Physics Guided Multi-contrast Magnetic Resonance Reconstruction Diffusion Model

doi:10.11938/cjmr20263215

Abstract

Abstract:

Multi-contrast MRI allows for the simultaneous acquisition of multiple weighted images, which improves imaging efficiency and provides rich quantitative information. However, reconstructing these images under high undersampling while ensuring anatomical consistency and physical plausibility remains a significant challenge. Existing methods often rely on scarce fully-sampled data or suffer from performance degradation due to domain shift. To address this, we propose a physics-prior-guided diffusion model that encodes MR signal evolution via Bloch equations into a dictionary-matching constraint. This constraint is directly coupled into the reverse sampling process, enabling physical correction of data-driven priors without retraining. Experimental results demonstrate superior generalization performance over supervised and self-supervised approaches, while the estimated parameter maps validate its high quantitative accuracy, highlighting its potential for clinical multi-contrast imaging.

Key words: magnetic resonance imaging (MRI), multi-contrast imaging, diffusion models, unsupervised learning, physics priors

CLC Number:

O482.53

SU Yilin, LIU Yuanyuan, CUI Zhuoxu, LIANG Dong. Physics Guided Multi-contrast Magnetic Resonance Reconstruction Diffusion Model[J]. Chinese Journal of Magnetic Resonance, 2026, 43(3): 321-338.

Figures/Tables 15

Fig. 1

Table 1

Fig. 2

Table 2

Fig. 3

Table 3

Fig. 4

Table 4

Fig. 5

Fig. 6

Table 5

Fig. 7

Table 6

Fig 8

Table 7

References 34

[1]	BROWN M A, SEMELKA R C. MRI: basic principles and applications[M]. Hoboken: John Wiley & Sons, 2011.
[2]	SALTARELLI G, DI CERBO G, INNOCENZI A, et al. Quantitative MRI in neuroimaging: a review of techniques, biomarkers, and emerging clinical applications[J]. Brain Sci, 2025, 15(10): 1088. doi: 10.3390/brainsci15101088
[3]	JONES D K, ALEXANDER D C, CHETCUTI K, et al. Low field, high impact: democratizing MRI for clinical and research innovation[J]. BJR Open, 2025, 7(1): tzaf022.
[4]	FARRAR T C. Pulse nuclear magnetic resonance spectroscopy: an introduction to the theory and applications[M]. Madison: Farragut Press, 1997.
[5]	CHEN Q, YANG Z J, CHENG X Y, et al. Application of magnetic resonance imaging technology in pediatric exercise intervention research[J]. Chinese J Magn Reson, 2025, 42(2): 195-204.
	陈群, 杨子剑, 程心怡, 等. 磁共振成像技术在儿童运动干预研究中的应用[J]. 波谱学杂志, 2025, 42(2): 195-204. doi: 10.11938/cjmr20243129
[6]	MA L C, BAO Q J, MARTINHO R P, et al. Fast T₁ mapping MRI in preclinical and clinical settings using subspace-constrained joint-domain reconstructions[J]. Magn Reson Lett, 2024, 4(4): 200134.
[7]	MIAO J L, WAN X Y, FU J Y, et al. Application of multi-contrast quantitative MR imaging in central nervous system[J]. Chinese J Magn Reson Imaging, 2024, 15(4): 165-170.
	缪佳丽, 万欣月, 付君言, 等. 多对比度定量磁共振成像在中枢神经系统中的应用[J]. 磁共振成像, 2024, 15(4): 165-170.
[8]	ZHANG J, THANH D N, EDDY S, et al. McLARO: Multi-contrast learned acquisition and reconstruction optimization for simultaneous multi-contrast and multi-parametric mapping[J]. Magn Reson Med, 2023, 91(1): 344-356. doi: 10.1002/mrm.v91.1
[9]	LIU Y, CUI Z, QIN S, et al. Score-based diffusion models with self-supervised learning for accelerated 3D multi-contrast cardiac MR imaging[J]. IEEE Trans Med Imaging, 2025, 44(6): 2436-2448. doi: 10.1109/TMI.2025.3534206
[10]	LEI P, HU L, FANG F, et al. Joint under-sampling pattern and dual-domain reconstruction for accelerating multi-contrast MRI[J]. IEEE Trans Image Process, 2024, 33: 4686-4701. doi: 10.1109/TIP.2024.3445729
[11]	YAGHOODI N, CHAND J R, CHEN Y, et al. Fast multi-contrast MRI using joint multiscale energy model[C]// 2025 IEEE 22nd International Symposium on Biomedical Imaging (ISBI), Houston, TX, USA. Piscataway: IEEE, 2025: 1-5.
[12]	QU X, HOU Y, LAM F, et al. Magnetic resonance image reconstruction from undersampled measurements using a patch-based nonlocal operator[J]. Med Image Anal, 2014, 18(6): 843-856. doi: 10.1016/j.media.2013.09.007 pmid: 24176973
[13]	BUSTIN A, DA CRUZ G L, JAUBERT O, et al. High-dimensionality undersampled patch-based reconstruction (HD-PROST) for accelerated multi-contrast MRI[J]. Magn Reson Med, 2019, 81(6): 3705-3719. doi: 10.1002/mrm.27694 pmid: 30834594
[14]	SONG P, WEIZMAN L, MOTA J F C, et al. Coupled dictionary learning for multi-contrast MRI reconstruction[J]. IEEE Trans Med Imaging, 2019, 39(3): 621-633. doi: 10.1109/TMI.42
[15]	LIANG D, CHENG J, KE Z, et al. Deep magnetic resonance image reconstruction: Inverse problems meet neural networks[J]. IEEE Signal Process Mag, 2020, 37(1): 141-151. doi: 10.1109/MSP.79
[16]	KNOLL F, HAMMERNIK K, ZHANG C, et al. Deep-learning methods for parallel magnetic resonance imaging reconstruction: a survey of the current approaches, trends, and issues[J]. IEEE Signal Process Mag, 2020, 37(1): 128-140.
[17]	SRIRAM A, ZBONTAR J, MURRELL T, et al. End-to-end variational networks for accelerated MRI reconstruction[C]// MARTEL A L, et al. Medical Image Computing and Computer Assisted Intervention-MICCAI 2020, Lima, Peru. Cham: Springer, 2020: 64-73.
[18]	LI G, LV J, TIAN Y, et al. Transformer-empowered multi-scale contextual matching and aggregation for multi-contrast MRI super-resolution[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, New Orleans, LA, USA. Piscataway: IEEE, 2022: 20636-20645.
[19]	FENG C M, YAN Y, YU K, et al. Exploring separable attention for multi-contrast MR image super-resolution[J]. IEEE Trans Neural Netw Learn Syst, 2024, 35(9): 12251-12262. doi: 10.1109/TNNLS.2023.3253557
[20]	LI B, HU W, FENG C M, et al. Multi-contrast complementary learning for accelerated mr imaging[J]. IEEE J Biomed Health Inform, 2023, 28(3): 1436-1447. doi: 10.1109/JBHI.2023.3348328
[21]	HO J, JAIN A, ABBEEL P. Denoising diffusion probabilistic models[J]. Adv Neural Inf Process Syst, 2020, 33: 6840-6851.
[22]	PENG W, ADELI E, BOSSCHIETER T, et al. Generating realistic brain mris via a conditional diffusion probabilistic model[C]// Medical Image Computing and Computer-Assisted Intervention-MICCAI 2023, Vancouver, BC, Canada. Cham: Springer, 2023: 14-24.
[23]	XIE T, CUI Z X, LUO C, et al. Joint diffusion: mutual consistency-driven diffusion model for PET-MRI co-reconstruction[J]. Phys Med Biol, 2024, 69(15): 155019. doi: 10.1088/1361-6560/ad6117
[24]	GUNGOR A, DAR S U H, OZTURK Ş, et al. Adaptive diffusion priors for accelerated MRI reconstruction[J]. Med Image Anal, 2023, 88: 102872. doi: 10.1016/j.media.2023.102872
[25]	DARESTANI M Z, LIU J, HECKEL R. Test-time training can close the natural distribution shift performance gap in deep learning based compressed sensing[C]// International Conference on Machine Learning, Baltimore, MD, USA. Brooklyn: PMLR, 2022: 4754-4776.
[26]	SONG Y, SOHL-DICKSTEIN J, KINGMA D P, et al. Score-based generative modeling through stochastic differential equations[PP/OL]. arXiv(2020-11-26)[2026-04-30]. https://arxiv.org/abs/2011.13456.
[27]	SANDINO C M, LAI P, VASANAWALA S S, et al. Accelerating cardiac cine MRI using a deep learning-based ESPIRiT reconstruction[J]. Magn Reson Med, 2021, 85(1): 152-167. doi: 10.1002/mrm.v85.1
[28]	KORKMAZ Y, CUKUR T, PATEL V M. Self-supervised MRI reconstruction with unrolled diffusion models[C]// Medical Image Computing and Computer-Assisted Intervention-MICCAI 2023, Vancouver, BC, Canada. Cham: Springer, 2023: 491-501.
[29]	YAMAN B, HOSSEINI S A H, MOELLER S, et al. Self-supervised learning of physics-guided reconstruction neural networks without fully sampled reference data[J]. Magn Reson Med, 2020, 84(6): 3172-3191. doi: 10.1002/mrm.v84.6
[30]	CHUNG H, RYU D, MCCANN M T, et al. Solving 3D inverse problems using pre-trained 2D diffusion models[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Vancouver, BC, Canada. Piscataway: IEEE, 2023: 22542-22551.
[31]	WANG C, LYU J, WANG S, et al. CMRxRecon: a publicly available k-space dataset and benchmark to advance deep learning for cardiac MRI[J]. Sci Data, 2024, 11(1): 687. doi: 10.1038/s41597-024-03525-4 pmid: 38918497
[32]	SLAVKOVA K P, DICARLO J C, WADHWA V, et al. An untrained deep learning method for reconstructing dynamic MR images from accelerated model-based data[J]. Magn Reson Med, 2023, 89(4): 1617-1633. doi: 10.1002/mrm.v89.4
[33]	CHUNG H, LEE S, YE J C. Decomposed diffusion sampler for accelerating large-scale inverse problems[PP/OL]. arXiv(2023-03-10)[2026-04-30]. https://arxiv.org/abs/2303.05754.
[34]	HOU R, LI F, ZENG T. Fast and reliable score-based generative model for parallel MRI[J]. IEEE Trans Neural Netw Learn Syst, 2023, 36(1): 953-966. doi: 10.1109/TNNLS.2023.3333538

R	Method	PSNR/dB	SSIM	RMSE
6	PGDS	23.651±2.582	0.715±0.043	0.068±0.018
	BCNN	21.579±2.643	0.767±0.042	0.087±0.022
	SSDU	20.736±2.963	0.748±0.046	0.097±0.028
	DLESPIRIT	17.990±2.608	0.617±0.063	0.131±0.033
	D3IP	17.936±2.902	0.606±0.049	0.133±0.039
14	PGDS	21.214±2.714	0.651±0.047	0.091±0.024
	BCNN	19.603±2.651	0.689±0.047	0.109±0.028
	SSDU	19.398±2.945	0.685±0.053	0.113±0.032
	DLESPIRIT	16.924±2.656	0.579±0.060	0.148±0.037
	D3IP	16.963±2.980	0.582±0.053	0.149±0.044

R	Method	PSNR/dB	SSIM	RMSE
4	PGDS	37.656±1.125	0.694±0.012	0.013±0.002
	BCNN	29.582±4.947	0.620±0.096	0.037±0.017
	SSDU	28.863±2.528	0.560±0.079	0.037±0.011
	DLESPIRIT	26.394±2.304	0.449±0.048	0.049±0.014
	D3IP	33.191±2.038	0.614±0.037	0.022±0.006
6	PGDS	34.688±2.094	0.615±0.035	0.019±0.005
	BCNN	27.944±2.824	0.511±0.083	0.041±0.012
	SSDU	27.629±2.878	0.475±0.078	0.043±0.015
	DLESPIRIT	24.076±0.263	0.384±0.020	0.063±0.002
	D3IP	29.613±2.136	0.521±0.036	0.034±0.009
8	PGDS	32.143±2.355	0.556±0.041	0.025±0.007
	BCNN	26.354±2.023	0.446±0.081	0.049±0.011
	SSDU	26.385±2.874	0.426±0.077	0.050±0.016
	DLESPIRIT	22.234±0.785	0.338±0.020	0.078±0.007
	D3IP	27.411±2.279	0.477±0.042	0.044±0.012

R	Method	PSNR/dB	SSIM	RMSE
8	PGDS	37.537±1.174	0.963±0.003	0.013±0.001
	DDS	36.573±1.278	0.926±0.010	0.015±0.002
	FRSGM	33.021±0.733	0.873±0.008	0.022±0.002
12	PGDS	37.007±1.132	0.951±0.003	0.014±0.002
	DDS	35.874±1.163	0.910±0.010	0.016±0.002
	FRSGM	30.351±0.605	0.834±0.016	0.030±0.002

Reference	Method	R=6	R=14
T₁ (1476.2±599.7)	PGDS	1473.2±38.4	1475.7±40.2
	BCNN	1415.0±28.0	1396.4±26.2
	SSDU	1484.9±27.6	1637.0±41.8
	DLESPIRIT	1230.9±28.7	1353.2±29.6
	D3IP	1459.7±29.9	1494.1±22.9
T_1ρ (58.5±22.5)	PGDS	57.4±1.6	58.1±1.7
	BCNN	54.5±12.7	50.4±7.9
	SSDU	58.3±9.2	52.3±12.1
	DLESPIRIT	58.9±10.1	63.7±7.3
	D3IP	66.5±10.1	65.4±13.5

Reference	Method	R=4	R=6	R=8
	PGDS	37.1±2.4	37.7±2.0	40.1±3.4
T₂	BCNN	39.2±3.6	38.9±3.4	40.4±3.2
(36.0±2.7)	SSDU	38.9±2.2	39.3±2.3	32.0±3.9
	DLESPIRIT	49.5±17.3	91.7±6.0	99.8±7.1
	D3IP	39.5±2.0	43.8±1.9	42.1±4.3