MRI-Based Centiloid Methods Prove Most Reliable for Amyloid PET Scoring

Quantifying amyloid plaques in the brain depends on a standardised unit called the Centiloid scale, yet the software used to compute those values has varied widely between hospitals and research centres. A study published in EJNMMI Physics has now compared leading processing methods head to head and found that the choice of platform can materially change whether a patient is classified as amyloid positive or negative. The work, led by Xiaoling Cao, Zhiqiang Tan and colleagues with senior author Hao Xu, systematically evaluated three software platforms—SPM, PMOD and BTXBrain—using two amyloid tracers, 18F-florbetaben and 18F-florbetapir, and validated the results against visual assessments by readers.

The Centiloid scale was designed to make amyloid PET measurements comparable across tracers, scanners and analysis pipelines, but variability in processing methods has complicated its adoption in routine clinical practice.

To test how much that matters, the team drew on datasets from the Global Alzheimer’s Association Interactive Network, a public repository of amyloid PET images, together with 127 subjects from a local database—64 scanned with 18F-florbetaben and 63 with 18F-florbetapir. Centiloid values were derived on each of the three platforms using their distinct processing methods. The first stage of validation compared Centiloid values computed from the public database images against published reference values using linear regression, a standard way of checking whether a new pipeline reproduces established numbers.

The results split the methods into two camps. MRI-based approaches, which use a patient’s structural magnetic resonance image to guide the spatial normalisation of the PET data, consistently met the established performance criteria across platforms. Their regression slopes fell between 0.98 and 1.02, intercepts stayed within minus 2 to plus 2 Centiloid units, and coefficients of determination exceeded 0.98—in other words, they reproduced the published reference values almost exactly. When the researchers applied these pipelines to the local datasets and classified amyloid status using cut-off values of 12 or 24 Centiloid units, diagnostic performance was strong, with sensitivity and negative predictive value both reaching 100 percent and accuracy of at least 94 percent, taking visual assessment as the reference standard.

PET-only methods, which must process images without the anatomical guidance of an MRI, told a different story. These pipelines proved highly tracer-specific, meaning a method that performed well with one radioligand did not necessarily transfer to the other. For 18F-florbetaben, SPM and BTXBrain delivered robust and comparable outcomes, with all key metrics exceeding 90 percent.

With 18F-florbetapir, the gaps between platforms widened considerably. BTXBrain showed the strongest diagnostic performance among the PET-only pipelines, achieving accuracy between 82.5 and 85.7 percent, sensitivity between 77.3 and 84.1 percent, and negative predictive value between 64.3 and 70.8 percent. SPM landed in an intermediate position, with accuracy of 73.0 to 81.0 percent, sensitivity of 63.6 to 77.3 percent and negative predictive value of 52.9 to 63.0 percent. PMOD, despite achieving perfect specificity and positive predictive value of 100 percent, showed the lowest accuracy at 60.3 to 66.7 percent, sensitivity of 43.2 to 52.3 percent and negative predictive value of 43.2 to 47.5 percent. The pattern indicates that PMOD rarely mislabelled amyloid-negative scans as positive, but it missed a substantial share of genuinely positive cases—a serious limitation when the clinical question is whether a patient carries amyloid pathology at all.

The diagnostic metrics were calculated with visual assessment results as the reference standard, and the researchers computed sensitivity, specificity, positive predictive value, negative predictive value and accuracy for each method at both the 12 and 24 Centiloid cut-offs. The ranges reported for each platform capture how performance shifted between those two thresholds.

Centres relying on PET-only processing need to match their software choice to their tracer. The study’s conclusion singles out BTXBrain as having a relative advantage in PET-only processing, particularly with 18F-florbetapir, while cautioning that further multicentre validation is warranted before the findings are generalised broadly.

Limitations remain. The local validation cohort comprised 127 subjects, and the authors themselves flag the need for further multicentre validation, particularly for the PET-only pipelines whose performance varied with tracer. The published reference values against which the public-database images were compared anchor the regression analysis, but centres adopting these methods should still verify performance on their own scanner and population.

Subject of Research: Validation of Centiloid scale processing methods for quantifying amyloid beta PET in Alzheimer's disease

Article Title: Comparison and clinical validation of different centiloid processing methods in amyloid beta PET

Article References: Cao, X., Tan, Z., Zhang, W., Zhang, P., Tang, Y., Wu, B., Cai, Q., Shang, J., Gong, J., Wang, L., Ling, X., & Xu, H. (2026). Comparison and clinical validation of different centiloid processing methods in amyloid beta PET. EJNMMI Physics. https://doi.org/10.1186/s40658-026-00951-8

Image Credits: AI Generated

DOI: 10.1186/s40658-026-00951-8

Keywords: amyloid PET, Centiloid scale, Alzheimer's disease, 18F-florbetaben, 18F-florbetapir, SPM, PMOD, BTXBrain, diagnostic accuracy, neuroimaging, quantitative PET, EJNMMI Physics