A radiologist stares at gray-white smudges on the screen — dozens of body slices where every suspicious spot must be found. Manual segmentation of one patient takes half an hour, and a large clinic sees hundreds of such patients. Fatigue and varying experience inevitably introduce variability: two specialists will outline the same lesion slightly differently.
This routine is now handled by the new open-source model GLOW-FDG. It was trained on 1563 fused PET and CT scans (one shows where cells avidly take up glucose, the other shows anatomy) across different cancer types: breast, lung, head and neck, and melanoma. On 185 external scans from other clinics, the model found lesions more clearly than publicly available competing programs and produced fewer false alarms. Its estimate of the overall "glucose hunger" of tumors, a marker of aggressiveness, closely matched the assessments of experienced oncologists.
Such an assistant won't replace the doctor but will take over the most monotonous part of the work. Automatic segmentation paves the way for rapid extraction of quantitative biomarkers (measurable signs of disease) directly from images — without unnecessary delays and subjectivity.
Tumors often "light up" on PET because cancer cells consume glucose several times more actively than healthy cells — a phenomenon discovered by Otto Warburg back in the 1920s.