New AI Model Turns Short Amyloid PET Scans into Full Late-Phase Images

Researchers have developed a generative artificial intelligence framework called MOVIE — Modeling ODEs with Visually Interpretable Evolution — that synthesizes late-phase [11C]-Pittsburgh compound B (PiB) PET images from a short one-minute early-phase acquisition. The work, published in EJNMMI Physics, aims to address a practical problem in amyloid imaging: dynamic PiB PET provides valuable temporal information for studying amyloid-related diseases, but the long acquisition causes patient discomfort and adds operational costs.

The motivation is straightforward. If clinically useful late-phase images can be reliably generated from data collected in the first minute of a scan, multiple biomarkers could potentially be obtained in a single, much shorter imaging session. The team describes MOVIE as the first framework, to their knowledge, to couple early-to-late PET synthesis with a temporally continuous, physiology-guided generative model, moving beyond conventional black-box regression.

Technically, MOVIE incorporates neural ordinary or stochastic differential equations (ODEs/SDEs) to model the evolution of image appearance over time. The framework also uses mid-phase frames to guide the prediction of late-phase images. A distinctive feature is the step-by-step visualization of the synthesis trajectory, allowing observers to watch how the model transforms early-phase uptake patterns into late-phase ones rather than receiving only a final output from an opaque network.

The study used 78 dynamic PiB PET scans from a mixed cohort of patients with cerebral amyloid angiopathy, Alzheimer’s disease, and hypertension to derive early- and late-phase images. This mixed clinical population is an important detail, since the training data reflect a range of amyloid-related and vascular conditions rather than a single diagnostic group.

Quantitative evaluation relied on several established image-quality and task-level metrics: peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), LPIPS, effect size, and SUVR-based regional analyses. Classification performance for amyloid positivity was assessed using ROC AUC and decision-curve analysis (DCA). The step-by-step synthesis from early to late phase was also visualized as part of the evaluation.

The results showed consistent quantitative gains. The best-performing configuration, the SDE-based model, reached a highest PSNR of 41.46 ± 2.96 and an SSIM of 0.969 ± 0.0093, with an effect size greater than 0.8, and it outperformed a UNet baseline model. These figures indicate that the generated images closely matched the actual late-phase acquisitions in the test data.

On the clinical task of classifying amyloid positivity, the covariate-adjusted AUC in the global Centiloid composite ROI was higher for the MOVIE SDE implementation (AUC = 0.881) than for the baseline model, and the framework also delivered a net benefit above the baseline under decision-curve analysis. Visualizations of the generative trajectories showed biologically plausible early-to-late evolution of uptake spatial patterns, which the authors present as evidence of the model’s interpretability.

The authors conclude that MOVIE shows potential to reduce scan time, alleviate hospital operations, improve interpretability, and support clinical decision-making in amyloid PET imaging. As with any newly proposed synthesis framework, these findings come from a single study in a mixed patient cohort, and the reported metrics reflect performance on that dataset rather than established clinical utility. The article was released as a peer-reviewed, accepted manuscript shared early to provide faster access, subject to further edits before the final version of record.

Subject of Research: Deep learning synthesis of late-phase [11C]-PiB PET images from short early-phase acquisitions

Article Title: A PET MOVIE for interpretable synthesis of late-frame [11C]-PiB PET images from early-frame counterparts

Article References: Tsai, B.-W., Lin, Y.-N., Lin, H.-T., Li, Y.-S., Huang, G.-L., Liu, C.-J., Ko, C.-L., Yen, R.-F., Tsai, H.-H., & Chen, K. T. (2026). A PET MOVIE for interpretable synthesis of late-frame [11C]-PiB PET images from early-frame counterparts. EJNMMI Physics. https://doi.org/10.1186/s40658-026-00910-3

Image Credits: AI Generated

DOI: 10.1186/s40658-026-00910-3

Keywords: PET imaging, amyloid, PiB, deep learning, neural ODEs, image synthesis, Alzheimer's disease, cerebral amyloid angiopathy, Centiloid, EJNMMI Physics, MOVIE, interpretable