AI-Augmented Thyroid Scintigraphy for Robust Classification of Disease

Technical Summary: AI-Augmented Thyroid Scintigraphy for Robust Classification of Disease

Problem Statement
Thyroid scintigraphy is a critical diagnostic tool for evaluating thyroid disorders, providing unique functional insights that other modalities like ultrasound or CT cannot. However, the interpretation of these images is often subjective and time-consuming. While deep learning (DL) offers a path toward automated, objective diagnosis, its application in this domain is hindered by data scarcity and class imbalance. Medical datasets are frequently limited in size due to privacy constraints and logistical challenges, and conventional augmentation techniques (e.g., rotation, flipping) often fail to sufficiently address these limitations or may generate unrealistic variations. There is a need for advanced generative strategies that can synthesize high-fidelity, anatomically accurate images to robustly train DL classifiers on limited data.

Methodology
The study investigates three data augmentation strategies to enhance the classification of thyroid scintigraphy images into four categories: Diffuse Goiter (DG), Nodular Goiter (NG), Normal (NL), and Thyroiditis (TI). The research utilizes a dataset of 2,954 anterior thyroid scintigraphy images collected from nine medical centers.

1. Data Preprocessing and Prompt Engineering:

   • Thyroid regions were manually segmented by an experienced nuclear medicine physician.
   • To guide generative models, the authors developed a pipeline to extract structured clinical information from physician reports. Using GPT-4 Turbo, they converted unstructured reports into standardized prompts (under 77 tokens) containing specific clinical attributes (e.g., thyroid function, nodule type, uptake patterns).

2. Augmentation Strategies:

   • Conventional Augmentation (CA): Applied standard geometric transformations (rotation, flipping, scaling) and noise addition.
   • Stable Diffusion (SD): The authors implemented four variations of SD-based augmentation:
     • SD1: Image + Prompt to Image.
     • SD2: Prompt to Image.
     • SD3: Image to Image (no prompt).
     • SD4: Mask + Prompt to Image.
       These models were fine-tuned on the dataset using mixed precision training.
   • Flow Matching (FM): A diffusion-based approach that learns a continuous-time velocity field to define a near-optimal transport between the source data distribution and a target prior. This method allows for efficient sampling via Ordinary Differential Equations (ODEs) with fewer integration steps (10 steps) compared to traditional diffusion processes. FM was conditioned on class labels and masks.

3. Classification and Evaluation:

   • A ResNet18 classifier, initialized with ImageNet weights, was trained on 18 distinct scenarios combining the original dataset (O) with various augmentation strategies (e.g., O+FM, O+SD1, O+CA+FM).
   • Models were trained on data from six centers (A–F) and evaluated on an external dataset from three centers (G–I) to assess generalization.
   • Performance was measured using precision, recall, F1-score (micro, macro, weighted), and ROC AUC.
   • Image fidelity was assessed using Fréchet Inception Distance (FID) and Kernel Inception Distance (KID).
   • Statistical significance was determined using bootstrapping and the Wilcoxon rank sum test with Benjamini-Hochberg correction.

Key Contributions

• Novel Application of Flow Matching: This study is among the first to apply Flow Matching for medical image augmentation in thyroid scintigraphy, demonstrating its superiority over both conventional methods and Stable Diffusion variants in terms of image fidelity and classification performance.
• Clinically Informed Prompting: The authors integrated physician-generated prompts derived from actual case reports into the Stable Diffusion pipeline, demonstrating that domain-specific textual context significantly enhances the realism and diagnostic utility of generated images.
• Comprehensive Comparative Analysis: The work provides a rigorous comparison of 18 augmentation scenarios, establishing that high-fidelity generative augmentation (specifically FM) can outperform and even supersede the need for combining generative methods with conventional heuristics.

Results

• Classification Performance: The O+FM (Original + Flow Matching) configuration achieved the highest performance across all metrics, with micro, macro, and weighted F1-scores of 0.78, 0.77, and 0.78, respectively, and AUC values of 0.95, 0.93, and 0.94. This configuration statistically outperformed all other methods, including the combination of FM with conventional augmentation (O+FM+CA).
• Stable Diffusion Variants: Among SD methods, SD1 (Image + Prompt) was the most effective, achieving a macro F1 of 0.76. This highlights the critical value of combining visual inputs with clinically derived prompts. SD3 (Image-to-Image without prompts) and SD4 (Mask + Prompt) performed significantly worse, indicating that prompts are essential for SD to generate diagnostically relevant variations.
• Image Fidelity: FM generated the most realistic images, achieving the lowest overall FID (0.66) and KID (0.83). In contrast, SD methods showed higher FID/KID scores, with SD4 exhibiting particularly poor performance (FID: 17.99, KID: 33.59).
• Efficiency: FM demonstrated superior computational efficiency, requiring only 10 integration steps to generate images (0.19 seconds/image) compared to the 50 steps required by Stable Diffusion (1.25 seconds/image) on the same hardware.
• Interpretability: Grad-CAM visualizations confirmed that models trained with FM correctly localized attention to the anatomical position of the thyroid gland, even in challenging cases like Thyroiditis where functional uptake is low, suggesting the models learned true pathophysiological features rather than background artifacts.

Significance and Claims
The paper claims that integrating advanced generative models, particularly Flow Matching, substantially improves the robustness of thyroid scintigraphy classification in the face of limited and imbalanced datasets. The authors assert that FM-based augmentation preserves key image features and enhances model generalization more effectively than conventional heuristics or standard Stable Diffusion methods.

A central finding is that high-fidelity generative augmentation can supersede conventional augmentation; the O+FM model outperformed O+FM+CA, suggesting that the realistic variations generated by FM are sufficient on their own without the addition of random geometric transformations. Furthermore, the study underscores the necessity of domain knowledge in generative AI, as physician-derived prompts were shown to be critical for guiding Stable Diffusion models to produce clinically relevant synthetic data.

The authors conclude that their methodology offers a pathway toward more objective, consistent, and efficient thyroid disease diagnosis, potentially aiding nuclear medicine physicians. However, they modestly acknowledge limitations, noting that the dataset is restricted to a single ethnic group and that future work should validate these strategies on more diverse, multi-ethnic global cohorts and potentially integrate multimodal data (e.g., laboratory results and ultrasound).