U-Net and YOLOv8 Artificial Intelligence Models for Automated Recognition of Internal Jugular Veins and Radial...

This study compares the U-Net and You Only Look Once version 8 (YOLOv8) models for identifying internal jugular veins (IJVs) and radial arteries (RAs) in longitudinal and/or transversal ultrasound views, evaluating their vascular recognition capabilities for artificial intelligence-guided cannulation systems under point-of-care ultrasound (POCUS) visualization.

A retrospective study.

Six teaching hospitals.

Data from 1,122 ultrasound images (612 IJVs, 510 RAs) between January and December 2024.

No intervention.

U-Net was employed for pixelwise segmentation with a combined dice-cross-entropy loss, while YOLOv8s-seg incorporated attention mechanisms for object detection. Model performance was evaluated using precision, recall, F1-score, mean average precision at intersection over union (IoU) threshold 0.5 (mAP@0.5), IoU, Dice coefficient, and inference time metrics. YOLOv8 demonstrated superior performance with a precision of 0.996, recall of 1.00, mAP@0.5 of 0.995, IoU of 0.739, Dice coefficient of 0.834, and inference time of 26.3 ms, outperforming U-Net, which achieved a precision of 0.988, recall of 0.998, mAP@0.5 of 0.993, IoU of 0.719, Dice coefficient of 0.816, and inference time of 38.9 ms. In the validation set, YOLOv8 exhibited higher accuracy across all categories, with values of 0.96 for V-S, 0.90 for V-L, 0.89 for A-S, and 0.86 for A-L, compared with U-Net's respective accuracies of 0.81, 0.88, 0.87, and 0.82.

YOLOv8 outperformed U-Net in accuracy, precision, and speed, proving more suitable for real-time vascular localization in clinical settings. These findings provide critical algorithmic foundations for developing automated vascular puncture navigation platforms in the future.