Brain oxygen metabolism: a GAN reconstructs it from two standard MRI scans, in 95 healthy adults from a single centre
A team from Erlangen and Munich trained a generative network to produce maps of the brain's oxygen consumption from two routine MRI sequences, without any contrast injection. In 41 healthy test adults, the synthetic maps closely track the reference measurement (correlation of 0.89) and recover the decline of metabolism with age, but exaggerate it by about a third. It is a careful, open proof of feasibility that is still far from a clinical tool: no patients were studied and accuracy drops on another MRI vendor.
The context
The brain uses roughly one fifth of the body's oxygen. The rate of that consumption, called CMRO2 (cerebral metabolic rate of oxygen, in micromoles per 100 g of tissue per minute), is a direct marker of the energy activity of neurons. It falls with age and is thought to be altered early in neurodegenerative diseases such as Alzheimer's disease, which makes it a sought-after measurement.
The problem is measuring it. The historical reference is oxygen-15 positron emission tomography (PET), which requires an on-site cyclotron and radiation exposure, so it remains confined to a few research centres. MRI methods exist, such as the mqBOLD (multiparametric quantitative BOLD) approach used here, but they combine several sequences, including a cerebral blood volume measurement that needs a gadolinium injection. Hence this team's idea: teach a neural network to infer the CMRO2 map from fewer sequences, without injection.
Synthesising one quantitative image from others with deep learning is an active field in imaging, including on Tatakoto with synthetic thin-slice CT. The risk is always the same: producing maps that look plausible but poorly reproduce what clinicians care about, here the differences between individuals.
The method
The model is a conditional GAN (generative adversarial network: two networks trained against each other, a generator that produces the map and a discriminator that tries to tell it apart from a real one). The generator is a 3D U-Net, a funnel-shaped architecture that compresses then rebuilds the image, with residual blocks and skip connections filtered by an attention mechanism. Training combines three penalties: the voxel-wise absolute error, the discriminator's adversarial loss, and a structural similarity measure (SSIM).
The reference is CMRO2 computed by mqBOLD using Fick's principle: the oxygen extraction fraction multiplied by cerebral blood flow (CBF, measured by arterial spin labelling, ASL, without injection) and by arterial oxygen content, estimated from each participant's haematocrit and pulse oximetry. The authors tested several input combinations: all mqBOLD maps, without the injected blood volume, CBF and T2* only (two standard contrast-free sequences), CBF alone, an anatomical T1 MRI alone, or a connectivity map from resting-state functional MRI.
The data come from a single Munich centre, on 3-tesla scanners, between May 2015 and June 2025: 95 healthy adults (58 women, mean age 44) from four earlier studies, manually split into 54 for training and 41 for testing, balancing age and sex. The test set spans 21 to 84 years in two separate groups: 20 young adults aged 21 to 41 and 21 older adults aged 61 to 84. The model was then applied without retraining to two external sets: 24 young men scanned with the same protocol and 12 people scanned on another vendor's machine. Analyses cover cortical grey matter at 4 mm resolution.
The results
On the test set, the full model reproduces the reference map with a correlation r = 0.92 (95% confidence interval 0.91 to 0.93) and a mean absolute error (MAE) of 12.7 µmol/100 g/min at group level. The minimal model, CBF plus T2*, does almost as well: r = 0.89 (0.88 to 0.90), MAE 17.9. CBF alone is clearly worse (r = 0.79), and anatomical T1 or functional MRI are not enough (r = 0.38 and 0.51). Metabolism cannot be guessed from anatomy: physiological information is required.
At the individual level, the minimal model's median error (38.1) stays below the variation seen when the same person is scanned with the reference method on two different days (46.5). Within-session repeatability, measured in 9 participants, matches that of the reference (median intraclass correlation coefficient of 0.87 in both cases). During a mental arithmetic task, the model recovers the activation map with an overlap coefficient (Dice) of 0.78 to 0.81, against 0.59 with CBF alone.
Outside the training data, performance holds on the same protocol (r = 0.89, MAE 26.8) but drops on the other vendor (r = 0.81, MAE 40.0). There, the median individual error (58.9) exceeds day-to-day variability (52.8).
On the age-related decline, the reference gives −5.0 µmol/100 g/min per decade and the model −6.8, a significant overestimate of 1.8 (0.4 to 3.2). After an age recalibration the gap is no longer significant (1.1; −0.4 to 2.4), which does not prove equivalence, as the authors acknowledge. In concrete terms: between ages 30 and 70, the reference indicates a drop of about 20 µmol/100 g/min, the raw model about 27. A study using these maps to compare young and older people would get the direction of the effect right but exaggerate its size.
What is good
First, a well-posed question and an honestly tested hierarchy of inputs. Rather than announcing one model, the authors compare six sequence combinations and show that a two-sequence, contrast-free model (CBF and T2*) approaches the full model, while a plain anatomical MRI fails. That negative result on T1 is as useful as the positive one.
Second, an evaluation that goes beyond group-level correlation. Individual error compared with test-retest, repeatability, response to a cognitive task, age effect, and two external sets including a cross-vendor one: most image-synthesis papers stop at visual similarity and SSIM.
Third, open and independent science. The code and full pipeline are on GitHub (NeuroenergeticsLab/synthCMRO2), two of the datasets are already on OpenNeuro, funding is public (European Research Council and Deutsche Forschungsgemeinschaft) and no conflicts of interest are declared. The limitations, including the overestimated decline, are written in the text.
What is less good
First, a marked population bias: healthy adults only, from a single centre, with 41 people in the test set. Yet the clinical interest of CMRO2 lies precisely in disease, where blood volume and oxygen extraction can vary independently of flow. The model learned in healthy subjects the usual relationship between flow, T2* and metabolism; nothing guarantees it holds in a patient after a stroke or with dementia, exactly where that relationship breaks down.
Second, a risk of shortcut learning tied to a confounder. In the test set, young and older participants were not scanned with the same ASL protocol (different labelling delay and readout). Age is therefore confounded with acquisition technique, and the authors acknowledge they cannot separate the two. The overestimated decline may come from atrophy at 4 mm, as they suggest, but also from the protocol.
Third, a misleading metric and an imperfect reference. The correlations of 0.89 to 0.92 are computed over regions pooled across the group, mixing anatomical differences between regions with differences between people; the more telling individual error is about twice as high. And the ground truth is not oxygen-15 PET but mqBOLD, itself an estimate with its own systematic errors, which the model inherits. Finally, the drop on another vendor indicates that deployment will require retraining or site-specific fine-tuning.
What it changes
For the research community, this preprint provides a reproducible pipeline and a demonstration that two contrast-free sequences carry most of the information needed to estimate cortical metabolism in healthy subjects. Next steps: patients (dementia, stroke, tumours), comparison with oxygen-15 PET, multiple centres and between-day repeatability.
For clinicians, nothing changes today. This is a research prototype with no regulatory status, not evaluated in patients; the authors themselves state that the preprint should not guide clinical practice.
For patients and the public, the prospect is a measurement of brain metabolism during an ordinary MRI, without injection or radioactivity, making it available beyond the few centres with a cyclotron. This study shows it is plausible in healthy adults; it does not yet show that the map would be reliable in someone who is ill.
Further reading
The preprint: Generative Deep Learning Maps Brain Oxygen Metabolism and Its Age-Related Decline from Two Standard Noninvasive MRI Scans, Castrillon, Marcolongo, Riedl et al., Uniklinikum Erlangen (FAU), Technical University of Munich, University of Padua and Imperial College London, medRxiv, posted 19 November 2026, DOI 10.64898/2026.11.17.26368289.
The code: NeuroenergeticsLab/synthCMRO2 on GitHub (the licence is not stated in the preprint). Public data: OpenNeuro ds004873 and ds008786.
On Tatakoto: AI-synthesised thin-slice CT and label leakage in Alzheimer's MRI models.
Editorial transparency: this article was produced by Tatakoto's automated editorial pipeline from version 1 of the preprint (medRxiv PDF, main text; Appendix S1 and supplementary tables not consulted). The "between 30 and 70" conversion is our own calculation from the published slopes. The English, Spanish and Chinese versions are AI-generated. Spotted an error? Contact us.