Automated cardiac chamber segmentation using non-contrast CT images

Accurate assessment of cardiovascular risk using low-dose computed tomography (CT) scans is crucial for managing cardiovascular diseases (CVD), the leading cause of death in the United States. However, automatic segmentation of non-contrast CT images is challenging due to the scarcity of labels and the complexity of manual annotation. Existing segmentation methods primarily focus on contrast-enhanced CT and do not apply to patients with only non-contrast scans. In this study, we address this gap by leveraging a dataset of 55 patients who each underwent both contrast-enhanced and non-contrast ECG-gated CT scans. After co-registering these paired datasets, we transferred labels from contrast-enhanced images to their spatially aligned non-contrast counterparts. Due to limited availability of paired data, we additionally employed a generative adversarial network (GAN) to synthesize non-contrast CT images from the contrast-only scans, thereby expanding our training set. We trained an nnUNet model on the combined real and synthesized non-contrast dataset and incorporated Hausdorff distance (HD) loss to improve boundary delineation. This approach achieved an averaged Dice Similarity Coefficient of 0.89 on non-contrast CT images, with predicted left and right ventricular chamber sizes showing strong correlations (R² = 0.91) compared to ground truth measurements from contrast-enhanced scans. Notably, the synthesized images further improved segmentation accuracy for the left ventricular chamber. This fully automated method enables precise boundary segmentation and robust, large-scale evaluation of cardiac structures on non-contrast CT scans, thereby facilitating earlier and more effective cardiovascular risk stratification.