Unsupervised Probabilistic Model Averaging across Nested Transport Models for Brain Permeability Mapping at 0.55 Tesla
Abstract
Purpose
Dynamic-contrast–enhanced (DCE) MRI enables quantitative assessment of microvascular permeability but is constrained at low field by reduced SNR and ambiguous voxel-wise model selection. This study constructs and evaluates an unsupervised probabilistic-nested-model-selection (PNMS) framework for voxel-wise brain permeability mapping using 0.55T DCE-MRI on a Siemens MAGNETOM-Free.Max system, and benchmarks its performance against deterministic nested-model-selection.
Methods
Four untreated patients with brain metastases underwent whole-brain DCE-MRI (IRB#: 15622-01) using a sequence-optimized clinical protocol designed for low-field stability. Baseline T1-mapping was performed using dual-flip-angle VIBE (2°/15°), and dynamic imaging employed TWIST (TR/TE/FA=4.48/2.02/15°) with 7.04-s temporal resolution and whole-brain coverage (60-slices, 0.97×0.97×3mm³). Signal time courses were converted to longitudinal relaxation-rate changes (ΔR1). Voxel-wise deterministic nested-model-selection (NMS) evaluated three physiologically nested vascular models: normal vasculature without leakage (Model-1), tumor tissue with unidirectional leakage (Model-2), and tumor vasculature with bidirectional exchange (Model-3). Standard NMS pharmacokinetic analysis estimated permeability parameters for each model (vp1 for Model-1; vp2 and Ktrans2 for Model-2; vp3, Ktrans3 and kep3 for Model-3). To augment training and increase model variability, ±30% perturbations around the mean parameter values and their combinations were analytically simulated to generate additional ΔR₁ profiles. In total, 292,630 normalized voxel-wise ΔR₁ profiles with corresponding NMS labels were used to train a Kohonen self-organizing map (K-SOM; 8×8 topology). K-SOM–based PNMS (50% probability-threshold) was compared with deterministic NMS across all patients using Dice-similarity-coefficients (DSC).
Results
PNMS demonstrated strong agreement with deterministic NMS in leaky tumor regions (DSC=0.782 [0.739-0.825] for Model-2 and 0.873 [0.837-0.909] for Model-3). PNMS showed less sensitivity to arterial input dispersion and noise-induced overfitting of Model-1 regions, producing spatially coherent probability maps that captured voxel-wise heterogeneity and model uncertainty at 0.55T.
Conclusion
The proposed unsupervised PNMS enables robust and physiologically meaningful brain permeability mapping at 0.55T by explicitly accounting for model uncertainty, supporting quantitative DCE-MRI on modern low-field MRI systems.