Beyond visual realism: assessing coronary artery dominance diversity in Al-generated synthetic coronary computed tomography angiography images

Generative AI has the potential to enhance downstream AI tasks in cardiovascular medicine by augmenting or replacing datasets with realistic synthetic data [1]. However, for synthetic data to be reliable, it must also capture true anatomical variability, not just visual realism. This challenge could be more pronounced for complex yet clinically relevant structures such as coronary arteries. Traditional metrics and even expert radiological evaluations that focus on realism or diagnostic quality can overlook critical distributional biases. Consequently, imbalances in anatomical diversity present in the training data can be amplified in the synthetic data without getting noticed, introducing biases into downstream applications, a critical issue given that underrepresented variants can correlate with a more severe prognosis [2].

This study assesses if a generative model's distributional accuracy reflects the training data's coronary artery dominance or exhibits mode collapse [1] by overproducing the most common variant.

Following institutional ethics approval (ID xxx), a dataset of 136 Photon Counting Coronary Computed Tomography Angiography (PCCTA) volumes was used. A latent diffusion model [3] was trained on 120 scans, validated on 16, and generated 60 synthetic volumes. An experienced radiologist confirmed the high perceptual quality of these volumes using four distinct metrics (Fig. 1), and afterwards we assessed their distributional accuracy. A trained observer labeled all real and synthetic volumes for coronary artery dominance. The resulting distributions were compared using a Chi-Squared Goodness of Fit test.

The real data’s coronary dominance distribution (78.4% right-, 13.0% left-, 8.4% co-dominant) was consistent with population data (p=0.335) (Fig. 2) [2, 4]. In stark contrast, the synthetic dataset exhibited severe biases. (98.2%, 1.8%, 0%) of synthetic volumes were (right, left, co)-dominant (p<0.005). Four of 60 volumes were excluded due to anatomically implausible errors, such as vessels being disconnected from the aortic root. Conclusion: Our findings reveal a crucial disconnect between perceptual quality and distributional accuracy. Despite receiving high expert ratings, the synthetic model exhibited severe mode collapse, almost exclusively generating the most common right-dominant variant. Fundamental errors, such as vessels disconnected from the aortic root, reveal that images can pass initial quality assessments, while still containing critical clinical flaws. Notably, our analysis focused on a single anatomical feature. Similar mode collapse may affect other clinically relevant structures. This has serious implications for the use of generative models in data synthesis, underscoring the need for rigorous validation, including distributional analyses and proactive bias mitigation strategies to ensure dataset diversity and reliability.