AI-augmented thyroid scintigraphy for robust classification of disease.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Thyroid scintigraphy is crucial for diagnosing thyroid disorders.
• Deep learning (DL) models face challenges with limited and imbalanced datasets in this field.
• Data augmentation is essential for improving DL model performance.

Purpose of the Study:

• To evaluate the impact of Stable Diffusion (SD), Flow Matching (FM), and Conventional Augmentation (CA) on DL-based thyroid scintigraphy classification.
• To compare the effectiveness of advanced generative augmentation techniques against traditional methods.
• To assess image fidelity and classification performance using various metrics.

Main Methods:

• Utilized anterior thyroid scintigraphy images from 2954 patients across nine centers.
• Classified images into four categories: Diffuse Goiter (DG), Nodular Goiter (NG), Normal (NL), and Thyroiditis (TI).
• Applied CA, SD, and FM models for data augmentation, creating 18 scenarios, and trained ResNet18 classifiers.

Main Results:

• Flow Matching (FM) based methods, particularly Original dataset + FM (O+FM), achieved the highest F1-scores (0.78) and AUC values (0.95).
• FM generated the most realistic images, indicated by the lowest FID (0.66) and KID (0.83).
• Stable Diffusion (SD1) with image and prompt inputs showed strong performance (macro F1: 0.76), emphasizing the value of clinical context.

Conclusions:

• Advanced generative models like FM and clinically-informed SD substantially enhance thyroid scintigraphy classification accuracy.
• High-fidelity generative augmentation can surpass conventional augmentation strategies for limited datasets.
• These findings underscore the importance of generative AI in improving diagnostic tools for thyroid disorders.