Automated placement of the bolus-tracking region of interest using U-Net-based aortic root segmentation in coronary CT angiography.

Abstract

In coronary CT angiography (CCTA), deviation of the bolus-tracking region of interest (ROI) from the aortic root (AR) during CT fluoroscopy may lead to inappropriate scan timing. This study developed and evaluated an automated ROI placement system based on the U-Net segmentation of the AR in CT fluoroscopic images. A total of 1663 CT fluoroscopic images from 57 patients who underwent CCTA were analyzed. A patient-level five-fold cross-validation was performed; in each iteration, one fold was used for testing, one for validation, and the remaining three for training. The training images were augmented by rotation. A 2D U-Net was trained to segment the AR region, and postprocessing consisted of Otsu thresholding, median filtering, and connected-component analysis. A circular ROI corresponding to 25% of the predicted AR area was placed at the centroid of the predicted AR region. Segmentation accuracy was assessed using the Dice similarity coefficient (DSC), and ROI placement accuracy was evaluated using centroid distance between the predicted and ground-truth regions and complete containment of the ROI mask within the ground-truth AR region. The median DSC was 0.957, and the median centroid distance was 0.7 mm. All automatically placed ROIs were completely contained within the ground-truth AR region. The U-Net inference time was 72.55 ± 15.72 ms per image, and the total processing time from image loading to ROI placement was 94.8 ± 14.6 ms per image. The proposed method can enable accurate, real-time, and operator-independent dynamic ROI placement for CCTA bolus tracking.