Flow-matching generative AI augments thyroid scintigraphy and boosts disease classification

Flow Matching (FM) augmentation of thyroid scintigraphy data led to the highest classifier F1-scores (macro 0.77) and AUC (macro 0.93). FM-generated images were most realistic (FID 0.66), outperforming Stable Diffusion and conventional augmentation on a multi-center dataset of 2…

• Multi-center retrospective study of 2954 patients across 9 centers; 4-category classification (Diffuse Goiter, Nodular Goiter, Normal, Thyroiditis) using a ResNet18 model.
• FM-based augmentation achieved statistically superior performance over conventional methods, with the O+FM model yielding a class-wise AUC range of 0.93–0.95.
• Critically, the model was not externally or prospectively validated, limiting certainty regarding performance on unseen clinical data from centers not in the training set.

RadPigeon summaries are original and for information only. They are not clinical advice.