A Unified Framework for Comprehensive Cardiac CT Segmentation and Phenotyping: Human-in-the-Loop Data Annotation, Vision Foundation Model Development, Multicenter Evaluation and Clinical Validation

Cardiac computed tomography (CT) is a standard diagnostic tool that captures extensive anatomical information beyond the specific clinical indication for which it was ordered. However, manual quantification of these structures is time-consuming and prone to observer variability, limiting its routine use. This study aimed to develop a scalable, automated, and accurate framework for comprehensive cardiac segmentation and phenotyping that can generalize across different clinical populations and imaging protocols.

The researchers employed a data-centric strategy to build a unified framework. They curated a large-scale, expert-annotated dataset of 1,598 cases covering 14 distinct cardiac structures. To address the scarcity of labeled data, they implemented a human-in-the-loop (HITL) annotation pipeline and developed a self-supervised foundation model (CCT-FM) pre-trained on over 60,000 unlabeled cardiac CT scans. Additionally, they created a specialized augmentation library, CTAug, to improve model robustness against imaging artifacts. The framework was benchmarked against various architectures—including convolutional, transformer, and state-space models—and compared against established open-source tools like TotalSegmentator, Atlas, and MOOSE.

The study demonstrates that self-supervised pre-training significantly boosts segmentation performance, particularly in low-data regimes and for challenging, thin, or low-contrast structures like the pulmonary arteries. When fully trained, the framework achieved superior accuracy compared to existing open-source tools across all shared structures, with the most dramatic improvements observed in the segmentation of coronary arteries. Notably, the benchmarking revealed that once high-quality data and pre-training are utilized, the choice of model architecture (e.g., U-Net vs. Transformer) has a negligible impact on overall segmentation accuracy. The framework also successfully enabled population-level phenotyping, providing functionally relevant insights into ventricular function and disease severity.