Cardiac CT foundation model for whole-heart segmentation: integrating human-in-the-loop and self-supervised learning
P Mohammadi Kazaj, L Fridolin Weber, W Xie, A Mokhtari, S A A Safavi-Naini, G Baj, A Stark, T Pilgrim, G C M Siontis, X Xu, M Hundertmark, R R Buechel, S Windecker, C Graeni, I ShiriAbstract
Background
Cardiac computed tomography (CCT) provides high-resolution data essential for personalized medicine and clinical evaluations; however, manual segmentation of cardiac structures is prohibitively time-consuming and prone to intra- and inter-observer variability.
Objective
To manually label a comprehensive, multi-center dataset using a human-in-the-loop approach and develop a deep-learning foundation model for the precise, automated 3D segmentation of diverse cardiac substructures, including chambers, vasculature, and adipose tissues.
Methods
A total of 1,559 CT scans were used to develop several segmentation models within the nnUZoo framework, including Swin-Mamba, U2Net, SwinUNETR, and the default nnU-Net. A human-in-the-loop approach with active learning was implemented to iteratively train models and refine inference labels with the help of two physicians for the coronary arteries, myocardium, major vascular structures, and peri-/epicardial fat. To ensure robustness, CT-specific augmentations, such as metal artifacts and calcification simulations were employed. Additionally, a self-supervised foundation model was trained on more than 60,000 cases to enhance performance, particularly in data-scarce scenarios.
Results
The model achieved a mean Dice score of 0.96 across all substructures, outperforming existing open-source tools. Specialized augmentations ensured robust performance in complex cases, while the foundation model provided an average 5 percent performance boost in low-data regimes.
Conclusion
We present a highly accurate, automated 3D segmentation tool for CCT. By enabling the consistent extraction of quantitative metrics, this model facilitates large-scale clinical research. Both the models and the labeled datasets have been made openly accessible.ct-seg-esc