A Manually Annotated, Open‐Source Dataset for Coronary, Aortic, and Valvular Calcium Segmentation in Non‐Gated Chest CT

ABSTRACT Automated cardiovascular calcium scoring on non‐gated chest CT offers a major opportunity for opportunistic cardiovascular risk stratification. However, progress is hindered by the scarcity of publicly available, annotated datasets that differentiate between coronary, aortic, and valvular sources. This study aimed to develop, characterize, and release a manually annotated, open‐source dataset for cardiovascular segmentation and to provide a baseline deep learning model performance benchmark. A dataset of 203 non‐gated chest CT scans from the Stanford AIMI COCA dataset was manually annotated for calcium across eight anatomical classes, with all segmentations verified by an expert radiologist. A YOLOv8m‐Seg model was trained on an 80% patient‐level split (171 scans) and evaluated on the remaining 20% (42 scans). Performance was assessed using instance segmentation metrics, and the dataset's utility was demonstrated with an illustrative risk stratification benchmark based on a total thoracic calcium score. The final dataset comprised 1649 distinct calcium instances, exhibiting significant class imbalance (thoracic aorta: 68.65%). The baseline YOLOv8m‐Seg model yielded moderate segmentation performance (Mask mAP50: 14.4%). The illustrative benchmark demonstrated that models trained on this data can learn features correlated with high calcium burden. This publicly available, expertly verified dataset is a valuable resource for developing and validating algorithms for comprehensive cardiovascular calcium analysis on non‐gated CT. The primary value lies in the dataset itself, which enables the development of next‐generation AI models capable of anatomically precise distinction between coronary and non‐coronary calcifications for more cardiovascular risk stratification, warranting further research and external validation.