Source-linked AI summary
DuDoRNet: Learning a Dual-Domain Recurrent Network for Fast MRI Reconstruction with Deep T1 Prior
Bo Zhou, S. Kevin Zhou
TL;DR
Fast MRI reconstruction is limited by structural, non-local aliasing and underuse of fully sampled short-protocol information. DuDoRNet addresses both issues by recurrently restoring image and k-space domains with embedded T1 priors, and experiments show improved reconstruction quality and stability across sampling patterns and acceleration rates.
Problem
Existing fast MRI methods inadequately address structural, non-local aliasing and rarely use fully sampled short-protocol data such as T1 as complementary information.
Method
DuDoRNet recurrently restores image and k-space domains using DRD-Nets, embeds T1 priors in both domains, and applies data consistency regularization.
Results
Experiments across sampling patterns and acceleration rates show consistently improved reconstruction quality and stability, including PSNR of 32.511 dB with T1 prior for Cartesian acceleration and SSIM above 0.99 across spiral undersampling rates.
Takeaways & Limitations
The method efficiently restores undersampled MRI, while T1 priors further improve structural recovery across the evaluated acquisition settings.
Abstract
from arXiv · showhide
MRI with multiple protocols is commonly used for diagnosis, but it suffers from a long acquisition time, which yields the image quality vulnerable to say motion artifacts. To accelerate, various methods have been proposed to reconstruct full images from under-sampled k-space data. However, these algorithms are inadequate for two main reasons. Firstly, aliasing artifacts generated in the image domain are structural and non-local, so that sole image domain restoration is insufficient. Secondly, though MRI comprises multiple protocols during one exam, almost all previous studies only employ the reconstruction of an individual protocol using a highly distorted undersampled image as input, leaving the use of fully-sampled short protocol (say T1) as complementary information highly underexplored. In this work, we address the above two limitations by proposing a Dual Domain Recurrent Network (DuDoRNet) with deep T1 prior embedded to simultaneously recover k-space and images for accelerating the acquisition of MRI with a long imaging protocol. Specifically, a Dilated Residual Dense Network (DRDNet) is customized for dual domain restorations from undersampled MRI data. Extensive experiments on different sampling patterns and acceleration rates demonstrate that our method consistently outperforms state-of-the-art methods, and can reconstruct high-quality MRI.
1. Introduction
Fast MRI reconstruction is motivated by long, motion-sensitive acquisitions, but prior methods inadequately restore non-local aliasing and underuse short-protocol T1 information. DuDoRNet addresses these gaps through recurrent restoration in image and k-space domains with embedded T1 priors.
- Motivation: Long MRI acquisitions can introduce motion artifacts and delay patient care, motivating reconstruction from undersampled k-space data.The acquisition burden is especially relevant for multi-protocol examinations.
- Limitations of prior work: Existing deep-learning methods primarily restore images alone, although undersampling can distort or remove detailed structures and produce structural, non-local aliasing.This limits the adequacy of image-domain restoration for undersampled MRI.
- Limitations of prior work: Prior studies also underexplore fully sampled short-protocol MRI, such as T1-weighted imaging, as a deep prior for reconstructing longer protocols.T1WI is described as requiring less acquisition time than the longer T2WI and FLAIR protocols.
- Proposed approach: DuDoRNet learns image- and k-space-domain DRD-Nets recurrently while embedding T1 priors and data-consistency regularization.The design uses complementary information exchanged between the two domains linked by Fourier Transform.
2. Problem Formulation
The formulation reconstructs a fully sampled MRI image and k-space from undersampled measurements, coupling dual-domain restoration with data consistency. A recurrent optimization strategy addresses the complexity of jointly optimizing these objectives.
- Problem definition: The task is to reconstruct fully sampled image x from undersampled k-space ku and its reconstructed image xu.The image–k-space relationship is defined through Fourier and inverse Fourier transforms.
- Image-domain restoration: Image-domain restoration minimizes prediction error for estimating the fully sampled image from the undersampled image.A data-consistency constraint can be added to the prediction objective.
- K-space restoration: K-space restoration analogously estimates fully sampled k-space from undersampled k-space and combines with image restoration and data consistency.The combined target enforces the dual-domain reconstruction objectives.
- Recurrent optimization: The method optimizes image- and k-space-restoration parameters recurrently because directly optimizing multiple objective terms is computationally difficult and prone to overfitting and local optima.The recurrent strategy is introduced as an alternative to direct joint optimization.
3. Methods
DuDoRNet combines recurrent dual-domain restoration, deep T1 priors, and recurrent data consistency. Its DRD-Net backbone uses dilated residual dense modules and global feature fusion for both image and k-space restoration.
- DuDoRNet architecture: The pipeline feeds image- and k-space T1 priors into recurrent blocks containing separate restoration networks and recurrent data-consistency regulation.Each block uses one image-restoration DRD-Net, one k-space-restoration DRD-Net, and interleaved consistency operations.
- DRD-Net: DRD-Net is customized for MRI reconstruction and uses Squeeze-and-Excitation Dilated Residual Dense Blocks as its backbone.The same network design supports both image-domain and k-space-domain restoration.
- Global structure: DRD-Net comprises initial feature extraction, multiple SDRDBs with global feature fusion, and global residual learning.Initial extraction uses two sequential 3 × 3 convolutions, while fused features feed the reconstruction output.
- SDRDB: Each SDRDB combines four densely connected atrous convolutions, local feature fusion, Squeeze-and-Excitation, and local residual learning.The atrous convolutions use dilation rates of 1, 2, 4, and 4 to form a feature pyramid.
DRD Block SE
The method combines specialized SDRDB-based restoration with recurrent dual-domain processing, T1 priors, and data consistency to reconstruct MRI from undersampled data.
- SDR Block design: SDRDB uses densely connected atrous layers, local feature fusion, squeeze-and-excitation, and local residual learning.The block’s atrous layers form a feature pyramid with dilation rates 1, 2, 4, and 4.
- Dual-domain recurrent learning: DuDoRNet recurrent blocks contain image- and k-space-restoration DRD-Nets with interleaved data consistency layers.The network feeds T1 priors into both domains during recurrent processing.
- Dual-domain recurrent learning: Data consistency preserves fidelity at sampled k-space locations by combining measured and predicted values under the undersampling mask.The mixing coefficient λ controls this linear combination; λ = 0 directly substitutes sampled values.
- Dual-domain recurrent learning: Each recurrent block optimizes image- and k-space losses, while the final loss sums these losses across recurrent blocks.The implementation uses shared parameters across recurrent blocks and obtains the final image by inverse Fourier transform.
4. Experiments
Experiments evaluate DuDoRNet across protocols, sampling trajectories, acceleration rates, and component configurations. The method achieves strong reconstruction quality and stability, with T1 prior, dual-domain learning, recurrent learning, and spiral sampling providing important benefits.
- Experimental Settings: The dataset contains 20 patients scanned with fully sampled T1, T2, and FLAIR protocols, split patient-wise into 252 training, 36 validation, and 72 test images per protocol.Inputs use randomly undersampled T2/FLAIR and fully sampled T1 images, with acceleration factors from 1 to 6 across three sampling patterns.
- Experimental Settings: Evaluation compares DuDoRNet with TV-CS, GRAPPA, sequential convolution, DeepCas, RefGAN, and UF-T2 using SSIM, PSNR, and MSE against fully sampled images.CNN-based baselines are trained under the same settings as DuDoRNet.
- Reconstruction Results: At R = 5, T1 prior raises Cartesian PSNR from 27.834 dB to 32.511 dB and radial PSNR from 37.27 dB to 40.815 dB.For Cartesian sampling, DuDoRNet with T1 prior improves over UF-T2 from 30.594 dB to 32.511 dB; radial DuDoRNet also exceeds RefGAN's 35.11 dB.
- Reconstruction Results: At R = 5, spiral sampling gives the highest reported image quality, with PSNR increasing from 48.418 dB without T1 prior to 49.186 dB with it.Figure 5 presents sampling masks, difference images, enlarged details, and SSIM values for Cartesian, radial, and spiral trajectories.
- Reconstruction Stability: DuDoRNet maintains SSIM above 0.95 for Cartesian sampling and above 0.98 with T1 prior for radial sampling at R = 6.Without T1 prior, SSIM remains above 0.89 for Cartesian and 0.97 for radial sampling at R = 6.
- Reconstruction Stability: Spiral sampling provides the best stability, with DuDoRNet maintaining SSIM above 0.99 across the full undersampling range regardless of T1 prior.Radial and spiral trajectories sample k-space more uniformly than random Cartesian sampling, producing less aliasing in the initial reconstruction input.
- Ablation Studies: Recurrent learning and dual-domain learning each improve the baseline by 0.015, while combining them gives the largest component-combination boost.The full model containing recurrent, dual-domain, and dilated residual dense components produces the best reconstruction results.
- Ablation Studies: SSIM increases monotonically with recurrent-block count, although improvement begins converging after Nrec = 3 across all three sampling patterns.Spiral sampling remains best across different recurrent-block counts.
5. Conclusion
The paper presents DuDoRNet, which recurrently restores MRI in image and k-space domains while embedding T1 priors. Experiments show effective reconstruction and further structural recovery from T1 guidance; future work targets pathological data and other recovery tasks.
- Conclusion: DuDoRNet recurrently restores image and k-space domains with DRD-Nets while embedding T1 priors in both domains.The architecture uses large receptive fields to guide restoration across recurrent blocks.
- Conclusion: Experiments demonstrate that DuDoRNet can efficiently reduce reconstruction artifacts, while T1 prior further improves structural recovery.The conclusion contrasts the method with prior single-domain approaches for individual protocols.
- Conclusion: Future work includes testing DuDoRNet on pathological MR data and applying it to noise reduction and super-resolution.