Computational decision support systems could provide clinical value in whole-body FDG PET/CT workflows, but labeled data is scarce and PET/CT imaging exams are cumbersome. Here, the authors describe a weak supervision framework that extracts regional abnormality labels from free-text radiology reports.