Open-Source Pipeline Brings Reproducible Tau PET Quantification to Alzheimer’s Research

Tau proteins clumping inside brain cells are among the most telling molecular signatures of neurodegenerative disease, and positron emission tomography (PET) tracers designed to light up these aggregates have transformed how researchers visualize Alzheimer’s disease and related disorders. Yet turning raw PET scans into reliable numbers has long been a stubborn bottleneck. A team of Japanese researchers has now unveiled TAME-Q, an open-source preprocessing pipeline that automates the semi-quantification of florzolotau (18F) PET images, and their results, published in BMC Medical Imaging, suggest that reproducible tau imaging may finally be within reach for laboratories everywhere.

The challenge that TAME-Q addresses is deceptively simple to state but notoriously difficult to solve. To measure how much tau is deposited in a given brain region, researchers typically compare tracer uptake in that region against a reference region assumed to be free of tau pathology. But in diseases such as Alzheimer’s and progressive supranuclear palsy, tau spreads so widely that finding a truly unaffected reference region becomes nearly impossible. Choosing the wrong reference can distort measurements and undermine comparisons across patients and studies, a problem that has hampered efforts to standardize tau imaging across the growing number of clinical trials targeting tau pathology.

To sidestep this dilemma, the group had previously developed a histogram-based grey-matter reference strategy. Instead of prespecifying a single anatomical region, the method examines the intensity distribution of grey-matter voxels across the whole brain and uses statistical fitting of that histogram to estimate a reference value. In patients with extensive tau burden, the distribution of grey-matter uptake becomes bimodal, with one peak representing affected tissue and another representing relatively spared tissue, and Gaussian fitting can separate these populations. In healthier brains the distribution is monomodal, and the approach adapts accordingly. The strategy worked well in practice, but its implementation depended on in-house code and proprietary software, which meant that other laboratories could not easily reproduce the results or integrate the method into their own workflows.

TAME-Q, which stands for Tomographic Image Preprocessing Using Automated Multi-Reference Estimation for Quantification, was built to break down that barrier. The pipeline takes dynamic florzolotau PET images and T1-weighted magnetic resonance images as input and carries out the entire processing chain automatically: image alignment, tissue segmentation, Gaussian fitting of grey-matter intensity histograms, reference estimation, generation of standardised uptake value ratios (SUVRs), FreeSurfer-based parcellation of the brain into defined regions, extraction of SUVRs for 132 regions of interest, and the production of quality-assurance outputs that allow users to inspect what the pipeline has done at every step. Because the code is openly accessible, any laboratory with the requisite imaging data can run the same analysis and expect comparable results.

The researchers evaluated TAME-Q on a cohort of 133 participants scanned at a single center: 37 individuals on the Alzheimer’s disease continuum, 46 patients with progressive supranuclear palsy–Richardson syndrome (PSP-RS), and 50 healthy controls. Every single participant’s scan completed automated preprocessing without failure, and all outputs passed both quantitative screening and structured visual review. Of the 133 histograms, 121 showed the bimodal pattern typical of tau-bearing brains and were fitted accordingly, while 12 showed monomodal distributions. That flawless completion rate matters in a field where preprocessing failures and manual interventions often introduce silent variability between studies.

The biological signals that emerged were exactly where theory predicted. After adjusting for age and sex, SUVRs were significantly elevated in the inferior temporal gyrus among participants on the Alzheimer’s continuum, a region known to accumulate tau early in that disease, and in the globus pallidus among patients with PSP-RS, consistent with the basal ganglia involvement characteristic of that disorder. Both findings reached statistical significance at Holm-adjusted p values below 0.001. The concordance between the pipeline’s disease-specific patterns and established neuropathological knowledge provides strong evidence that the automated processing preserves the biological signal it is meant to measure.

Because the histogram-based reference and a conventional cerebellar reference produced broadly similar group-level scaling, the team also examined where the two approaches diverged, finding differences in some individual participants even though group averages aligned. This individual-level divergence underscores why the choice of reference strategy matters for applications such as single-subject diagnosis or longitudinal monitoring, where group-level agreement offers little comfort. For clinical decision-making or trial enrollment based on an individual scan, a reference method that adapts to each brain’s actual intensity distribution may capture information that a fixed anatomical reference misses.

To test whether TAME-Q could support the kind of machine-learning classification that tau imaging increasingly demands, the researchers derived composite tau scores using Elastic Net regularization, producing separate AD-tau and PSP-tau scores from the regional SUVR data. When compared with scores from the original benchmark implementation, the correlations were strikingly strong: r = 0.979 for the AD-tau score and r = 0.960 for the PSP-tau score. Discrimination performance was equally impressive. Across the full cohort, the area under the receiver operating characteristic curve reached 0.999 for distinguishing Alzheimer’s participants and 0.967 for distinguishing PSP-RS patients. In held-out testing across ten iterations of a 75/25 percent shuffle-split, mean AUCs remained at 0.996 plus or minus 0.009 for Alzheimer’s disease and 0.935 plus or minus 0.010 for PSP-RS, indicating that the pipeline’s outputs generalize to data the model has never seen.

The team went further, stress-testing the robustness of their classification framework. When they truncated the Elastic Net coefficients to retain only the top 20 regions of interest, full-cohort AUCs remained essentially unchanged, showing that the discriminative power does not depend on an unwieldy number of features. Retraining with an unmerged version of the FreeSurfer atlas likewise produced similar results, suggesting that the approach is not hypersensitive to the precise anatomical parcellation used. These sensitivity analyses matter for any laboratory hoping to adopt the pipeline with slightly different tools or atlas versions, since they indicate a method resilient to reasonable variations in implementation.

The authors are careful to note the limits of what they have shown. Their validation comes from a single-center dataset, and external multicenter validation will be required to establish that TAME-Q generalizes across different scanner platforms, acquisition protocols, and patient populations. PET quantification is notoriously sensitive to variations in tracer production, injection protocols, and reconstruction algorithms, so a pipeline proven at one site must earn its stripes at many more before it can anchor multi-site trials. Still, the significance of what has been released is hard to overstate. By packaging a sophisticated, previously proprietary quantification strategy into an automated, inspectable, and openly available tool, TAME-Q lowers the entry barrier for tau imaging research worldwide. As florzolotau and other tau tracers move from research settings toward clinical use in diagnosing and tracking Alzheimer’s disease and tauopathies such as progressive supranuclear palsy, tools that guarantee that a SUVR means the same thing in Tokyo, Boston, or London will be indispensable. In a field racing toward anti-tau therapies, the ability to measure the target reliably and reproducibly may prove as consequential as the drugs themselves.

Subject of Research: An open-source preprocessing pipeline for reproducible semi-quantification of florzolotau (18F) tau PET in Alzheimer's disease and progressive supranuclear palsy

Article Title: TAME-Q: an open-source preprocessing pipeline for reproducible semi-quantification of florzolotau (18 F) PET

Article References: Nakayama, K., Nemoto, K., Endo, H., Tagai, K., Oyama, A., Nemoto, M., Tamura, M., Ota, M., Tokuda, T., Higuchi, M., & Arai, T. (2026). TAME-Q: an open-source preprocessing pipeline for reproducible semi-quantification of florzolotau (18 F) PET. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02775-5

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02775-5

Keywords: TAME-Q, florzolotau PET, tau imaging, Alzheimer's disease, progressive supranuclear palsy, SUVR, open-source software, neuroimaging, PET quantification, FreeSurfer, machine learning, BMC Medical Imaging