Source-linked AI summary
DoReMe: Dual optimization of Reconstruction and Motion estimation for self-navigated free-breathing abdominal MRI using Gaussian Splatting
Majd Helo, Dominik Nickel, Thomas Kuestner
TL;DR
Respiratory motion makes free-breathing abdominal MRI susceptible to artifacts, while existing reconstruction pipelines may rely on fixed motion estimates or respiratory surrogate signals. DoReMe jointly refines image estimates and Gaussian motion models from self-navigated data; in volunteer experiments it improved visual image quality over compressed sensing, including in shortened acquisitions.
Problem
Free-breathing abdominal MRI combines data acquired while anatomy moves, causing k-space inconsistencies and artifacts; many motion-resolved pipelines use fixed pre-estimated fields, while some coupled methods require respiratory surrogate information.
Source: Key words · p. 2Method
DoReMe alternates data-consistent image reconstruction with refitting a compact, groupwise Gaussian representation of respiratory deformation to jointly refine images and motion fields.
Source: Key words · p. 3, 2.2Optimization algorithm · p. 20Results
In volunteer experiments, DoReMe improved visual image quality over CS, preserved structure under retrospective undersampling, and completed reconstruction and motion updates in approximately 1.5 minutes.
Source: 2.2Optimization algorithm · pp. 18, 19, 21Takeaways & Limitations
The evaluated Gaussian primitives provided a compact, explicit motion parameterization suitable for repeated fitting during reconstruction.
Source: 2.2Optimization algorithm · p. 18The study involved a small volunteer cohort, and reader assessments are needed before clinical use.
Source: 2.2Optimization algorithm · p. 20
Abstract
from arXiv · showhide
Purpose: Free-breathing abdominal MRI is limited by respiratory motion, which introduces image artifacts and k-space inconsistencies. Existing reconstruction approaches often depend on pre-estimated motion fields or external respiratory surrogate signals. We propose a coupled reconstruction-motion estimation algorithm for self-navigated free-breathing acquisition, in which motion is represented by compact Gaussian primitives and jointly refined with the image estimates during iterative reconstruction. Methods: An iterative reconstruction that alternates between data consistency updates and motion modelling was proposed. Respiratory motion was represented using Gaussian primitives. A Gaussian splatting was integrated into the reconstruction algorithm. The method was evaluated on self-navigated free-breathing Cartesian gradient-echo (GRE) acquisition in four healthy volunteers, including retrospective scan shortening experiments, and was compared with regularized compressed sensing (CS) reconstruction. Results: Respiratory motion states were reconstructed in approximately 1.5 min. Compared with CS, the proposed method improved visual image quality, particularly in coronal liver views, and better preserved anatomical structures in retrospectively shortened acquisitions. These findings provide proof of concept for the proposed method. Conclusion: The proposed coupled reconstruction-motion estimation algorithm improved visual image quality compared with CS alone. Gaussian primitives provided a compact and sufficiently flexible representation of nonrigid respiratory motion with practical runtime, supporting further scanner integration and validation in larger cohorts.
Key words:
DoReMe jointly refines image reconstruction and Gaussian-based respiratory motion estimates through alternating data-consistent reconstruction and motion estimation. The passages describe its evaluation, runtime and convergence, and qualitative image findings in comparison with CS.
2.2Optimization algorithm:
- The algorithm uses inexact FISTA, updating motion parameters after each inertial extrapolation step, and alternates data-consistency updates with motion estimation from current images.Data were assigned to eight respiratory states using stable amplitude binning with readouts distributed equally across bins.Source: 2.2Optimization algorithm · pp. 7–9
- The proposed algorithm converged consistently in the experiments and can be guaranteed to converge to a local minimum with monotone FISTA.The motion-estimation objective is non-convex, so the alternating optimization does not guarantee a global optimum.Source: Key words · p. 5, 2.2Optimization algorithm · p. 8
- Evaluation involved four healthy volunteers and directly compared the proposed joint reconstruction-motion estimation algorithm with CS reconstruction after retrospectively retaining the first 50% and 25% of sequentially acquired readouts.Separately, Gaussian primitive counts ranged from 50 to 50000 with 100 optimization iterations per fitting step. ANTs registration used SyN and the middle motion state as reference [57].Source: 2.2Optimization algorithm · p. 10
- Compared with CS, DoReMe had slightly higher DC error but consistently lower TV regularization values and a lower total objective across four volunteers.The motion regularization term rose initially, peaked after approximately 5–7 iterations, and then stabilized.Source: Figure 2 · p. 11, 2.2Optimization algorithm · p. 12
- Approximately 1:30 min:s was required for the complete joint reconstruction-motion estimation algorithm.For 30 reconstruction iterations, FISTA took approximately 8 s; adding Gaussian motion estimation with 5000 primitives added approximately 45 s, and Jacobian-based regularization added a further 45 s. Maximum GPU memory was 14 GB for datasets of size 8 x 88 x 320 x 312 (motion states x slices x width x height).Source: 2.2Optimization algorithm · pp. 11–12
- DoReMe improved depiction of hepatic boundaries and internal structures over CS, especially in coronal views.The comparison showed breath-hold images alongside end-inhalation free-breathing reconstructions; CS was also described as providing acceptable image quality.Source: Figure 3 · p. 12, 2.2Optimization algorithm · p. 12
- Joint image and motion estimates converged around iterations 18–20, with only minor visual changes afterward.The motion fields became spatially coherent as they were refined together with image content.Source: Figure 4 · p. 13, 2.2Optimization algorithm · p. 14
- DoReMe achieved SSIM 0.9866 ± 0.0078 and PSNR 44.27 ± 3.09 dB, with lower NRMSE, HFEN, and photometric error than ANTs, but 0.0143 ± 0.0192% non-positive Jacobians.ANTs produced no voxels with non-positive Jacobian determinants; the registration metrics compare alignment to the middle respiratory state.Source: Table 2 · p. 14, 2.2Optimization algorithm · pp. 14, 19
- Increasing Gaussian primitives from 50 to 50000 improved all evaluated registration metrics, but 5000 primitives offered a practical quality–runtime compromise.From 5000 to 50000 primitives, SSIM improved by only 0.0009 and PSNR by 0.46 dB, while mean reconstruction time rose from 170.84 to 462.60 s; these metrics do not measure accuracy against independent ground truth.Source: Figure 5 · p. 15, 2.2Optimization algorithm · pp. 15, 19
- DoReMe preserved more consistent anatomy and higher visual image quality than CS across respiratory states and retrospectively shortened acquisitions, especially in coronal views and states 4–8.The undersampled acquisitions retained 50% (1:26 min) or 25% (43 s) of readouts.Source: Figure 6 · p. 16, 2.2Optimization algorithm · pp. 16, 18, 20, Figure 7 · p. 17
- The method alternates data-consistent image reconstruction with refitting a compact, shared Gaussian motion model, completing updates in approximately 1.5 minutes on evaluated datasets.The model shares Gaussian centers, covariance parameters, and blending weights across motion states while using state-specific displacement vectors.Source: 2.2Optimization algorithm · pp. 18–20
- Clinical translation requires larger-cohort and patient validation, reader assessments, known-motion phantom testing, and scanner integration; respiratory binning and computational cost remain limitations.The study involved a small volunteer cohort, and the current Python implementation was not integrated on-scanner for prospective studies.Source: 2.2Optimization algorithm · pp. 20–21