts from severe undersampling.The advantage was observed in retrospective and in vivo comparisons.
  • Practical implications: A pre-trained network can be fine-tuned with little MR data, avoiding additional source-domain pre-training when an available network exists.The authors identify this as an advantage over MR approaches requiring considerably larger datasets.
  • Abdominal reconstruction: Abdominal reconstruction was stable despite more complex structures than brain images, with 15 MR datasets and 3,000 epochs identified as reasonable.The reported abdominal setting used more data and epochs than the brain setting.
  • Fine-tuning data dependence: More MR fine-tuning data improved restored-image quality and reduced average NMSE, while domain adaptation remained more stable than MR-only training.The MR-only network surpassed CS methods after more than three MR slices.
  • Acceleration scope: The current method requires at least 36 projection views for brain data and 75 for abdominal data.The paper contrasts this static-MR setting with dynamic compressed sensing methods that exploit temporal redundancy.
  • Limitations and future work: Dynamic MR extension and higher acceleration using temporal redundancy remain outside the scope of the present paper.The authors state that this topic will be reported elsewhere.
  • Generalization: The network can remove streaking artifacts even when training and test trajectories differ, including fan-beam CT training and parallel-beam MR testing.This generalization was observed even without domain adaptation in the cited comparison.

Conclusion

The proposed domain-adaptation network reconstructs high-quality radial MR images from undersampled k-space using limited MR data. It improves restoration and computational efficiency, while similar organ structure is more important than matching imaging modality for pre-training.

  • The proposed deep learning approach reconstructs high-quality MR images from sub-sampled k-space data using domain adaptation.It pre-trains on CT or synthetic radial MR datasets and fine-tunes with a small number of radial MR datasets.
  • With insufficient MR training data, combining pre-training and fine-tuning produced the best restoration performance.When sufficient radial MR data were available, an MR-only deep network also showed good restoration results.
  • Similar organ structure was more important than imaging modality when selecting pre-training data.The experiments therefore favored pre-trained networks from similar organ structures.
  • The method was much faster than conventional compressed sensing because it avoids computationally heavy projection or back-projection operations.The cited comparison concerns conventional compressed sensing methods and their projection-related computation.
  • Domain adaptation can mix medical imaging systems when their artifacts are similar and topologically simple.The paper connects this possibility to relationships among CT, MRI, and optical diffraction tomography in Fourier space.
Loading 1703.01135v2…