Search papers, labs, and topics across Lattice.
This paper introduces FlowPET, a physics-informed framework for low-count Positron Emission Tomography (PET) reconstruction that addresses the issue of signal wash-out in weak lesions caused by conventional generative models. By reformulating reconstruction as volume-preserving transport in symplectic phase space and employing a Separable Hamiltonian System, FlowPET ensures that weak signals are preserved against probability mass collapse. Extensive experiments demonstrate that FlowPET outperforms existing methods in both image quality metrics and the recovery of low-contrast lesions, highlighting the effectiveness of Hamiltonian constraints in medical imaging under high noise conditions.
FlowPET not only recovers low-contrast lesions better than current methods but does so by leveraging Hamiltonian dynamics to preserve weak signals in high-noise environments.
Low-count Positron Emission Tomography (PET) reconstruction is severely hindered by the dissipative nature of prevailing generative models, where the inherent phase-space contraction leads to the numerical extinction (``wash-out'') of weak but diagnostically critical lesion signals. To overcome this geometric limitation, we propose \textbf{FlowPET}, a physics-informed framework that reformulates reconstruction as volume-preserving transport in a symplectic phase space. By parameterizing the posterior dynamics via a Separable Hamiltonian System, our approach guarantees a divergence-free vector field by construction, theoretically immunizing weak signals against probability mass collapse. To steer this conservative flow, we introduce conjugate boundary conditions based on the Range-Null space decomposition of the PET operator; this strictly enforces data consistency in the range space while confining stochastic uncertainty injection to the unobserved null space. We train the model via symplectic flow matching and perform inference using a symplectic leapfrog integrator. Extensive experiments on BrainWeb, clinical pediatric, and UDPET datasets demonstrate that \textbf{FlowPET} not only surpasses state-of-the-art deterministic and stochastic baselines in SSIM and PSNR but, more crucially, exhibits superior recovery of low-contrast lesions. The results confirm that imposing Hamiltonian structural constraints offers a robust geometric safeguard for medical inverse problems in high-noise regimes.
