AI-Augmented Thyroid Scintigraphy for Robust Classification of Disease

Researchers demonstrate that Flow Matching generative models outperform Stable Diffusion and conventional augmentation techniques for classifying thyroid scintigraphy images, achieving F1-scores of 0.78 and AUC of 0.95. The study validates that advanced AI-generated synthetic medical images can effectively address dataset limitations in diagnostic imaging tasks.

This research addresses a critical challenge in medical AI: training robust deep learning models when clinical datasets are small and imbalanced. Thyroid imaging classification typically relies on limited patient data across multiple institutions, making traditional machine learning approaches unreliable. The study's comparison of three augmentation strategies—conventional techniques, Stable Diffusion variants, and Flow Matching models—provides actionable insights for medical imaging applications beyond thyroidology.

The superior performance of Flow Matching represents a meaningful advancement in generative AI for healthcare. While Stable Diffusion has dominated recent AI discourse, this work demonstrates that domain-specific generative models can produce higher-fidelity synthetic data more suitable for clinical applications. The finding that physician-generated prompts improve Stable Diffusion performance (macro F1 of 0.76) underscores that clinical context enhances synthetic data generation quality. Flow Matching's lowest image fidelity scores (FID: 0.66, KID: 0.83) indicate the synthetic images more closely resemble real medical scans, a crucial factor for training classifiers that generalize to actual patient data.

For healthcare organizations and medical AI developers, this research validates synthetic data generation as a legitimate strategy for improving diagnostic systems. The open-source code release democratizes access to these techniques, enabling smaller institutions and research groups to implement advanced augmentation. Insurance companies and healthcare systems may benefit from more reliable diagnostic tools. The broader implication extends beyond thyroid imaging: generative models tailored to specific medical imaging modalities could accelerate development of diagnostic AI across radiology, pathology, and other data-constrained specialties, potentially reducing costs and improving care quality.

• →Flow Matching generative models achieved superior thyroid classification performance (F1: 0.78, AUC: 0.95) compared to Stable Diffusion and conventional augmentation.
• →Synthetic medical images generated by advanced models can effectively augment limited clinical datasets without compromising model performance.
• →Physician-provided clinical context through prompts significantly improves Stable Diffusion's synthetic image quality and diagnostic utility.
• →Image fidelity metrics (FID, KID) are reliable indicators of generative model quality for medical imaging applications.
• →Open-source implementation enables broader adoption of advanced augmentation techniques across healthcare organizations with limited data.