Phosphorylated tau (pTau) is a defining pathology in Alzheimer’s disease. Establishing
how pTau is reflected in magnetic resonance imaging (MRI) is important because MRI is
scalable, repeatable, and noninvasive, making it a practical modality for tracking pathology
burden over space and time. However, it remains unclear whether an MRI signature of
pTau exists, and whether such a signature could enable accurate voxelwise detection of tau
concetration. We investigate whether tau-related tissue injury has a unique signature based
on tissue microenvironments in diffusion–relaxation space, and whether this signature can
be validated against histological ground truth.

Using ex vivo human cortical tissue (14 postmortem slices; internal cohort, n = 8; independent external cohort, n = 6), we reconstructed voxelwise T1–MD and T2–MD joint distributions and co-registered them with histology-derived pTau measurements. We compared three representation strategies, divergence descriptors, direct principal component analysis (PCA), and transport-based morphometry followed by PCA (TBM-PCA), within nested cross-validation with Bayesian hyperparameter optimization.

We report that pTau-related structure is present in MRI, as captured by divergence analysis
in native space. TBM-PCA yields the strongest overall accuracy. In the internal cohort, random
forest with TBM-PCA achieved an R2 of 0.883 for T1–MD regression, binary accuracy
of 92.6 % and three-class accuracy of 89.3 % on unseen data. In the external cohort, classification remained comparatively stable, and predicted maps retained spatial localization of pTau-enriched regions.

Taken together, we present a proof-of-concept that the signature of pTau can be accurately
detected on multidimensional MRI. These data indicate that multidimensional MRI
distributions carry histology-linked pTau signal, and that this signal is recovered more effectively when geometric structure is modeled rather than reduced to variance-only embeddings. Looking ahead, this framework could help enable scalable, radiation-free MRI markers of tau pathology for longitudinal monitoring, biologically grounded staging, and treatment-response tracking as disease-modifying therapies expand.

Phosphorylated tau (pTau) is a defining pathology in Alzheimer’s disease. Establishing
how pTau is reflected in magnetic resonance imaging (MRI) is important because MRI is
scalable, repeatable, and noninvasive, making it a practical modality for tracking pathology
burden over space and time. However, it remains unclear whether an MRI signature of
pTau exists, and whether such a signature could enable accurate voxelwise detection of tau
concetration. We investigate whether tau-related tissue injury has a unique signature based
on tissue microenvironments in diffusion–relaxation space, and whether this signature can
be validated against histological ground truth.

Using ex vivo human cortical tissue (14 postmortem slices; internal cohort, n = 8; independent external cohort, n = 6), we reconstructed voxelwise T1–MD and T2–MD joint distributions and co-registered them with histology-derived pTau measurements. We compared three representation strategies, divergence descriptors, direct principal component analysis (PCA), and transport-based morphometry followed by PCA (TBM-PCA), within nested cross-validation with Bayesian hyperparameter optimization.

We report that pTau-related structure is present in MRI, as captured by divergence analysis
in native space. TBM-PCA yields the strongest overall accuracy. In the internal cohort, random
forest with TBM-PCA achieved an R2 of 0.883 for T1–MD regression, binary accuracy
of 92.6 % and three-class accuracy of 89.3 % on unseen data. In the external cohort, classification remained comparatively stable, and predicted maps retained spatial localization of pTau-enriched regions.

Taken together, we present a proof-of-concept that the signature of pTau can be accurately
detected on multidimensional MRI. These data indicate that multidimensional MRI
distributions carry histology-linked pTau signal, and that this signal is recovered more effectively when geometric structure is modeled rather than reduced to variance-only embeddings. Looking ahead, this framework could help enable scalable, radiation-free MRI markers of tau pathology for longitudinal monitoring, biologically grounded staging, and treatment-response tracking as disease-modifying therapies expand.