Voxel-wise tracer kinetic model selection for DCE-MRI measurements of blood-brain barrier leakage

To apply voxel-wise tracer kinetic model selection, characterise the spatial distribution of best-fitting models across the brain, and evaluate whether model selection improves sensitivity for differentiating normal-appearing tissue from pathological tissue compared to the Patlak model.

Methods: Extended Tofts, Patlak, and intravascular models were fit to DCE-MRI data from stroke survivors and controls, as well as simulated data. The best-fitting model was chosen for each voxel using the Akaike Information Criterion, and model selection $K^{\text{trans}}$ (estimates from the best-fitting model for each voxel) compared to Patlak model $K^{\text{trans}}$ .

In simulated data, the Extended Tofts model was best-fitting at $K^{\text{trans}}>{10}^{-3}{min}^{-1}$ , where the Patlak model systematically underestimated $K^{\text{trans}}$ . Patlak was optimal at $K^{\text{trans}}$ between ${10}^{-4}-{10}^{-3}{min}^{-1}$ , where Extended Tofts estimates had greater variability. The intravascular model was selected for $K^{\text{trans}}\sim {10}^{-4}{min}^{-1}$ . The Patlak model was chosen in most control voxels. In chronic stroke, the Extended Tofts model was preferred in most cortical and white matter hyperintensity voxels, while the Patlak model was selected in most deep grey matter and normal-appearing white matter voxels. Model selection $K^{\text{trans}}$ estimates were significantly greater than Patlak estimates in the cortex and white matter hyperintensities, with greater inter-patient variability, likely reflecting biological variability in blood-brain barrier leakage resulting from stroke.

Voxel-wise model selection may provide more accurate estimates of a wider range of $K^{\text{trans}}$ values than any single model, revealing greater differences between normal and pathological tissue and offering a more sensitive and physiologically appropriate framework for DCE-MRI analysis of blood-brain barrier dysfunction.