Feynman Kac Reweighted Schr\"odinger Bridge Matching for Surface-Based Tau PET Harmonization

A new computational method aims to reduce scanner- and protocol-driven variability in tau PET brain scans, a key tool for tracking Alzheimer's disease, according to a paper submitted on 16 Jun 2026 [1]. The technique, called Feynman Kac Reweighted Schrödinger Bridge Matching (FKRSBM), learns a direct transport process between imaging distributions to harmonize data across different clinical sites [1]. Tau PET imaging is central to monitoring Alzheimer's disease progression, but systematic differences between scanners, protocols, and radiotracers introduce nonbiological variability that can inflate biomarker variance and bias clinical assessments [1]. Existing harmonization methods such as ComBat and CycleGAN often struggle when the source and target patient cohorts differ in subgroup composition, risking the conflation of site effects with biological variation like tau-positivity status [1]. The FKRSBM model addresses this by learning a direct stochastic transport process between source and target distributions via entropy-regularized optimal transport, rather than routing data through a Gaussian noise prior as in diffusion-based approaches [1]. To enforce biologically consistent transport, the method incorporates a subgroup-aware endpoint proposal derived from a Feynman Kac reweighting of the reference bridge measure, implemented through stratified importance sampling at the data level without requiring changes to the underlying bridge-matching solver or network architecture [1]. For surface-based neuroimaging, FKRSBM uses a spherical convolutional backbone operating on cortical meshes to perform vertex-level harmonization [1]. The researchers evaluated the method on tau PET SUVR maps, harmonizing PI-2620 data from the HABS-HD cohort into the AV-1451 domain of the ADNI dataset [1]. Compared against ComBat, CycleGAN, a diffusion-based method, and unregularized Diffusion Schrödinger Bridge Matching, FKRSBM achieved superior distributional alignment, reduced tau-positivity sign mismatch, stronger APOE subgroup alignment, and improved downstream disease classification performance [1]. The work builds on broader trends in transfer learning across scientific domains, where models trained on one dataset are adapted to improve performance on smaller or differently distributed datasets, a strategy that has shown utility in fields such as catalysis informatics [3].