PaperlayerView exact revision on arXiv

Source-linked AI summary

Highly accelerated 3D Cartesian MPnRAGE with implicit neural representation reconstruction

Natascha Niessen, Ana Beatriz Solana, Carolin M. Pirkl, Tim Sprenger, Hannah Eichhorn, Veronika Spieker, Wenqi Huang, Rolf F. Schulte, Florian Wiesinger, Tobias C. Wood, Marion I. Menzel, Julia A. Schnabel on behalf of the PREDICTOM consortium

arXiv:2609.18589v1eess.IV

TL;DR

MPnRAGE can provide multiple inversion contrasts and quantitative T1 information, but accelerating 3D acquisitions is important for reducing scan time. This work combines Cartesian 3D MPnRAGE with tailored sampling and joint INR reconstruction; the method outperforms state-of-the-art reconstruction methods and produces high-quality images at R = 20 in 5.39 minutes.

  • Problem

    MPnRAGE provides multiple contrasts and quantitative T1 information, while 3D scan times motivate accelerated acquisition.

    Source: §1 · p. 2

  • Method

    The method combines Cartesian 3D MPnRAGE with complementary variable-density undersampling, calibration data, and joint INR reconstruction.

    Source: §1 · p. 2

  • Results

    INR outperforms state-of-the-art reconstruction methods and provides high-quality images at R = 20 in 5.39 min.

    Source: §5 · p. 15

  • Takeaways & Limitations

    The framework produces high-quality multi-contrast images from highly accelerated Cartesian 3D MPnRAGE acquisitions.

    Source: §5 · p. 15

    Slice-wise INR processing may introduce subtle slice-to-slice artifacts, and R = 20 whole-volume reconstruction takes 13.8 minutes, too long for clinical applications.

    Source: §4.4 · p. 14

Abstract

from arXiv · show

MPnRAGE enables multiple inversion contrast images in a single scan, allowing quantitative T1 mapping, tissue nulled contrasts, and standard MPRAGE synthesis. However, current 3D scan times remain clinically impractical, motivating accelerated 3D MPnRAGE. This work provides a highly accelerated Cartesian 3D MPnRAGE sequence with joint implicit neural representation (INR) reconstruction. The sequence uses a tailored view-ordering strategy, flip angle schedule and complementary variable-density Poisson-disk undersampling. Calibration data acquired during otherwise unused delay time are used for sensitivity map estimation and complementary high-frequency sampling. Ten INR-reconstructed inversion images at 1.5 mm$^3$ are evaluated against fully sampled references via retrospective undersampling. Prospectively accelerated 1 mm$^3$ images at R = 20 (5.39 min) demonstrate clinical feasibility. INR reconstruction outperforms subspace and iterative local low rank reconstruction on highly accelerated data. The proposed highly undersampled 3D Cartesian MPnRAGE with INR reconstruction generates multiple high-quality inversion contrasts in substantially reduced scan time. Scan-specific INR reconstruction improves image quality while reducing reconstruction time versus state-of-the-art methods.

1 INTRODUCTION

MPnRAGE captures multiple inversion contrasts for T1 mapping and tissue-nulled images, while scan-specific INR offers a route to reconstructing these contrasts jointly. This work applies complementary undersampling and tailored acquisition strategies to Cartesian 3D MPnRAGE.

1 INTRODUCTION

  • MPnRAGE samples signal evolution throughout inversion recovery, enabling quantitative T1 mapping and multiple contrasts from one dataset.These include standard MPRAGE-like, gray-matter-nulled, and white-matter-nulled images.Source: §1 · p. 2
  • The Cartesian 3D sequence combines tailored view ordering with complementary variable-density Poisson-disk undersampling for accelerated acquisition.The approach is designed to maintain TI continuity in k-space while limiting eddy-current artifacts.Source: §1 · p. 2
  • Calibration data acquired during otherwise unused delay time support sensitivity-map estimation and add denser sampling to the joint INR reconstruction.The data are acquired with low flip angles and higher sampling density.Source: §1 · p. 2
  • A custom flip-angle schedule is introduced to improve SNR near tissue null points.The schedule is designed to increase signal for low-SNR inversion times.Source: §1 · p. 2
  • The study evaluates scan-specific INR reconstruction against established approaches for MPnRAGE.INR methods are self-supervised and scan-specific, making them attractive for sequences without large training datasets.Source: §1 · p. 2

2.1 Accelerated Cartesian 3D MPnRAGE sequence

The sequence combines complementary undersampling and continuity-aware view ordering with calibration data acquired during recovery delays. Dictionary-based signal modeling guides T1 estimation and a flip-angle schedule intended to improve low-SNR inversion images.

2.1.1 Complementary undersampling with tailored view-ordering

  • Complementary variable-density Poisson-disk masks distribute high-frequency sampling across TI contrasts while densely sampling each contrast's k-space center.TI bins are acquired interleaved after each inversion; view ordering alternates ky traversal direction and places the k-space center midway through each bin.Source: §2.1.1 · p. 3, Figure 1 · p. 4

2.1.2 Integrated calibration TI for sensitivity maps

  • The clinically used recovery delay supplies two low-flip-angle calibration intervals with denser sampling and a larger fully sampled k-space center.The center can be used for sensitivity-map estimation; two intervals limit signal variation within an interval.Source: §2.1.2 · p. 3

2.1.3 Iterative dictionary for T1 mapping and FA schedule optimization

  • A steady-state dictionary models longitudinal magnetization across inversion cycles and uses signal at each TI center for T1 estimation.The custom flip-angle schedule increases signal for low-SNR TI 2–5 while reducing flip angles at higher-signal TI 1 and TI 6–10.Source: §2.1.3 · pp. 3–4

2.2 Image reconstruction

The reconstruction models coil-specific undersampled k-space measurements and jointly represents multi-TI images with a scan-specific implicit neural representation. Data consistency and spatial total-variation regularization guide reconstruction.

2.2.1 Forward model

  • The forward model represents each coil's measurements as sensitivity-weighted Fourier-encoded images sampled by TI-specific masks, with additive noise.Readout Fourier transformation can be performed first, splitting reconstruction into smaller independent image-slice problems.Source: §2.2.1 · pp. 4–5

2.2.2 Joint INR reconstruction

  • A single INR maps spatial coordinates to complex signals across contrasts, exploiting shared anatomy in complementary undersampled TI images.The method applies coil sensitivities, FFT, and sampling masks to learned images for k-space data consistency; spatial TV regularization is applied to the INR output.Source: §2.2.2 · p. 5, Figure 3 · p. 6
  • The scan-specific INR combines hashgrid encoding with a lightweight multilayer perceptron and is trained from scratch for each subject and slice.The architecture uses two MLP layers with 64 neurons each.Source: §2.2.2 · p. 5

2.3 Data acquisition

Phantom and in-vivo acquisitions were performed on 3T scanners using multiple head-coil configurations. The in-vivo protocol includes fully sampled 1.5 mm isotropic references and prospectively accelerated acquisitions at 1.5 mm and 1 mm isotropic resolution.

2.3 Data acquisition

  • Phantom and in-vivo scans used 3T GE scanners with 8-, 24-, or 48-channel head coils; 24- and 48-channel data were compressed to 12 virtual coils.The acquisition series comprised 10 TI contrast images followed by two calibration images.Source: §2.3 · p. 6

2.3.1 Phantom data

  • NIST/ISMRM phantom acquisitions were used to test the protocols, with spin-echo inversion-recovery data collected to assess quantitative T1 values against documentation.The additional spin-echo acquisition took 13.36 hours.Source: §2.3.1 · p. 6

2.3.2 In vivo data

  • Two healthy volunteers provided fully sampled 1.5 mm isotropic references, while prospective scans included 1.5 mm and 1 mm isotropic resolutions.At 1 mm isotropic resolution, five healthy volunteers were scanned at R = 14, 17, and 20, with the R = 20 scan taking 5.39 minutes.Source: §2.3.2 · pp. 6–7

2.4 Comparison methods

The section compares individual-contrast and joint-contrast reconstruction methods, then describes the evaluation and T1-mapping procedures used to assess them.

2.4 Comparison methods

  • INR is compared with zero-filled iFFT and PICS, which reconstruct each TI separately, and with subspace and iterative LLR, which jointly use complementary information.All methods apply a readout-direction 1D FFT and reconstruct image slices individually.Source: §2.4 · p. 7
  • Subspace reconstruction represents the TI series in a dictionary-derived low-dimensional temporal basis, estimates spatial coefficient images with L2-regularized least squares, and evaluates different basis counts.The coefficient problem is solved by conjugate-gradient descent; λ = 10^-2 was selected by parameter sweep.Source: §2.4 · p. 7
  • Iterative LLR minimizes the nuclear norm in 32×32 voxel neighborhoods and adds global singular-value-thresholding denoising.The approach follows Allen et al. [14].Source: §2.4 · p. 7

2.4.1 Image processing and evaluation metrics

  • Retrospective undersampling is evaluated against fully sampled iFFT references using PSNR and SSIM, with masking and percentile normalization before metric calculation.Phantom backgrounds and volunteer brain masks are used; prospective accelerated data are reconstructed with the most promising retrospective methods.Source: §2.4.1 · p. 7
  • T1 maps are generated by dictionary matching with inversion efficiency set to 1 for phantom data and 0.85 for in-vivo data.Phantom T1 percentage error is assessed against fully sampled iFFT maps and documented phantom values; TI 6 is the MPRAGE-like T1-weighted contrast at 984 ms.Source: §2.4.1 · pp. 7–8

3 RESULTS

Across accelerated evaluations, INR maintains image detail and contrast where competing reconstructions blur or become noisy. The results also assess acquisition choices, phantom T1 accuracy, and the contrasts available from 1 mm³ MPnRAGE.

3.1 Phantom data

  • For phantom T1 estimates, shortest and longest values showed increased bias and variability, with short-T1 errors up to 150% for K = 2; intermediate values were slightly underestimated across methods.Figure 5B showed trends similar to fully sampled data; mean percentage error approached 50% even with K = 3 subspace bases.Source: §3.1 · p. 8

3.2.1 Acquisition aspects

  • The custom flip-angle schedule increased signal at TI 2–TI 5, strengthening the WM–GM contrast difference; complementary Poisson-disk sampling improved average PSNR to 32.57 dB and SSIM to 0.944, versus 29.47 and 0.904 without complementarity.Source: §3.2.1 · pp. 8–9

3.1 Phantom data

  • For phantom data at R = 20, zero-filled iFFT and PICS were strongly degraded, while LLR and INR retained sharp, reference-like contrast; INR also removed reference ringing.The supplied Figure 4 passage describes both methods as achieving high PSNR and SSIM.Source: §3.1 · p. 8, Figure 4 · p. 9

3.2.2 Evaluation of reconstruction and T1 mapping

  • INR showed sharper early-TI images than subspace reconstruction; for TI 1, PSNR/SSIM were 32.93 dB/0.923 versus 27.75 dB/0.820, and INR remained robust as prospective acceleration increased.Source: §3.2.2 · pp. 9–10
  • At R = 20, all joint methods produced generally good in-vivo images and T1 maps, but subspace images were blurry at TI 1 and noisier at TI 2–TI 3; CSF with long T1 was an exception in the error maps.Source: §3.2.2 · p. 9

3.2.3 MPnRAGE derived imaging

  • At 1 mm³ and R = 20, MPnRAGE TI 6 had contrast similar to MPRAGE, alongside WM-, GM-, and CSF-nulled contrasts and a quantitative T1 map.Source: §3.2.3 · p. 10

4 DISCUSSION

The discussion presents acquisition choices for accelerated Cartesian MPnRAGE and evaluates INR against established reconstructions. It also describes the resulting imaging contrasts and identifies computational and study-scope limitations.

4.1 Acquisition aspects

  • Cartesian readout enables slice-wise 2D INR reconstruction, while scan-matched calibration data provide sensitivity maps, background phase correction, and denser sampling for joint reconstruction.The calibration data use otherwise unused delays after further longitudinal recovery, reducing phase errors associated with magnetization regrowth.Source: §4.1 · p. 10
  • Continuous alternating k-space traversal across successive TI bins minimizes abrupt phase-encoding transitions, and complementary Poisson-disk sampling better preserves detail and improves PSNR and SSIM.The proposed variable flip-angle schedule is intended to increase signal in inherently low-SNR TI contrasts, though systematic optimization remains necessary.Source: §4.1 · p. 10, Figure 6 · p. 11

4.2 Evaluation of reconstruction and T1 mapping

  • INR is more robust to undersampling than subspace and iterative LLR, especially at shorter TIs, while preserving anatomy and more accurate T1 estimates for very short T1 values.Subspace with two bases can overestimate short T1 values; three bases better represent signal evolution but produce noisier, less sharp undersampled images.Source: Figure 7 · p. 11, §4.2 · p. 13
  • Fixed INR hyperparameters across subjects, resolutions, and acceleration factors, with reconstructions shown for five volunteers, suggest robustness across acquisition settings.INR reconstruction took about 13.8 minutes for a 1 mm^3 whole-brain R = 20 volume, versus approximately 6.6 minutes for subspace and 1.23 hours for iterative LLR.Source: §4.2 · p. 14

4.3 MPnRAGE derived imaging

  • A 5.39-minute MPnRAGE acquisition provides quantitative T1 maps and tissue-nulled contrasts alongside a T1-weighted image resembling clinical MPRAGE.The derived T1 maps show slight overestimation of longer T1 values, likely due to incomplete longitudinal recovery before subsequent inversion pulses.Source: §4.3 · p. 14

4.4 Limitations & Future directions

  • Slice-wise INR processing addresses memory and computation challenges but may introduce subtle slice-to-slice artifacts; whole-volume reconstruction time also remains too long for clinical image-quality assessment or rescanning.The study included healthy volunteers only; evaluation in patient data is planned.Source: §4.4 · p. 14

5 CONCLUSION

The conclusion presents a Cartesian 3D MPnRAGE acquisition with complementary variable-density Poisson-disk sampling and joint INR reconstruction. It reports high-quality accelerated images with multiple contrasts, quantitative T1 fitting, and a T1-weighted image closely matching clinical MPRAGE.

5 CONCLUSION

  • The proposed Cartesian MPnRAGE with joint INR reconstruction produces high-quality images at R = 20 in a 5.39-minute scan and outperforms state-of-the-art reconstruction methods.The sequence uses tailored complementary variable-density Poisson-disk sampling and calibration data acquired during otherwise unused delay time.Source: §5 · p. 15
  • Tailored view ordering keeps consecutively acquired TI samples spatially close in k-space while minimizing abrupt phase-encoding gradient changes.Source: §5 · p. 15
  • The resulting images support quantitative T1 fitting, tissue-nulled contrasts, and a T1-weighted image closely matching clinical MPRAGE.Source: §5 · p. 15

SUPPORTING INFORMATION

The supporting information includes phantom and in-vivo reconstructions, T1 maps, and an additional reconstruction-time table. Several figures compare reconstruction methods, sampling approaches, and acceleration settings.

Supporting information

  • Phantom figures show fully sampled 1.5 mm^3 and prospective 1 mm^3 data reconstructed with multiple methods, including INR, with short-T1 vial contrast differences noted for subspace variants.The prospective 1 mm^3 phantom results show sharp images with reference-like contrast for iterative LLR and INR.Source: Figure S1 · p. 18, Figure S2 · p. 18
  • At 1.5 mm3, retrospective in-vivo MPnRAGE undersampling at R = 28 and 34 degrades subspace and iterative LLR image quality more than INR, especially for the first three TI.The fully sampled 1.5 mm3 dataset had a 29.22 min scan time, and reconstructions were sharp with high PSNR and SSIM relative to iFFT.Source: Figure S3 · p. 19, Figure S6 · p. 21
  • A supplemental figure shows 1 mm^3 prospective R = 20 reconstructions from five volunteers using the same INR hyperparameters across volunteers, resolutions, and accelerations.The supporting information also includes a table of reconstruction times for prospective R = 20 3D data.Source: Table S1 · p. 19, Figure S7 · p. 22
Loading 2609.18589v1…