MlPET: A Localized Neural Network Approach for Probabilistic Post-Reconstruction PET Image Analysis Using Informed Priors

TL;DR

The paper addresses the noise-resolution trade-off in PET by introducing MlPET, a fast localized probabilistic post-reconstruction framework that uses a neural network to estimate the posterior mean voxel activity from small neighborhoods while integrating scanner PSF, spatially correlated noise, and informed priors. By replacing computationally intensive posterior sampling (extended Metropolis) with a localized neural estimator, MlPET achieves near-equivalence to sampling-based posterior statistics with a substantial speed-up, enabling full-volume voxelwise probabilistic PET analysis. The method is validated on NEMA phantom data across three scanners, showing higher contrast recovery near unity, reduced PSF blur by about $2.5\times$, and the ability to match 900 s conventional PET quality with only 40–80 s acquisitions; a clinical breast cancer example further illustrates potential gains in lesion detectability and noise suppression. The results suggest MlPET can provide uncertainty-quantified post-reconstruction PET analyses without altering reconstruction algorithms, with promise for reduced scan times and improved quantitative reliability in clinical imaging, pending broader patient validation and contextual priors development.

Abstract

We develop and evaluate MlPET, a fast localized machine learning approach for probabilistic PET image analysis addressing the noise-resolution trade-off in conventional reconstructions. MlPET replaces computationally demanding Markov chain Monte Carlo sampling with a localized neural network trained to estimate posterior mean voxel activity from small image neighborhoods. The method incorporates scanner-specific point spread functions, spatially correlated noise modeling, and flexible priors. Performance was evaluated on NEMA IEC phantom data from three PET systems (GE Discovery MI, Siemens Biograph Vision 600, and Quadra) under varying reconstruction settings and acquisition times. On phantom data, MlPET achieved contrast recovery coefficients consistently higher than standard PET and close to 1.0 (including 10 mm spheres), while reducing background noise and improving spatial definition. Effective pointspread function full width at half maximum decreased from approximately 2 mm in standard PET to below 1 mm with MlPET, a 2.5 fold reduction in blur. Comparable image quality was obtained at 40-80 s acquisition time with MlPET versus 900 s with conventional PET. MlPET provides an efficient approach for quantitative probabilistic post-reconstruction PET analysis. By combining informed priors with neural network speed, it achieves noise suppression and resolution enhancement without altering reconstruction algorithms. The method shows promise for improved small-lesion detectability and quantitative reliability in clinical PET imaging. Future studies will evaluate performance on patient data.

Figures (14)

• Figure 1: Schematic of a localized neighborhood $\mathbf{\Phi}_l$ (red voxels) centered on the voxel of interest $\phi_{ic}$ (black). Gray voxels indicate the full model domain $\mathbf{\Phi}$.
• Figure 2: Example of a fully connected NN architecture used to estimate the posterior mean and standard deviation of the marginal posterior distribution of the central voxel $\phi_{ic}$ in a neighborhood $\mathbf{\Phi}_l$.
• Figure 3: Estimated power-law relation between $\phi_{\text{std}}$ and $\phi_{\text{mean}}$ for GE Discovery and Vision 600 (left) and Vision Quadra Edge (right). Dots show measured pairs; lines show fitted power-law models (Eq. \ref{['eq:powerlaw']}).
• Figure 4: Example realizations from the NEMA phantom prior model $\rho_p(\mathbf{\Phi})$. The black square marks the analyzed neighborhood, and the red square marks the central voxel.
• Figure 5: Example realizations from the clinical prior model $\rho_c(\mathbf{\Phi})$ representing expected activity distributions in breast cancer patients. (a) Realizations from $\rho_c(\mathbf{\Phi})$. (b) Full marginal prior distribution of the central voxel (black square in a). (c) Marginal distribution for activity values above 10 kBq/ml for a patient with a liver mean of 10 kBq/ml and standard deviation of 2 kBq/ml.