A new AI-powered MRI framework called multiplexed MRI (MRx), developed by researchers at the University of Illinois Urbana-Champaign and published in Nature on 6 May 2026, can simultaneously map more than 20 quantitative structural, physiological and molecular biomarkers of the whole brain in a single 14-minute scan on standard clinical hardware – offering clinicians an unprecedented window into brain tumour grading, multiple sclerosis lesion staging, and early disease detection that conventional MRI simply cannot provide.
For decades, MRI has been the workhorse of neurological diagnosis. Yet for all its power, the technology has remained stubbornly limited in one critical respect: clinicians typically acquire one qualitative biomarker at a time, piecing together a picture of brain pathology from a sequence of separate scans that can collectively take up to an hour. A research team led by Zhi-Pei Liang, professor of electrical and computer engineering and a member of the Beckman Institute for Advanced Science and Technology at the University of Illinois Urbana-Champaign, has now fundamentally changed that equation.
The clinical implications are immediate and substantial. In patients with brain tumours, MRx distinguished between tumour grades that looked identical on conventional imaging. In multiple sclerosis, it staged lesions without the need for contrast agents – a meaningful advantage for patients who undergo repeated scanning over many years. And it did all of this within a timeframe that fits comfortably into routine clinical workflows.
What MRx actually does
Conventional MRI exploits magnetic resonance signals from water molecules to generate anatomical images. Separately, MR spectroscopy and spectroscopic imaging can detect signals from metabolites such as Nacetylaspartate (NAA), creatine, choline and neurotransmitters including GABA and glutamate – but these techniques require their own acquisition protocols, deliver far lower spatial resolution, and have never been feasible to run simultaneously with structural imaging in clinical settings.
MRx dissolves that barrier. Rather than selectively exciting one or a few molecular targets, the MRx acquisition framework uses wideband radio-frequency pulses to excite all detectable molecules in the proton spin system simultaneously. Rapid echo-planar sampling trajectories then capture the resulting signals, with complementary encoding modules handling transverse relaxation, longitudinal relaxation and J-evolution in parallel. To keep scan time clinically acceptable, the team designed a sparse sampling scheme that exploits the partial separability of high-dimensional signals, image sparsity, parallel imaging and mutual information between biomarkers.
The data processing challenge – essentially reconstructing a high-dimensional image from sparse, noisy measurements where water signals dwarf molecular signals by three to four orders of magnitude – is addressed using physics-based machine learning. Subspace models represent the multiparametric spatial-spectral structure of the signals parsimoniously, while neural networks incorporate data-driven spectral, spatial and parameter distribution priors. As study co-author Rong Guo, formerly in Liang’s group and now a senior scientist at Siemens Healthineers, explained: “With our integration of ultrafast data acquisition and physics-based machine learning methods for data processing, MRx overcomes several longstanding bottlenecks to fast, high-resolution multiplexed imaging.”
The result, in a healthy volunteer, is simultaneous whole-brain maps of 22 bio-markers: molecular maps of NAA, creatine, choline, myo-inositol, glutamine, glutathione, taurine, NAAG, glutamate and GABA at 2 × 3 × 3 mm³ resolution, alongside tissue property maps – oxygen extraction fraction, deoxygenated blood volume, T2′, T2*, quantitative susceptibility mapping, axonal water fraction, extracellular water fraction, myelin water fraction, T1, T2 and proton density – at 1 × 1 × 1 mm³ resolution. The entire acquisition takes approximately 14 minutes.
Validation and reproducibility
Before applying MRx to patients, the team rigorously validated the technology in phantom studies. Molecular biomarker accuracy was confirmed using a home-made spectroscopic phantom, with measured concentrations showing no significant biases even in tubes of 8 mm diameter. Relaxation parameter accuracy was assessed against the ISMRM/NIST relaxometry phantom, with MRx-derived T1 and T2 values achieving regression slopes of 0.9792 and 1.0116 respectively, and R² values above 0.99 in both cases.
Reproducibility was evaluated across intra-session, inter-session and cross-centre experiments in 48 standard neuroimaging regions of interest. For most water-based parameters, coefficients of variation remained below 4% within sessions and below 7% across centres. For the main metabolites, cross-centre variation stayed below 9%. For quantitative susceptibility mapping, intraclass correlation coefficients of 0.9853, 0.9102 and 0.9074 were achieved for intra-session, inter-session and cross-centre reproducibility respectively – performance the authors describe as comparable to or better than existing MR techniques.
Brain tumours: seeing what conventional MRI misses
The translational power of MRx becomes clear in the tumour results. In a patient with grade IV glioblastoma multiforme, the full panel of MRx biomarkers revealed a detailed map of the pathological processes underlying both tumour and surrounding oedema tissue: elevated choline indicating cellular proliferation, raised lactate reflecting anaerobic glycolysis, reduced NAA signalling neuronal loss, glutamine elevation consistent with alternative energy sourcing, increased glutathione marking oxidative stress, and hypoxia evidenced by changes in T2*, deoxygenated blood volume and oxygen extraction fraction, among others.
Perhaps more clinically striking were the findings in oligodendroglioma. Conventional MRI – including FLAIR and contrast-enhanced T1-weighted sequences – failed entirely to distinguish between grade II and grade III tumours in two patients, both exhibiting hyperintensity on FLAIR and no noticeable contrast enhancement. MRx, by contrast, revealed substantially higher choline (58.8%, P < 0.001) and lactate (84.3%, P < 0.001) in the grade III tumour, reflecting greater cellular proliferation and anaerobic glycolysis, findings corroborated by biopsy histopathology.
By deriving a composite tissue state index from the full biomarker set using machine learning, the team was able to clearly separate eight distinct tissue types – grey matter, white matter, cerebrospinal fluid, oedema, meningioma, low-grade oligodendroglioma, high-grade oligodendroglioma and glioblastoma – in a single scalar map. Standard multi-parametric MRI based on T1, T2 and proton density alone could not separate these states.
Multiple sclerosis: staging lesions without contrast
In MS, the challenge is not just detecting lesions but characterising them. Active, chronic active and chronic lesions have distinct pathological profiles and different implications for treatment, yet current clinical MRI modalities cannot reliably distinguish between them without contrast agents, or at all in some cases.
MRx addressed this directly. In one MS patient with both active and chronic lesions, the MRx tissue state index map correctly identified and differentiated the two lesion types, whereas the standard multiparametric MRI tissue state index mislabelled them. The mechanistic underpinning was clear in the biomarker data: lactate was significantly elevated in active lesions (210.2%, P < 0.001), consistent with increased aerobic hyperglycolysis during acute inflammation; the myo-inositol/ NAA ratio was higher in chronic lesions (32.4%, P < 0.001), reflecting gliosis accumulation; and active lesions showed increased T2 (32.7%, P < 0.001) and markedly reduced myelin water fraction (510.8%, P < 0.001), indicating ongoing demyelination. All of this was achieved without the contrast agents currently required in standard clinical practice.
MRx also showed predictive capability. In a separate MS patient, baseline MRx biomarkers obtained at an initial scan correctly predicted which lesion voxels would evolve, remain stable or shrink over a four-month follow-up. A pre-lesion invisible on FLAIR showed early elevation in myo-inositol/NAA, signalling gliosis and axonal dysfunction before any structural change was apparent.
A new era for brain imaging
Liang was measured but clear about where this technology leads: “MRx can be a powerful tool for noninvasive tissue characterisation, helping to advance personalised, precision and predictive medicine. By providing rich, multidimensional biomarkers to capture disease progression and treatment response, this capability could open new opportunities for more precise diagnosis, individualised treatment planning and improved patient outcomes.”
The authors also point beyond the bio-markers demonstrated in this study. MRx is extensible to diffusion, perfusion, magnetisation transfer and elastography-based tissue properties, and with multinuclear RF systems could eventually image biomarkers from deuterium, sodium and phosphorus simultaneously. The team also foresees a role in the development of digital brain twins – AI-powered generative models capable of predicting disease trajectories – for which MRx could supply the informative imaging biomarkers that current technology cannot provide.
As first author Yudu Li, professor of bioengineering at Illinois, noted: “Diseases such as tumours, multiple sclerosis and neurodegenerative disorders are highly heterogeneous. The rich set of biomarkers obtained using MRx has the potential to provide deeper insights into brain function and disease processes, while also improving the sensitivity and specificity of detection and diagnosis.”
With whole-brain scanning completed in 14 minutes on standard 3T clinical systems, the barrier to clinical adoption is considerably lower than might be expected for a technology of this scope. Whether the complexity of AI-driven data processing and the need for clinical validation at scale can be navigated efficiently will determine how quickly MRx moves from research to routine practice – but the case it makes for doing so is compelling.
Journal reference:
Li, Y., Guo, R., Zhao, Y., et al. (2026). Multiplexed magnetic resonance imaging. Nature, 653, 411–417.
https://doi.org/10.1038/s41586-026-10475-x
