DiffusionTBAD: Rendering CTA images for type B aortic dissection diagnosis.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Radiology

Background:

• Diffusion models show promise for generating synthetic medical data, addressing limitations in dataset size and patient privacy.
• Computed tomography angiography (CTA) is crucial for diagnosing type B aortic dissection (TBAD), but data scarcity poses challenges.

Purpose of the Study:

• To introduce DiffusionTBAD, a novel text-to-image diffusion pipeline for synthesizing diagnostically accurate CTA images of TBAD.
• To evaluate the utility of synthetic TBAD data for improving AI model performance.

Main Methods:

• Fine-tuning a diffusion model using few-shot learning and text prompts specific to TBAD features.
• Quantitative evaluation of synthetic data diversity and similarity, and assessment of downstream task performance (classification and segmentation).
• Qualitative assessment of image realism and diagnostic plausibility by healthcare professionals.

Main Results:

• Augmenting real TBAD datasets with synthetic images improved classification accuracy from 67% to 76%.
• Pre-training AI models on synthetic data increased segmentation DICE scores from 66% to 70%.
• Expert evaluation confirmed the high visual realism and diagnostic accuracy of generated images.

Conclusions:

• DiffusionTBAD effectively generates high-quality, diagnostically relevant synthetic TBAD CTA images.
• Synthetic data generated by DiffusionTBAD can significantly enhance the performance of medical imaging AI models.
• This privacy-preserving approach facilitates the development of robust AI tools for TBAD diagnosis.