Deep learning model for automated identification of ventrally positioned right hepatic artery in contrast-enhanced computed tomography of pediatric congenital biliary dilatation: development and clinical application.

BACKGROUND

Preoperative identification of a ventrally positioned right hepatic artery (vRHA) is critical in congenital biliary dilatation (CBD), as unrecognized variants increase surgical risk. Detection on computed tomography (CT) is challenging in routine pediatric practice.

OBJECTIVE

To develop and validate a You Only Look Once version 12 (YOLOv12)-based model for vRHA identification in contrast-enhanced CT using a targeted key-slice strategy.

MATERIALS AND

METHODS

In this retrospective single-center study, 232 CBD patients (116 vRHA, 116 controls) were divided into training (n=186) and test (n=46) sets. Five YOLOv12 sub-models were trained as second-stage classifiers using 1,452 radiologist-selected key arterial-phase slices. Performance was assessed by precision, recall, F1-score, mean average precision (mAP), and area under the curve (AUC). Diagnostic performance was compared with two radiologists using DeLong's test.

RESULTS

All sub-models showed perfect precision (1.000) with recall ranging from 0.684 to 0.895. YOLOv12n achieved the best performance (recall 0.842, F1-score 0.914, mAP50 0.989, AUC 0.977; 95% confidence interval, 0.913-1.000). It significantly outperformed the junior radiologist (AUC 0.737, P<0.001) and demonstrated comparable performance to the senior radiologist (AUC 0.947, P=0.515).

CONCLUSION

The YOLOv12n model achieved excellent diagnostic performance for vRHA identification on key CT slices and performed at a senior radiologist level, supporting its potential role in preoperative assessment.