PaperlayerView exact revision on arXiv

Source-linked AI summary

PGDM-MRSRGAN: Physics-Guided Degradation Model with an SRGAN Framework for Magnetic Resonance Image Super-Resolution: Applications in Low-Field MRI

Yashwant Kurmi, Charlotte R. Sappo, Sai Abitha Srinivas, Selin D. Kandemir, Malvika Viswanathan, Zhongliang Zu

arXiv:2609.30431v1eess.IVphysics.med-ph

TL;DR

Low-field MRI has reduced SNR and spatial resolution, motivating an approach tailored to its degradation. The paper integrates a physics-guided degradation model with an MRI super-resolution GAN framework; it reports improvements across image-quality metrics and validation on simulated and in vivo datasets.

  • Problem

    Low-field MRI has reduced SNR and spatial resolution, limiting image quality for diagnosis.

    Source: §1 · p. 3

  • Method

    The paper integrates a physics-guided degradation model with a deep-learning framework and uses MSAM loss to preserve spectral relationships and reconstruct texture and intensity variations.

    Source: §1 · pp. 3–4, §4 · p. 16

  • Results

    The framework achieves significant improvements across SNR, PSNR, SSIM, NQM, MSSSIM, MSAM, and NIQE, with reliability validated on simulated and in vivo datasets.

    Source: §5 · pp. 16–17

  • Takeaways & Limitations

    The conclusion presents the framework as advancing low-field MRI image resolution for faster, accurate, and clinically impactful imaging.

    Source: §5 · p. 17

    Broader applicability depends on addressing challenges such as dataset-related constraints.

    Source: §4 · p. 16

Abstract

from arXiv · show

Magnetic Resonance Imaging (MRI) often suffers from low signal-to-noise ratio (SNR) and limited spatial resolution, which compromise clinical precision. This study aims to address these challenges by developing a physics-guided degradation model (PGDM) and a novel deep learning framework, Magnetic Resonance Super-Resolution GAN Imaging (MRSRGAN), for improved super-resolution reconstruction of MRI images. The proposed approach consists of two stages: (1) a physics-guided degradation model that simulates low-field MRI conditions by incorporating artifacts such as blur, B0 inhomogeneity, chemical shift effects, down-sampling, and noise to generate paired low-resolution (LR) and high-resolution (HR) training datasets; and (2) an MRSRGAN-based reconstruction network utilizing a hybrid loss function, including modified spectral angle mapper (MSAM) and sharpened ground truth images based on de-convolution of slice profile-based kernels to restore high-SNR and high-resolution images. Experimental validation demonstrates that the proposed method achieves superior spatial resolution, enhanced SNR, improved peak signal-to-noise ratio (PSNR), higher structural similarity index (SSIM), reduced MSAM error, and better non-reference image quality evaluation (NIQE) metrics. Furthermore, it effectively reduces artifacts and demonstrates robustness on real MRI datasets. The MRSRGAN framework, guided by the physics-informed degradation model, provides a significant improvement in MRI image quality, enhancing spatial resolution and diagnostic accuracy. Its demonstrated robustness and effectiveness on real MRI datasets highlight its potential to resolution enhancement for precise diagnostic imaging.

1. INTRODUCTION

Low-field MRI faces reduced SNR, lower spatial resolution, and longer scan times, while computational enhancement methods can lose fine structural detail. The paper proposes a physics-guided degradation model and MRSRGAN to simulate low-field artifacts and reconstruct higher-SNR, higher-resolution images.

1. INTRODUCTION

  • Low-field MRI has reduced SNR, lower spatial resolution, and longer scan times, affecting diagnostic quality.These constraints are especially relevant in resource-limited settings and for patients with metallic implants incompatible with high-field systems.Source: §1 · p. 3
  • The paper positions computational post-processing as a scalable, cost-effective alternative to hardware upgrades that can be impractical or cost-prohibitive.The introduction also connects detailed anatomical information to accurate diagnosis and image processing.Source: §1 · p. 3
  • Interpolation-based enhancement methods often fail to preserve fine structural details needed in complex clinical cases.Traditional GAN training may also rely on optical degradation models that fail to capture MRI-specific properties, contributing to a performance gap on real MRI data.Source: §1 · p. 3
  • The proposed PGDM simulates low-field MRI artifacts, and MRSRGAN reconstructs high-SNR, high-resolution images.The degradation model addresses blur, noise, distortions, and chemical shift effects; the reconstruction network uses dense blocks and sensitive activation functions for artifact suppression and structural-detail preservation.Source: §1 · p. 3
  • The framework uses MSAM loss based on angle images from Fourier-transformed real and imaginary components to mimic MRI formation.The paper also sharpens ground-truth images by de-convolving slice profile-based kernels, with the stated aim of improving generalization across MRI modalities and conditions.Source: §1 · p. 4

2. METHOD

The study introduces two components for improved MRI reconstruction: a physics-guided degradation model and MRSRGAN, organized as a generator-discriminator framework.

2. METHOD

  • The study introduces a PGDM and an MRSRGAN generator-discriminator as its two components for improved MRI reconstruction.The passage associates PGDM with Fig. 1A and MRSRGAN with Figs. 1B and 1C.Source: §2 · p. 4

2.1.MRI Physics-Guided Degradation Model

The physics-guided degradation pipeline creates low-resolution MRI inputs by applying simulated artifacts and image-quality degradations to high-resolution images.

2.1.MRI Physics-Guided Degradation Model

  • Three randomly selected artifact types from 11 are sequentially applied to each high-resolution image to create clinically relevant degraded inputs.The authors limit each image to three artifacts to keep reconstruction tractable.Source: 2.1.MRI Physics-Guided Degradation Model · p. 4

2.1.1. Artifacts Degradation

  • Artifact degradation includes MRI effects such as slice-thickness and frequency-encoding blur, motion, chemical shift, ringing, field inhomogeneity, and shading.The section also describes magnetic susceptibility and edge-related artifacts.Source: §2.1.1 · pp. 5–6

2.1.2. Down Sampling (D1)

2.1.3. Noise Degradation (N1)

  • Noise degradation combines Rician and Gaussian noise, with Gaussian noise modeled using a standard deviation and a zero-mean, unit-variance distribution.The noisy image is formed by adding Rician and Gaussian noise to the original image.Source: §2.1.3 · p. 8

2.1.4. JPEG compression noise (N2)

  • JPEG degradation uses 8×8 discrete cosine transform blocks and quantization, simulating compression artifacts including blocking and detail loss.The image is reconstructed by inverse quantization and inverse DCT.Source: §2.1.4 · p. 8

2.2.Network Architectures and Training

MRSRGAN adapts an ESRGAN-based architecture for MRI reconstruction, pairing a 19-RRDB generator with a sensitive discriminator and MRI-specific training targets.

Generator

MRI-specific reconstruction

Discriminator

  • The discriminator uses spectral normalization and LeakyReLU with negative slope 0.25, forming the authors’ ESDiscriminator.The authors describe it as designed for enhanced sensitivity.Source: 2.2.Network Architectures and Training · p. 8

2.3.Loss Functions

The reconstruction objective combines content, perceptual, adversarial, and modified spectral-angle losses to address image similarity and phase-angle structure.

2.3.1. Content Loss (𝑙𝑀𝑆𝐸

  • The content loss compares the high-resolution ground truth with the generator’s reconstructed image.Source: §2.3.1 · p. 9

2.3.2. Perceptual Loss (𝑙𝑉𝐺𝐺

  • The perceptual loss emphasizes feature-level similarity using feature maps from a pre-trained VGG19 network.The feature-map width and height are defined for the jth VGG19 layer.Source: §2.3.2 · p. 9

2.3.3. Adversarial Loss (𝑙𝐴𝑑𝑣

2.3.4. Modified Spectral Angle Mapper (MSAM) Loss (𝑙𝑀𝑆𝐴𝑀

  • MSAM loss compares phase-angle images using Fourier transformation, frequency shifting, and normalization while preserving k-space features.The hybrid loss combines multiple components to target structural accuracy, visual realism, and phase-angle images.Source: §2.3.4 · p. 9

2.4.Training strategy

Training proceeds from content-loss pretraining to hybrid-loss MRSRGAN pretraining and artifact-specific fine-tuning, using simulated MRI degradations.

Three-stage training

  • Training has three stages: generator training on degraded MRI, MRSRGAN pretraining with a hybrid loss, and fine-tuning on artifact-specific degraded data.Source: 2.4.Training strategy · pp. 9–10

MRSRGAN pretraining

  • The hybrid pretraining loss combines content, perceptual, GAN, and MSAM terms, while slice-profile deconvolution sharpens high-frequency ground-truth components.Source: 2.4.Training strategy · pp. 9–10

Artifact-specific fine-tuning

  • Fine-tuning on data degraded by specific artifact types is used to optimize the pretrained model for accurate prediction.Source: 2.4.Training strategy · p. 10

Degradation augmentation

  • Degradation training randomly shuffles artifact simulation, downsampling, noise addition, and JPEG compression, using randomly selected bicubic scale factors.The authors state this strategy is intended to enhance training robustness.Source: 2.4.Training strategy · p. 10

2.5. Performance parameters

The section lists the metrics used to evaluate reconstructed images and specifies which metric directions indicate better performance.

2.5. Performance parameters

  • Performance is assessed with SNR, PSNR, SSIM, MSSSIM, NQM, MSAM, and NIQE.NIQE evaluates reconstructed-image naturalness.Source: §2.5 · p. 10
  • Higher SNR, PSNR, SSIM, NQM, and MSSSIM values indicate better performance.Source: §2.5 · p. 10
  • Lower MSAM and NIQE values signify improved reconstruction.Source: §2.5 · p. 10

3. Results

The results section evaluates MRSRGAN on simulated and real MRI data, reporting reconstruction quality and artifact reduction alongside an image-sharpening training technique and intermediate phase-image results. It also reports SNR and NIQE results on a 47.5 mT low-field MRI dataset, with structure and texture preservation and artifact correction under varying B0 inhomogeneity and noise.

3.1. Datasets and Implementation

  • The MRSRGAN framework used M4Raw data (47) for synthetic data and included in-vivo imaging experiments in ten healthy volunteers, with degradation using Gaussian, generalized Gaussian, and plateau-shaped kernels and Gaussian/Rician noise.Gaussian and Rician noise were each used with a probability of 0.5.Source: §3.1 · pp. 10–11

3.2.Training Data pairs

  • Sharpened reference images were introduced for fine-tuning using slice-profile deconvolution with profiles in the range [2-5].The stated aim was to enhance visual sharpness without noticeable artifacts.Source: 3.2.Training Data pairs · pp. 11–12

3.3.Intermediate Results on the Proposed MSAM Angle

  • MRSRGAN’s in-phase and Q-phase outputs showed less deviation from the reference than the other compared super-resolution methods.The comparison included SRGAN, BSRGAN, ESRGAN, and Real-ESRGAN.Source: 3.3.Intermediate Results on the… · p. 12

3.4.Comparison of Qualitative Results

  • MRSRGAN achieved the best reported performance, with PSNR 30.54 dB, SSIM 0.7332, NQM 26.12, MSSSIM 0.9476, and MSAM 1.5235.Compared with other methods, the proposed approach yielded the lowest residual artifacts.Source: 3.4.Comparison of Qualitative Results · p. 13
  • 28.08 dB PSNR, 0.7459 SSIM, and 25.79 NQM were reported for MRSRGAN on MR-ART, alongside improved reconstructions across motion severity levels and imaging views.The prose says MRSRGAN surpasses SRGAN, ESRGAN, BSRGAN, and Real-ESRGAN, and reports similar improvement trends to the simulated dataset.Source: 3.4.Comparison of Qualitative Results · pp. 13–14
  • In SLP simulated and MR-ART synthetic datasets (where MR-ART artifacts were simulated in ground-truth images), MRSRGAN reduced intensity distortions and improved reconstruction quality and artifact correction as B0-inhomogeneity severity rose from level 1 to 2, consistently outperforming other methods.Source: 3.4.Comparison of Qualitative Results · p. 14
  • As noise increased from level 1 to 2 in SLP simulated and MR-ART synthetic data, MRSRGAN had the least residuals and consistently delivered top performance across all images versus state-of-the-art methods.Source: 3.4.Comparison of Qualitative Results · p. 14

3.5. Experiments on our LF MRI Dataset

  • Compared with MRSRGAN, SRGAN amplified artifacts, BSRGAN smoothed away structural details, and Real-ESRGAN introduced graininess and hallucinated textures.These shortcomings were reported to compromise anatomical accuracy, texture, and diagnostic utility.Source: §3.5 · p. 15
  • On the low-field MRI images, MRSRGAN reduced noise while preserving structural details and texture, yielding higher-resolution, anatomically accurate, diagnostically useful reconstructions.Both visual and tabular analyses are reported as validating these results.Source: §3.5 · p. 15
  • MRSRGAN achieved an SNR of 120 dB for normalized images and the best NIQE value of 1.1.Figs. 10(A, B) present SNR and NIQE values for the methods.Source: §3.5 · p. 15

4. Discussion

The discussion presents MRSRGAN as an MRI-focused approach to low-field super-resolution, reporting improved reconstruction accuracy and image quality. It also identifies dataset diversity and generalizability as remaining challenges.

4. Discussion

  • MRSRGAN is described as optimizing the trade-offs among resolution, SNR, and acquisition time, a longstanding MRI challenge.Source: §4 · p. 15
  • MRSRGAN’s MSAM loss targets the spectral-fidelity gap identified in conventional super-resolution methods.Source: §4 · pp. 15–16
  • MSAM preserves spectral relationships, while slice-profile deconvolution enhances high-frequency components and image fidelity.Source: §4 · p. 16
  • The discussion positions MRSRGAN as a practical diagnostic-imaging tool for resource-constrained settings.Source: §4 · p. 16
  • K-space data improves reconstruction accuracy over image-space methods.Source: §4 · p. 16
  • At 47.5mT, models trained on 0.3T datasets can generate high-quality images under resource-limited conditions.Source: §4 · p. 16
  • Broader applicability remains dependent on addressing dataset diversity and generalizability across populations and imaging scenarios.The discussion proposes domain adaptation and deep-learning-based EMI mitigation as future research directions.Source: §4 · p. 16

5. Conclusion

The conclusion presents MRSRGAN as a physics-guided approach to low-field MRI enhancement, reporting gains across image-quality metrics and validation on simulated and in vivo data. The accompanying material also describes method comparisons and the availability of code and data on request.

5. Conclusion

  • MRSRGAN combines MRI-artifact, multiscale downsampling, and noise degradation with modified SAM loss and slice-profile deconvolution-based sharpening.Source: §5 · p. 16
  • The framework reports improvements in SNR, PSNR, SSIM, NQM, MSSSIM, MSAM, and NIQE, with validation on simulated and in vivo datasets.Source: §5 · pp. 16–17
  • The conclusion characterizes the framework as reliable for advancing low-field MRI toward faster, accurate imaging.Source: §5 · p. 17

DATA AVAILABILITY STATEMENT

Supplementary information

The supplementary information catalogs degradation-artifact parameters and their ranges, and lists supporting comparisons for several artifact-related enhancement settings.

Supplementary information

Related papers

Loading 2609.30431v1…