Dynamic PET model selection gets a consistency-model sampler: 97.97% accuracy in simulation, under 4 seconds per slice
Synopsis
Dynamic PET must pick a kinetic model for each voxel's tracer curve, and the standard route fits every candidate by nonlinear least squares and compares AIC or BIC, which is slow and favors simple models under noise. A consistency model folds the continuous kinetic parameters and the discrete model index into one target vector conditioned on the measured time-activity curve and input function, so one forward sampling pass yields the joint posterior. On 7,000 simulated FDG total-body PET curves it reached 97.97% selection accuracy across seven compartment models versus 58.67% for NLS+AIC, and 91.4% versus 73.5% for NLS+BIC in simulated brain dopamine-release detection.
Interpretation
Voxel-wise kinetic model selection in dynamic PET can now return posterior probabilities in one pass instead of fitting each candidate separately and comparing information criteria. The conventional route fits every candidate model by nonlinear least squares and picks a winner by AIC or BIC; those criteria assume Gaussian noise and apply a fixed parameter penalty, so they over-penalize complex models under high voxel-level noise. On 7,000 simulated test curves (1,000 per model across seven models), the consistency model reached 97.97% overall classification accuracy versus 58.67% for NLS+AIC; NLS+AIC fell to 17.4% for 3TC and 8.0% for 3TC-D.
The method concatenates the continuous kinetic parameters and the discrete model index into one target vector, so the network learns the joint posterior conditioned on the measured curve. Consistency models were built to map Gaussian noise to a posterior in very few steps, and had previously been used for voxel-wise parameter posterior estimation with two-tissue compartment models. The conditioning vector joins the 35-frame voxel time-activity curve with the arterial input function; inference uses 3-step multistep consistency sampling and processes a coronal slice of roughly one hundred thousand curves in under 4 seconds on a GPU workstation with 4 NVIDIA A100 GPUs.
The same joint posterior sampling separates baseline receptor kinetics from transient smoking-induced dopamine release in brain PET. That setting had relied on NLS+BIC to choose between MRTM and lp-ntPET, and the BIC complexity penalty suppresses sensitivity to activation curves under voxel noise. On 2,000 simulated test curves the consistency model reached 91.4% accuracy versus 73.5% for NLS+BIC; it identified 94.4% of activated lp-ntPET curves and 88.4% of baseline MRTM curves, while NLS+BIC identified only 48.1% of lp-ntPET curves.
Perspective
The demonstrated scope is voxel-wise selection among seven FDG compartment models in total-body PET, and selection between MRTM and lp-ntPET in [11C]raclopride brain PET; the clinical part uses scans from one 54-year-old man with a pancreatic neuroendocrine tumor and hepatic metastasis and from one smoker. The model is trained on simulated curves generated with this cohort's frame schedule and arterial input functions, so a different scanner protocol or tracer requires retraining or fine-tuning.
Parameter estimates currently use only the single model with the highest posterior probability, yet voxels in regions such as liver and skeletal muscle assign non-zero probability to several models; the authors propose Bayesian model averaging for this mixed kinetics but have not yet applied it in clinical studies. Clinical validation is limited to scans from one tumor patient and one smoker, and the simulation results depend on training data generated from this cohort's input functions and frame schedule.