MRI2Qmap: multi-parametric quantitative mapping with MRI-driven denoising priors

Mohammad Golbabaee; Matteo Cencini; Carolin Pirkl; Marion Menzel; Michela Tosetti; Bjoern Menze

physics.med-phMar 11, 2026

MRI2Qmap: multi-parametric quantitative mapping with MRI-driven denoising priors

Authors: Mohammad Golbabaee, Matteo Cencini, Carolin Pirkl, Marion Menzel, Michela Tosetti, Bjoern Menze

Organizations: University of Bristol · IRCCS Fondazione Stella Maris · GE Healthcare · TH Ingolstadt · University of Zurich

Abstract

Magnetic Resonance Fingerprinting (MRF) and other highly accelerated transient-state parameter mapping techniques enable simultaneous quantification of multiple tissue properties, but often suffer from aliasing artifacts due to compressed sampling. Incorporating spatial image priors can mitigate these artifacts, and deep learning has shown strong potential when large training datasets are available. However, extending this paradigm to MRF-type sequences remains challenging due to the scarcity of quantitative imaging data for training. Can this limitation be overcome by leveraging sources of training data from clinically-routine weighted MRI images? To this end, we introduce MRI2Qmap, a plug-and-play quantitative reconstruction framework that integrates the physical acquisition model with priors learned from deep denoising autoencoders pretrained on large multimodal weighted-MRI datasets. MRI2Qmap demonstrates that spatial-domain structural priors learned from independently acquired datasets of routine weighted-MRI images can be effectively used for quantitative MRI reconstruction. The proposed method is validated on highly accelerated 3D whole-brain MRF data from both in-vivo and simulated acquisitions, achieving competitive or superior performance relative to existing baselines without requiring ground-truth quantitative imaging data for training. By decoupling quantitative reconstruction from the need for ground-truth MRF training data, this framework points toward a scalable paradigm for quantitative MRI that can capitalize on the large and growing repositories of routine clinical MRI.

Figures & tables

Data/Method		MAPE (%) $\downarrow$			CD $\times 10^{-3}\downarrow$	PSNR (dB) $\uparrow$			SSIM (%) $\uparrow$
Data/Method		T1+T2	T1	T2	TSMI	T1w	T2w	PDw	T1w	T2w	PDw
In-vivo	SVDMRF	44.40	14.97 (0.94)	29.44 (2.22)	–	–	–	–	–
	ADMM	23.14	9.40 (0.79)	13.73 (1.27)	2.92 (0.50)	32.18 (0.62)	27.23 (0.65)	27.81 (0.60)	94.45 (0.78)	91.35 (1.05)	86.85 (1.20)
	LLR	16.17	8.38 (0.59)	7.79 (0.66)	3.25 (0.41)	32.93 (0.51)	28.44 (0.59)	29.09 (0.49)	95.13 (0.68)	92.97 (0.98)	86.43 (1.09)
	LRTV	14.18	6.53 (0.43)	7.65 (0.58)	2.71 (0.38)	33.53 (0.46)	28.84 (0.55)	30.58 (0.45)	97.02 (0.42)	93.98 (0.72)	90.65 (0.76)
	MRI2Qmap	12.01	4.93 (0.34)	7.08 (0.50)	0.85 (0.11)	35.91 (0.45)	30.03 (0.49)	30.63 (0.52)	97.95 (0.30)	95.26 (0.58)	91.42 (0.94)

TABLE I : Mean (±SD) performance on in-vivo and BrainWeb datasets at

R=8

(^{*})

indicates methods trained with ground-truth MRF maps. MRI2Qmap performs competitively with MRF-supervised baselines and outperforms remaining methods.

Panel of Figure 2 — Fig. 1 : MRI2Qmap vs. an ADMM reconstruction without spatial regularization. Quantitative T1/T2 maps and synthesized T1w, T2w, and PDw images are shown (electronic zoom recommended). By leveraging MRI-driven denoising priors, MRI2Qmap suppresses aliasing artifacts in synthesized contrasts and the underlying quantitative maps.

Panel of Figure 3 — Fig. 2 : Reconstructed T1 and T2 maps at $R=8$ acceleration. Relative to baselines ( ∗ indicates MRF-trained), MRI2Qmap exhibits reduced artifacts and improved anatomical delineation e.g. for deep brain structures (axial views) and the cerebellum (coronal views). Electronic zoom in boxed regions is recommended.

Fig. 3 : TSMI magnitudes across $s=5$ compressed coefficient channels (rows) for the subject in Fig. 2 , reconstructed by MRI2Qmap and baselines (columns). CG/init $\mathbf{x}_{\text{init}}$ initializes MRI2Qmap; ARNet was trained with MRF ground-truth data.

Fig. 4 : Reconstructed T1/T2 maps for additional in-vivo subjects: MRI2Qmap vs. representative baselines (LRTV/ARNet).

MRI2Qmap	MAPE (%) $\downarrow$			PSNR (dB) $\uparrow$
Modalities	T1+T2	T1	T2	T1w	T2w	PDw
All (base)	12.01	4.93	7.08	35.91	30.03	30.63
T2w PDw	14.59	7.50	7.09	33.75	29.94	30.66
T1w PDw	16.51	4.77	11.74	36.05	27.96	31.07
T1w T2w	19.19	10.44	8.75	36.01	29.36	27.28

TABLE II : MRI2Qmap performance with all or partial MRI modality priors, averaged over in-vivo data.

Fig. 5 : T1/T2 maps and synthesized MRIs (columns) from MRI2Qmap with one denoising modality prior disabled (rows); most affected images in each row are highlighted.

Fig. 6 : MRI2Qmap mapping performance on in-vivo data vs. T1w TR (blue) and T2w TE (orange) variations relative to the default synthesis values (TR ${}_{\text{base}}$ /TE ${}_{\text{base}}$ = 500/100 ms). Solid lines and shaded regions show mean MAPE ± standard deviation. Dashed lines mark practical parameter ranges.

	Total cost					MRI2Qmap cost/iter.
	LLR	LRTV	MRFPnP	ARNet	MRI2Qmap	Prox	DM	DN
Runtime $\approx$	778	528	547	89	684	40.1	9.9	10.6
GPU Mem.	33.9	33.9	36.9	34.8	36.9	36.9	19.5	26.5

TABLE III : (Left) Reconstruction times (s) and peak GPU memory (GB) usage of each method, averaged over in-vivo data. (Right) MRI2Qmap average per-iteration cost breakdown for

prox_{f}

, dictionary matching (DM), and denoising (DN).

Fig. 7 : MRI2Qmap denoising schedules $\sigma_{k}$ (top) and T1+T2 MAPEs (bottom) vs. iterations for in-vivo (left) and BrainWeb (right) reconstruction examples.

	in-vivo data					BrainWeb data
$\sigma_{\max}$	$K$ = 5	10	15	20	30	10	20	40	60	100
0.02	13.22	12.46	12.10	11.87	11.65	38.19	30.72	23.07	19.50	16.46
0.05	12.17	11.51	11.19	11.11	11.12	32.51	23.43	15.13	11.91	10.25
0.1	11.54	10.95	10.88	10.97	11.23	29.40	19.46	11.61	9.90	9.47
0.2	11.27	10.82	10.91	11.11	11.35	28.59	18.87	11.02	9.61	9.45
0.5	11.34	10.91	11.01	11.20	11.46	29.12	19.37	11.31	9.93	9.49

TABLE IV : Combined T1+T2 MAPE (%) across

K

\sigma_{\max}

on in-vivo /BrainWeb reconstruction examples.

$K_{cg}$	1	5	10	20	40
MAPE (%)	12.53	10.88	10.82	10.81	10.81
Avg./iter (s)	5.8	23.0	39.9	61.3	87.0

TABLE V : MRI2Qmap T1+T2 MAPE (%) and

prox_{f}

average runtime (s) vs. CG iterations (

K_{cg}

) on an in-vivo volume.

	Method	MAPE (%) $\downarrow$			PSNR (dB) $\uparrow$
	Method	T1+T2	T1	T2	T1w	T2w	PDw
$R=2$	LRTV	7.94	3.45	4.49	40.32	34.80	35.30
	MRFPnP	7.56	3.15	4.41	41.66	35.74	36.23
	MRI2Qmap	7.50	3.25	4.25	41.58	35.44	35.13
$R=4$	LRTV	11.01	4.88	6.13	35.88	31.18	32.54
	MRFPnP	10.07	4.20	5.87	38.62	32.85	34.02

TABLE VI : MRI2Qmap performance vs. selected baselines on acceleration factors

R=2,4,6

, averaged on in-vivo dataset.

$K_{cg}$	5	10	20	40	60
MAPE (%)	20.18	11.67	9.61	9.49	9.49
Avg./iter (s)	2.5	4.5	8.4	14.4	18.1

TABLE S.1 : MRI2Qmap T1+T2 MAPE (%) and

prox_{f}

average runtime (s) vs. CG iterations (

K_{cg}

) on BrainWeb data.

Panel of Figure 15 — Fig. S.1 : Reconstructed T1/T2 maps (axial and coronal views) at acceleration factor $R=8$ for a representative simulated BrainWeb subject, comparing MRI2Qmap with MRF baselines ( ∗ indicates MRF-trained methods). Percentage error (PE) maps relative to the ground-truth reference are shown below each reconstruction.

Panel of Figure 16 — Fig. S.2 : Reconstructed T1 and T2 maps (axial view) at $R=8$ for the remaining seven subjects in in-vivo dataset, comparing MRI2Qmap with representative baselines (LRTV and ARNet). Relative to baselines, MRI2Qmap exhibits reduced artifacts and improved anatomical delineation (electronic zooming recommended).

Panel of Figure 17 — Fig. S.3 : Reconstructed T1 and T2 maps (coronal view) at $R=8$ for the remaining seven subjects in in-vivo dataset, comparing MRI2Qmap with representative baselines (LRTV and ARNet). Relative to baselines, MRI2Qmap exhibits reduced artifacts and improved anatomical delineation (electronic zooming recommended).

Panel of Figure 18 — Fig. S.4 : Reconstructed T1 and T2 maps at acquisition acceleration factors $R=2,4,6,8$ (rows) for a representative in-vivo test image. Relative to baselines ( ∗ indicates MRF-trained), MRI2Qmap exhibits reduced artifacts and improved delineation of deep brain structures (electronic zoom in boxed regions is recommended).

Explore similar work

A physics-informed foundation model for quantitative diffusion MRI
May 29, 2026Zihan Li, Jialan Zheng, Ziyu Li +18Physics-Informed Diffusion Models Medical Imaging Foundation Models
School of Biomedical Engineering, Tsinghua University, Beijing, China. · Oxford Centre for Integrative Neuroimaging, FMRIB, Nuffield Department of Clinical Neurosciences, University of Oxford, Oxford, United Kingdom. · Department of Radiology, West China Second University Hospital, Sichuan University, Chengdu, China. +4
MoE-dqINR: A Unified Mixture-of-Experts Implicit Neural Representation Framework for Scan-Specific Dynamic and Quantitative MRI Reconstruction
May 29, 2026Yinzhe Wu, Fanwen Wang, Zhenxuan Zhang +3Implicit Neural Representations Mixture of Experts
Department of Bioengineering and I-X, Imperial College London, London, SW7 2AZ, United Kingdom · Cardiovascular Research Centre, Royal Brompton Hospital, London, SW3 6NP, United Kingdom · Shanghai Pudong Hospital and Human Phenome Institute, Fudan University, Shanghai, China +3
Pixelwise Uncertainty Quantification of Accelerated MRI Reconstruction
Jan 19, 2026Ilias I. Giannakopoulos, Lokesh B Gautham Muthukumar, Yvonne W. Lui +1Uncertainty Quantification Accelerated MRI Reconstruction
Bernard and Irene Schwartz Center for Biomedical Imaging, Department of Radiology, NYU Grossman School of Medicine, 10016, New York, NY, United States of America · Courant Institute of Mathematical Sciences, NYU, 10012, New York, NY, United States of America · Center for Advanced Imaging Innovation and Research (CAI2R), Department of Radiology, NYU Grossman School of Medicine, 10016, New York, NY, United States of America

May 29, 2026J/K move · Enter open · S save

A physics-informed foundation model for quantitative diffusion MRI

Zihan Li, Jialan Zheng, Ziyu Li, Xun Yuan+17

School of Biomedical Engineering, Tsinghua University, Beijing, China. · Oxford Centre for Integrative Neuroimaging, FMRIB, Nuffield Department of Clinical Neurosciences, University of Oxford, Oxford, United Kingdom. · Department of Radiology, West China Second University Hospital, Sichuan University, Chengdu, China.+4

Physics-Informed Diffusion Models Medical Imaging Foundation Models Magnetic Resonance Imaging Physics-Informed ML Diffusion MRI Brain MRI

Open paper AskPDF arXiv

Understanding the human brain requires access to its microscopic tissue architecture. Diffusion magnetic resonance imaging (MRI) provides the only noninvasive window into whole-brain microstructure in vivo, yet reliable quantitative mapping remains confined to specialized research settings requiring dense sampling and optimized acquisition protocols. To address this gap, we present a physics-informed generative microstructure network (PIGMENT) that learns a universal generative prior of human brain microstructure and adapts it zero-shot to each participant's measured data to recover subject-specific maps. Trained on 11375 scans spanning multiple sites, vendors, and field strengths, PIGMENT enabled reliable quantitative mapping for tensor, kurtosis, and NODDI models across external datasets from five independent centers. It remains effective where conventional fitting becomes unreliable, recovering meaningful maps from extremely sparse acquisitions while supporting downstream tractography and structural connectivity mapping. PIGMENT estimates demonstrated strong biological validity, preserving submillimeter cortical microarchitectural patterns and early-childhood white matter developmental trajectories from 10-fold accelerated scans. Furthermore, PIGMENT enables reliable quantitative tensor mapping on cost-efficient low-field systems and the extraction of tumor-related biomarkers using ultra-fast clinical protocols. Together, these results establish PIGMENT as a physics-informed foundation model that extends quantitative diffusion MRI into regimes traditionally too sparse, heterogeneous, or clinically constrained for reliable analysis.

Abstract

Explore similar work

A physics-informed foundation model for quantitative diffusion MRI

MoE-dqINR: A Unified Mixture-of-Experts Implicit Neural Representation Framework for Scan-Specific Dynamic and Quantitative MRI Reconstruction

Pixelwise Uncertainty Quantification of Accelerated MRI Reconstruction