Hamiltonian dynamics for stochastic reconstruction in emission tomography

TL;DR

This paper introduces a Hamiltonian Monte Carlo-based stochastic reconstruction method for emission tomography, enabling ensemble generation, uncertainty quantification, and insight into the inverse problem's conditioning in high-dimensional voxel space.

Contribution

It presents a novel stochastic reformulation using HMC sampling in high-dimensional space, providing practical ensemble generation and uncertainty analysis in emission tomography.

Findings

Ensemble analysis distinguishes intrinsic uncertainty from model inadequacy.

Method achieves point-estimate accuracy comparable to deterministic methods.

Provides physically interpretable uncertainty metrics in clinical imaging.

Abstract

The AMIAS/RISE framework formulates emission tomography as a probabilistic inverse problem in which reconstructed images are sampled from a distribution defined by the measurement model and counting statistics. In this work we present a stochastic reformulation of this approach based on gradient-driven optimization combined with Hamiltonian Monte Carlo (HMC) sampling directly in high-dimensional voxel space. This formulation enables practical ensemble generation for tomographic image reconstructions and provides direct access to image fluctuations within the sampled ensemble. Beyond point reconstruction, we introduce a spatially resolved operator-weighted diagnostic-the sampled data-visible variance-which quantifies how image fluctuations propagate through the imaging operator and thereby probes the local conditioning of the inverse problem under realistic acquisition physics. Using controlled software phantoms, experimental anthropomorphic phantom measurements, and a clinical DATSCAN SPECT acquisition, we demonstrate that while point-estimate accuracy under ideal conditions is comparable to deterministic reconstruction methods, the stochastic formulation provides additional physically interpretable insight. In particular, the ensemble analysis distinguishes uncertainty arising from intrinsic ill-posedness of the inverse problem from that associated with forward-model inadequacy. The clinical example is included to illustrate methodological applicability under realistic acquisition statistics rather than to assess diagnostic performance. The results establish the stochastic reconstruction framework as a practical ensemble-based approach for uncertainty quantification and forward-model validation in emission tomography.