Making blood oxygenation level dependent imaging interpretable: predicting regional cerebral metabolic rate of oxygen and blood flow via deep learning

BOLD fMRI is the most widely used neuroimaging modality for mapping brain function, yet its relationship to cerebral metabolism is indirect and remains incompletely understood. Direct measures of cerebral blood flow (CBF), oxygen extraction fraction (OEF), and cerebral metabolic rate of oxygen (CMRO) provide important insight into neuroenergetics and vascular coupling, but they require specialized quantitative MRI or PET acquisitions that are more costly and less widely available than standard BOLD imaging. Recent work has further shown that BOLD signal changes can diverge from underlying oxygen metabolic changes across cortex, highlighting the need for mechanistically interpretable BOLD-derived markers of brain physiology. Here, a CNN-Vision Transformer model was developed to predict measured CBF, OEF, and CMRO directly from BOLD fMRI.

The model was trained on the public OpenNeuro dataset ds004873 [1], which contains task-based multiparametric quantitative BOLD data from 40 healthy participants. The dataset includes reference maps derived from multiecho T₂ and T₂* measurements, pseudo-continuous arterial spin labeling for CBF, dynamic susceptibility contrast imaging for blood volume (CBV), and BOLD fMRI. OEF was estimated from relaxation rate differences, and CMRO was derived using the Davis model. The analysis was strictly restricted to gray matter voxels (GM probability > 0.50). White matter and CSF-prone regions, as well as voxels violating physiological reliability bounds (OEF > 0.9, T₂ > 90 ms, R₂’ > 9 s⁻¹, and CBV > 10%), were excluded in adherence with the dataset authors’ recommendations.

Preprocessed BOLD data were converted into 3D patches with 51-channel feature representations, capturing baseline BOLD intensity, spatial gradients, temporal signal fluctuations, and spectral features. Patches were extracted across task conditions from 37 subjects with sufficiently complete data. Data were split at the subject level into training (25 subjects), validation (5 subjects), and test (7 subjects) sets. The CNN-Vision Transformer architecture used convolutional downsampling, transformer-based spatial token integration, and decoder upsampling to generate voxelwise predictions of CBF, OEF, and CMRO. Both loss calculation and final evaluation metrics were strictly restricted to physiologically plausible target ranges (5 < CBF < 150 ml/100g/min; 0.1 < OEF < 0.9; 10 < CMRO₂ < 600 µmol/100g/min) to avoid artificial inflation of performance from background voxels.

On held-out test data, predicted values showed good to moderate agreement with measured targets, with Pearson correlations of approximately 0.7 for CBF, 0.6 for OEF, and 0.5 for CMRO₂. Permutation feature importance revealed that baseline BOLD intensity and spatial tissue gradients drove over 85% of the model’s predictive performance, demonstrating a heavy reliance on underlying tissue contrast and vascular boundaries. These findings establish the feasibility of synthesizing quantitative metabolic maps from standard BOLD acquisitions. Ongoing cross-dataset validation of this approach could ultimately unlock retrospective metabolic analyses of existing large-scale clinical cohorts, providing accessible, non-invasive biomarkers for neurovascular dysfunction.

References

1. Epp, S., Castrillon, G., Yuan, B., Andrews-Hanna, J., Preibisch, C., Riedl, V., 2025. BOLD signal changes can oppose oxygen metabolism across the human cortex. Nat. Neurosci.(2025)