Deep learning-based automated assessment of pulmonary artery indices and surgical approach triage for tetralogy of Fallot from multicenter cardiac computed tomography (CT).

AIM: To develop an automated framework for deep learning (DL)-based segmentation, pulmonary artery indices (PAIs) computation, and rule-based surgical triage from cardiac CT (CCT) in pediatric Tetralogy of Fallot (TOF). MATERIALS AND METHODS: Preoperative CCT scans of pediatric TOF patients were collected from the primary internal centers and two external centers. A modified 3D nnU-Net architecture was used for automated segmentation of the pulmonary arteries (PA) and descending aorta (DAO). Following segmentation, PAIs were automatically computed using predefined geometric algorithms. Segmentation performance, measurement agreement, and processing efficiency were evaluated. Surgical triage analysis was performed only in the primary internal dataset because the external dataset lacked palliative cases. RESULTS: 122 TOF patients were included (the training dataset n = 80, the internal validation dataset n = 20, the external dataset n = 22). The DL model achieved high segmentation performance, with PA/DAO dice similarity coefficient ranging (DSC) from 0.92 to 0.98 across datasets. DL-derived PAIs showed good agreement with manual measurements and reduced processing time by 90.11% (30.28 s vs. 205.00 s, P < 0.001). Automated PAIs demonstrated good rule-based surgical triage ability (AUC > 0.87). CONCLUSIONS: This automated framework enables rapid and automated PAI quantification from CCT in pediatric TOF across multi-center datasets. DL-derived PAIs demonstrated comparable performance to manual measurements for rule-based surgical triage.