Researchers have developed and validated an automated deep learning framework capable of identifying positron emission tomography radiotracers directly from reconstructed brain images, achieving robust classification across extensive multi-site neuroimaging cohorts without relying on digital metadata.
The automated classification pipeline addresses a persistent vulnerability in neuroimaging research, where manually recorded metadata records frequently introduce clerical errors, omissions, or ambiguities that compromise downstream computational analyses.
The unified development dataset encompassed ten individual radiopharmaceuticals spanning diverse biological processes: tau-targeting ligands AV-1451, MK-6240, and PI-2620; amyloid-targeting compounds Florbetapir, Florbetaben, NAV-4694, and Pittsburgh Compound B; two early-frame perfusion adaptations of Florbetapir and Florbetaben; and the metabolic fluorodeoxyglucose tracer.
Model development centered on the ConvNeXt-Large architecture, initialized with pre-trained ImageNet parameters to exploit transfer learning representations for spatial pattern recognition.
Automated PET tracer classification in multi-site, multi-cohort studies using deep learning.