⚡ electrical engineering

Feynman Kac Reweighted Schrödinger Bridge Matching for Surface-Based Tau PET Harmonization

The paper proposes Feynman Kac Reweighted Schrödinger Bridge Matching (FKRSBM), a novel harmonization method that uses entropy-regularized optimal transport and subgroup-aware reweighting to effectively remove site-induced variability in surface-based Tau PET imaging while preserving biologically meaningful signals and improving downstream disease classification.

Original authors: Jianwei Zhang, Xinyu Nie, Jiaxin Yue, Yonggang Shi

Published 2026-06-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jianwei Zhang, Xinyu Nie, Jiaxin Yue, Yonggang Shi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to compare the height of trees in two different forests. One forest is measured by a team using a new, high-tech laser scanner, and the other is measured by a team using an old, slightly bent tape measure. Even if the trees are the same size, the numbers you get will look different just because of the tools used.

In the world of brain imaging, scientists face a similar problem. They want to study Alzheimer's disease by looking at "Tau" proteins in the brain using PET scans. However, different hospitals use different scanners, different chemical dyes (called radiotracers), and different scanning protocols. This creates "noise" that looks like biological differences but is actually just a technical glitch. If you try to mix data from these different sites, you might accidentally think a healthy brain is sick, or vice versa, simply because of the machine used.

This paper introduces a new method called FKRSBM (a mouthful, so let's call it the "Smart Bridge") to fix this. Here is how it works, explained simply:

The Problem: Mixing Apples and Oranges

Imagine you have a pile of red apples (patients with a specific type of brain protein) and green apples (patients without it) from Forest A. You want to compare them to a pile of apples from Forest B. But Forest B has a different mix: maybe 80% are red and 20% are green, while Forest A is 50/50.

Old methods of "harmonizing" (fixing the data) often just throw all the apples from Forest A into a blender and try to make them look like Forest B. The problem? The blender might accidentally turn a red apple from Forest A into a green apple just to match the average of Forest B. This destroys the real biological truth: a patient who is "Tau-positive" (red apple) should stay "Tau-positive" after the fix, not get turned into a "Tau-negative" (green apple) patient.

The Solution: Building a "Smart Bridge"

The authors use a mathematical concept called a Schrödinger Bridge. Think of this as building a bridge to transport data from one forest to the other.

The Old Way (The Blind Bridge): Previous methods built a bridge where they randomly picked an apple from Forest A and tried to match it with any apple from Forest B. If Forest B had more red apples, the bridge might force a green apple from Forest A to turn into a red one just to fit the numbers. This confuses the scanner differences with real disease differences.
The New Way (The Smart Bridge - FKRSBM): The authors added a "traffic cop" to the bridge. This traffic cop looks at the "ID card" of each apple (whether it's a red or green apple, i.e., Tau-positive or Tau-negative).
- If you have a red apple from Forest A, the traffic cop only lets you cross the bridge to a red apple in Forest B.
- If you have a green apple, you only cross to a green apple.

This ensures that the "transport" respects the biology. A sick brain stays sick; a healthy brain stays healthy. The scanner differences are smoothed out, but the disease status is preserved.

How They Did It: The "Feynman-Kac" Trick

The paper uses a complex mathematical trick called Feynman-Kac reweighting. In simple terms, instead of changing the entire bridge-building machine, they just changed the list of who gets to cross the bridge.

They created a special rule for picking pairs: "We will only pair a patient from Site A with a patient from Site B if they have the same disease status." They did this by mathematically "weighting" the pairs. Good matches (same status) get a heavy weight (they are picked often); bad matches (different status) get a tiny weight (they are almost never picked).

This is clever because they didn't have to rebuild the whole computer program. They just changed the input list, and the computer learned to build a bridge that naturally respects the biology.

The Results: A Better Map

The team tested this on real brain scans from two different groups of people (HABS-HD and ADNI) using different chemical dyes.

Better Alignment: The "Smart Bridge" made the data from the two groups look much more similar to each other than previous methods did.
No Identity Theft: Most importantly, it rarely made a "Tau-positive" patient look "Tau-negative" (or vice versa). Other methods made this mistake much more often.
Genetic Consistency: The method also preserved the link between the brain scans and a specific genetic risk factor for Alzheimer's (called APOE ε4). People with the risk gene still showed higher levels of the protein after the fix, just like in the real target data.
Better Diagnosis: When they used the cleaned-up data to train a computer to diagnose Alzheimer's, the results were more accurate than when using data cleaned by other methods.

The Bottom Line

The paper claims that by using this "Smart Bridge" that respects biological subgroups (like Tau status), they can successfully mix brain scan data from different hospitals and different machines without losing the critical information about who actually has the disease. It's like fixing the translation between two languages without accidentally changing the meaning of the story.

Technical Summary: Feynman–Kac Reweighted Schrödinger Bridge Matching for Surface-based Tau PET Harmonization

Problem Statement

Multi-site neuroimaging studies are essential for increasing statistical power but are confounded by non-biological variability arising from differences in scanners, acquisition protocols, and radiotracers. This variability inflates biomarker variance, reduces sensitivity to disease effects, and can introduce systematic bias, particularly when site distributions covary with biological variables.

While harmonization methods aim to remove site-induced shifts, existing approaches face significant limitations in unpaired settings (where no common subjects exist across sites):

Statistical methods (e.g., ComBat): Assume linear, Gaussian effects and treat features independently, failing to capture nonlinear interactions.
Deep learning methods (e.g., CycleGAN): Suffer from training instability, mode collapse, and the theoretical impossibility of unsupervised disentanglement without strong inductive biases.
Diffusion-based methods: Rely on noise injection that presents a tradeoff between domain adaptation and content preservation.

A critical gap remains in Schrödinger Bridge (SB) based harmonization. Standard SB implementations (e.g., Diffusion Schrödinger Bridge Matching, DSBM) typically sample endpoint pairs independently from the product measure $\pi_0 \otimes \pi_1$ . When source and target cohorts differ in subgroup composition (e.g., varying proportions of tau-positive and tau-negative subjects), this "agnostic" coupling can map samples from one biological subgroup to a different subgroup in the target domain simply due to prevalence. This conflates tracer differences with biological variation, defeating the purpose of harmonization. Furthermore, applying SB to surface-based neuroimaging is challenging due to the non-Euclidean mesh topology of cortical surfaces.

Methodology

The authors propose Feynman–Kac Reweighted Schrödinger Bridge Matching (FKRSBM), a framework designed to harmonize surface-based tau PET data while respecting biological subgroup structures.

1. Theoretical Foundation: Endpoint-Regularized SB

The method formulates harmonization as an entropy-regularized optimal transport problem. To prevent the conflation of site effects with biological variation, the authors introduce an endpoint penalty $\Phi(X_0, X_1)$ that encodes prior knowledge about biologically meaningful pairings (e.g., penalizing mismatches in tau-positivity status).

The objective minimizes the Kullback-Leibler (KL) divergence between the learned path measure $P$ and a reference measure $R$ , augmented by a penalty term:
$P^*_\lambda \in \arg\min_{P \ll R, P_0=\pi_0} \left\{ KL(P \| R) + \lambda E_P[\Phi(X_0, X_1)] \right\}$
where $\lambda$ controls the penalty strength.

2. Feynman–Kac Reweighting

A key theoretical contribution is showing that this regularized objective is equivalent to standard bridge matching under a Feynman–Kac reweighted reference measure $R_\lambda$ .

Mechanism: The reweighting modifies the endpoint coupling (the distribution of source-target pairs) but leaves the conditional Brownian bridge dynamics between any fixed pair unchanged.
Implementation: This allows the constraint to be implemented entirely at the data-sampling level via stratified importance sampling. The model does not require changes to the underlying network architecture or the bridge-matching solver.
Sampling: During training, target endpoints are sampled from a conditional proposal $q_\lambda(x_1|x_0) \propto \exp(-\lambda \Phi(x_0, x_1)) \pi_1(x_1)$ , which preferentially pairs subjects with matching biological labels (e.g., tau-positive with tau-positive).

3. Surface-Based Architecture

For cortical surface data, FKRSBM employs a spherical convolutional backbone (Conditional Surface U-Net) operating on icosahedral meshes. This architecture respects the non-Euclidean geometry of the cortex, preserving surface topology during vertex-level harmonization.

4. Training and Inference

Training: The model uses a two-stage process (pretraining and finetuning) involving forward and backward drift regression against analytic Brownian bridge velocities. The endpoint pairs are drawn from the Feynman–Kac reweighted proposal.
Inference: Harmonization is performed by integrating the learned stochastic differential equation (SDE) from $t=0$ to $t=1$ using the Euler–Maruyama method, starting from a source sample.

Key Contributions

Schrödinger Bridge on Cortical Surfaces: The paper introduces the first SB-based harmonization framework specifically designed for cortical surface meshes, utilizing spherical convolutions to handle non-Euclidean topology.
Endpoint-Constrained Transport: The authors formulate an endpoint-regularized SB objective that prevents the transport from conflating site effects with biological variation. They prove that this can be implemented via Feynman–Kac reweighting of the endpoint proposal, requiring no architectural changes to the solver.
Modular Design: The reweighting operates entirely at the data-sampling level, making the constraint modular and compatible with future improvements in SB solvers.

Results

The method was evaluated on unpaired tau PET SUVR maps, harmonizing PI-2620 data from the HABS-HD cohort into the AV-1451 domain of ADNI. It was compared against ComBat, CycleGAN, a diffusion-based method (DF), and unregularized DSBM.

Distributional Alignment: FKRSBM achieved the lowest Wasserstein Distance (WD) and the strongest residual site separability reduction (lowest Somers' D) in both hemispheres, outperforming all baselines.
Subgroup Consistency: Unlike unregularized DSBM, which showed systematic under/over-correction in stratified subgroups, FKRSBM maintained alignment in both tau-positive and tau-negative subgroups.
Tau-Positivity Preservation: FKRSBM demonstrated the lowest rate of tau-status flips (2.6% left, 3.0% right hemisphere) compared to DSBM (17.5% left, 13.3% right). The unregularized DSBM exhibited a strong bias toward flipping tau-negative subjects to positive, a failure mode mitigated by the reweighting.
Biological Stratification: FKRSBM better preserved the separation between APOE ε4 carriers and non-carriers, showing lower Kolmogorov–Smirnov (KS) statistics against the target distribution within these genetic subgroups.
Downstream Performance: In diagnostic classification tasks (CN vs. AD, and CN vs. MCI vs. AD), FKRSBM yielded the highest F1 scores and accuracy, indicating that the harmonized data retained disease-relevant signal while reducing site noise.

Significance and Claims

The paper claims that FKRSBM addresses a fundamental flaw in standard unpaired domain translation: the risk of conflating technical site effects with biological heterogeneity when cohort compositions differ. By enforcing biologically consistent transport through Feynman–Kac reweighting, the method achieves superior distributional alignment and preserves critical clinical biomarkers (tau positivity and APOE status) better than existing statistical and deep learning approaches.

The authors conclude that endpoint-aware coupling is crucial for harmonization in scenarios where source and target cohorts differ in subgroup composition. They position FKRSBM as a principled solution that improves both theoretical alignment and practical performance in cross-tracer tau PET experiments, while noting that future work should extend the method to multi-site harmonization, continuous covariates, and other surface-based imaging modalities.