M-Gaussian: An Magnetic Gaussian Framework for Efficient Multi-Stack MRI Reconstruction

Kangyuan Zheng; Xuan Cai; Jiangqi Wang; Guixing Fu; Zhuoshuo Li; Yazhou Chen; Xinting Ge; Liangqiong Qu

cs.CVFeb 24, 2026

M-Gaussian: An Magnetic Gaussian Framework for Efficient Multi-Stack MRI Reconstruction

Authors: Kangyuan Zheng, Xuan Cai, Jiangqi Wang, Guixing Fu, Zhuoshuo Li, Yazhou Chen, Xinting Ge, Liangqiong Qu, +1 more

Organizations: School of Biomedical Engineering, Shenzhen Campus of Sun Yat-sen University, Shenzhen, 518107, China. · the School of Information Science and Engineering, Shandong Normal University, Shandong, China. · School of Computing and Data Science, The University of Hong Kong, Hong Kong, China.

Abstract

Magnetic Resonance Imaging (MRI) is a crucial non-invasive imaging modality. In routine clinical practice, multi-stack thick-slice acquisitions are widely used to reduce scan time and motion sensitivity, particularly in challenging scenarios such as fetal brain imaging. However, the resulting severe through-plane anisotropy compromises volumetric analysis and downstream quantitative assessment, necessitating robust reconstruction of isotropic high-resolution volumes. Implicit neural representation methods, while achieving high quality, suffer from computational inefficiency due to complex network structures. We present M-Gaussian, adapting 3D Gaussian Splatting to MRI reconstruction. Our contributions include: (1) Magnetic Gaussian primitives with physics-consistent volumetric rendering, (2) neural residual field for high-frequency detail refinement, and (3) multi-resolution progressive training. Our method achieves an optimal balance between quality and speed. On the FeTA dataset, M-Gaussian achieves 40.31 dB PSNR while being 14 times faster, representing the first successful adaptation of 3D Gaussian Splatting to multi-stack MRI reconstruction.

Figures & tables

Figure 1: Overview of the M-Gaussian framework. (a) The training pipeline integrates Gaussian rendering with a Neural Residual Field (NRF) refinement module. (b) Point cloud construction is performed via multi-stack registration and devoxelization to initialize the Gaussians. (c) The inference stage generates the final volume through Gaussian sampling and aggregation at the target resolution.

Figure 2: Comparison between original 3D Gaussian Splatting primitives and the proposed Magnetic Gaussian primitives. The original 3DGS utilizes view-dependent spherical harmonic coefficients for color representation, whereas M-Gaussian employs a single intensity value to model tissue-specific MRI signal properties, which significantly reduces memory overhead.

Methods	PSNR (dB) $\uparrow$	SSIM $\uparrow$	NCC $\uparrow$	NRMSE $\downarrow$	Runtime (min) $\downarrow$
FeTA Dataset
NiftyMIC [ 6 ]	36.08	0.9857	0.9932	0.0157	79.12
SVRTK [ 11 ]	34.64	0.9861	0.9916	0.0185	14.75
NeSVoR [ 17 ]	31.35	0.9588	0.9546	0.0414	31.75
M-Gaussian (Ours)	40.31	0.9936	0.9975	0.0096	5.63
FaBiAN Dataset

Table 1: Quantitative comparison with baseline slice-to-volume MRI reconstruction methods across three datasets. Metrics include PSNR, SSIM, NCC, and NRMSE for reconstruction quality, and runtime for computational efficiency. Best results are in bold , second-best are underlined .

Figure 3: Qualitative comparison of slice-to-volume MRI reconstruction methods. Each row shows results from a different dataset (FeTA, FaBiAN, HCP). The Input column displays one representative stack from the multi-stack acquisition. Our M-Gaussian method produces reconstructions with sharper anatomical boundaries and fewer artifacts compared to baseline methods.

Figure 4: Qualitative ablation results on Progressive resolution (P.R.) training and SSIM loss. Progressive resolution training stabilizes convergence, while the inclusion of SSIM loss significantly enhances the preservation of structural integrity and contrast. The effect is particularly significant on FeTA dataset.

Figure 5: Qualitative ablation results on Neural Residual Field (NRF). The NRF suppresses noise and refines high-frequency details, which is prominent on the HCP dataset.

Figure 6: Qualitative ablation results on Anisotropic Regularization (A.R.). The absence of A.R. leads to needle-like artifacts due to degenerate Gaussian geometries. These artifacts are effectively mitigated in the full model, which is most pronounced on the FaBiAN dataset.

Configuration	FeTA		FaBiAN		HCP
Configuration	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM
Full model	40.31	0.9936	32.26	0.9521	33.10	0.9776
w/o SSIM	32.24	0.9611	31.01	0.9388	32.56	0.9582
w/o SSIM & P.R.	31.07	0.9573	28.43	0.9272	31.36	0.9512
w/o NRF	39.11	0.9893	30.93	0.9387	31.96	0.9538
w/o A.R.	39.49	0.9902	30.02	0.9486	32.82	0.9606

Table 2: Ablation study quantifying the contribution of structural similarity loss (SSIM), progressive resolution training (P.R.), Neural Residual Field (NRF), and anisotropic regularization (A.R.).

Resolution	PSNR $\uparrow$	SSIM $\uparrow$	Time (min) $\downarrow$
$50^{3}$	30.27	0.9352	10.82
$150^{3}$	31.31	0.9505	16.43
$250^{3}$	31.50	0.9527	24.70
$350^{3}$	31.76	0.9547	27.07
$400^{3}$	31.63	0.9546	30.30

Table 3: Impact of Gaussian grid resolution on reconstruction fidelity and training time using the HCP dataset. The models were trained with fixed resolutions, excluding the effect of progressive training strategy.

Block Radius	PSNR $\uparrow$	SSIM $\uparrow$	Time (min) $\downarrow$
3	32.40	0.9733	7.62
4	32.67	0.9768	11.13
5	33.10	0.9776	17.28
6	32.77	0.9778	26.32
7	32.89	0.9780	34.24

Table 4: Quantitative assessment of reconstruction performance under different block radius settings. The results illustrate the trade-off between reconstruction quality and runtime.

Structure	Dice $\uparrow$				HD95 (mm) $\downarrow$				ASSD (mm) $\downarrow$
Structure	NiftyMIC	SVRTK	NeSVoR	M-Gaussian	NiftyMIC	SVRTK	NeSVoR	M-Gaussian	NiftyMIC	SVRTK	NeSVoR	M-Gaussian
External CSF	0.859	0.883	0.888	0.905	1.43	1.42	1.10	0.97	0.55	0.53	0.46	0.41
Grey Matter	0.832	0.816	0.843	0.846	0.82	0.85	0.68	0.66	0.37	0.38	0.34	0.32
White Matter	0.945	0.937	0.951	0.953	0.95	1.11	0.80	0.69	0.39	0.46	0.36	0.34
Ventricles	0.914	0.818	0.906	0.917	0.64	1.23	0.65	0.63	0.28	0.52	0.30	0.28
Cerebellum	0.924	0.934	0.941	0.931	1.12	0.94	0.87	1.05	0.46	0.40	0.36	0.42

Table 5: Quantitative evaluation of segmentation performance on the FeTA dataset. Metrics include Dice, HD95, and ASSD. Best results are bold , second-best are underlined .

Figure 9: Qualitative comparison of downstream segmentation results on the FeTA dataset. M-Gaussian reconstructions yield segmentation masks with the closest correspondence to ground truth, particularly for large-scale structures such as the cortical ribbon and ventricular boundaries.

Explore similar work

Physics-Driven 3D Gaussian Rendering for Zero-Shot MRI Super-Resolution
Mar 10, 2026Shuting Liu, Lei Zhang, Wei Huang +2Medical Image Super-Resolution Magnetic Resonance Imaging
School of Artificial Intelligence, Sichuan University, Chengdu 610065, China · Institute of High Performance Computing, A*STAR, Singapore 138632
SGAM: Shared Gaussian Geometry with Implicit Amplitude Modeling for Scan-Specific 3D Multi-Contrast MRI Reconstruction
Sep 21, 2026Jingran Xu, Dong Liang, Hairong Zheng +2Accelerated MRI Reconstruction Magnetic Resonance Imaging
Paul C. Lauterbur Research Center for Biomedical Imaging, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, Guangdong, China
GaussianPile: A Unified Sparse Gaussian Splatting Framework for Slice-based Volumetric Reconstruction
Mar 21, 2026Di Kong, Yikai Wang, Wenjie Guo +83D Reconstruction 3D Gaussian Splatting
Tsinghua University · Zhongguancun Academy · Beijing Normal University +3

Mar 10, 2026J/K move · Enter open · S save

Physics-Driven 3D Gaussian Rendering for Zero-Shot MRI Super-Resolution

Shuting Liu, Lei Zhang, Wei Huang, Zhao Zhang+1

School of Artificial Intelligence, Sichuan University, Chengdu 610065, China · Institute of High Performance Computing, A*STAR, Singapore 138632

Medical Image Super-Resolution Magnetic Resonance Imaging Physics-Guided Image Restoration 3D Gaussian Splatting

Open paper AskPDF arXiv

High-resolution Magnetic Resonance Imaging (MRI) is vital for clinical diagnosis but limited by long acquisition times and motion artifacts. Super-resolution (SR) reconstructs low-resolution scans into high-resolution images, yet existing methods are mutually constrained: paired-data methods achieve efficiency only by relying on costly aligned datasets, while implicit neural representation approaches avoid such data needs at the expense of heavy computation. We propose a zero-shot MRI SR framework using explicit Gaussian representation to balance data requirements and efficiency. MRI-tailored Gaussian parameters embed tissue physical properties, reducing learnable parameters while preserving MR signal fidelity. A physics-grounded volume rendering strategy models MRI signal formation via normalized Gaussian aggregation. Additionally, a brick-based order-independent rasterization scheme enables highly parallel 3D computation, lowering training and inference costs. Experiments on two public MRI datasets show superior reconstruction quality and efficiency, demonstrating the method's potential for clinical MRI SR.

Abstract

Explore similar work

Physics-Driven 3D Gaussian Rendering for Zero-Shot MRI Super-Resolution

SGAM: Shared Gaussian Geometry with Implicit Amplitude Modeling for Scan-Specific 3D Multi-Contrast MRI Reconstruction

GaussianPile: A Unified Sparse Gaussian Splatting Framework for Slice-based Volumetric Reconstruction