A Unified Framework for Comprehensive Cardiac CT Segmentation and Phenotyping

The paper presents an end-to-end framework for cardiac CT segmentation and phenotyping, combining human-in-the-loop annotation, a cardiac CT augmentation method, and self-supervised pre-training on 60,000 unlabeled scans. The resulting expert-annotated dataset contains 1,598 cases and 14 cardiac structures, with 1,000 training cases and 598 external-test cases. Across five external datasets, the framework reportedly outperformed existing open-source tools in accuracy and coverage. Comparisons among convolutional, transformer, and state-space architectures showed similar performance, suggesting that annotation quality and pre-training were more influential than architecture choice. The authors also evaluated population-scale phenotyping and released data, code, weights, augmentation tools, and software.