Brain oxygen metabolism without gadolinium: a generative model needs only perfusion and T2* scans, with errors below day-to-day retest
Synopsis
Regional cerebral metabolic rate of oxygen (CMRO₂) previously required quantitative BOLD MRI with a gadolinium contrast agent plus blood sampling for hematocrit. A conditional generative adversarial network reduced the input to two contrast-free standard scans, cerebral blood flow and T2*, and on a 41-person healthy-adult test set its synthetic maps reached a correlation of 0.89 with the analytical reference, the same as a three-input model that also used T2, with a median individual error of 38.1 versus the analytical method's between-day retest error of 46.5 µmol/100 g/min. Agreement held in new participants under the same protocol and fell to r = 0.81 on a scanner from another vendor.
Interpretation
Regional cerebral oxygen metabolism can now be synthesized from two standard MRI scans that use no contrast agent, removing the need for gadolinium, blood sampling for hematocrit and complex multiparametric post-processing. The earlier quantitative BOLD route needed four maps, cerebral blood flow, T2, T2* and cerebral blood volume, and the blood-volume map required an injected gadolinium agent that can leave trace amounts in the brain. On the 41-person healthy-adult test set, the model using only cerebral blood flow and T2* reached a voxel-wise correlation of 0.89 (95% CI 0.88–0.90) with the analytical reference, the same as the three-input model that also used T2, and its median individual error of 38.1 stayed below the analytical method's between-day retest error of 46.5 µmol/100 g/min.
The synthetic maps preserved physiological variation beyond image similarity, reproducing both task-evoked metabolic changes and the age-related decline in brain oxygen metabolism. Models trained on T1-weighted anatomy or resting-state functional MRI could not replace the quantitative BOLD inputs, reaching correlations of only 0.38 and 0.51, so anatomy alone does not explain the agreement. In 10 test participants who performed a calculation task, the task-evoked pattern overlapped the analytical one with a Dice coefficient of 0.78; across the 41 test participants, uncorrected synthetic values declined by 6.8 µmol/100 g/min per decade against 5.0 for the analytical values.
Applied without retraining to new participants under the same protocol and to a scanner from another vendor, the model kept its agreement on same-protocol data and lost some on cross-vendor data. Medical imaging models are often evaluated at the training site, and validation on independently acquired data is frequently missing. Correlation was 0.89 in the 24-person same-protocol dataset, comparable to the test set, and 0.81 in the 12-person cross-vendor dataset, where the median individual error of 58.9 exceeded that dataset's between-day retest error of 52.8 µmol/100 g/min.
Perspective
The findings rest on retrospective single-center data from healthy adults at 3 T and 4-mm resolution, restricted to cortical gray matter; the same-protocol and cross-vendor datasets were used for testing only and did not enter training.
Synthetic values inherit the model assumptions and systematic errors of quantitative BOLD; the inputs lack individual arterial oxygen content, so hematocrit differences between women and men are not captured; age coincided with the arterial spin labeling readout in the test set, so age effects cannot be fully separated from protocol effects; reliability was assessed within a single session; after age calibration the synthetic and analytical declines did not differ significantly, but the confidence interval still allows a gap of up to 2.4 µmol/100 g/min per decade, and agreement across the lifespan remains to be established in larger cohorts.
