The success of diffusion models in medical imaging highlights their potential to generate high-quality synthetic datasets that closely resemble real clinical data, addressing limited dataset availability and patient privacy concerns. We present DiffusionTBAD , a novel text-to-image (TtI) diffusion-based pipeline for synthesizing diagnostically accurate computed tomography angiography (CTA) images of type B aortic dissection (TBAD). Using few-shot learning, DiffusionTBAD fine-tunes a TtI diffusion model with guided textual prompts to capture the distinct features and variability of TBAD cases. The synthetic data are evaluated using quantitative diversity and similarity metrics, as well as downstream task performance. Augmenting real TBAD datasets with synthetic images improved supervised classification accuracy from 67% to 76%, and pretraining on synthetic images increased segmentation Dice scores from 66% to 70%. Additionally, qualitative assessment by eight healthcare professionals confirmed the high visual realism and diagnostic plausibility of the generated images. These results demonstrate that DiffusionTBAD can enhance model performance while reducing reliance on real patient data, enabling privacy-preserving development of medical imaging models. • We propose DiffusionTBAD, a text-to-image diffusion model for TBAD synthesis. • DiffusionTBAD uses LoRA and few-shot tuning for efficient synthetic data generation. • We release synthetic TBAD images across five key lumen-based class variants. • Extensive evaluation shows synthetic data supports ML tasks and visual realism. • Clinical experts validated realism via paired and class-specific evaluations.
Abaid et al. (Sun,) studied this question.