Deep learning-based automated segmentation of femoral vessels for mixed reality-guided vascular access
R Miyabe, T Nakamura, A Higaki, M Isshiki, Y Nakao, T Fujisawa, Y Akazawa, T Miyoshi, H Kawakami, H Higashi, S Tamaki, K Nishimura, K Inoue, S Ikeda, O YamaguchiAbstract
Background
Current ultrasound-guided central venous access has not sufficiently reduced the incidence of complications, such as accidental arterial puncture. Therefore, there is a clinical demand for more secure and precise techniques utilizing digital health technologies such as mixed reality (MR)-guided navigation.
Purpose
As a fundamental step toward achieving MR-guided vascular access, we aimed to develop an automated model to segment the contours of femoral arteries and veins from ultrasound images and reconstruct them into 3D models.
Methods
We prospectively enrolled 50 consecutive patients undergoing cardiac catheterization between June 2023 and December 2025. Ultrasound videos of the femoral region (50 × 100 mm field of view, 50 mm depth) were recorded during manual hand-scans. From these videos, 22 static frames per patient were extracted at regular intervals. Ground truth labels were generated by two board-certified cardiologists through independent contour tracing. A MedSAM architecture was utilized for semantic segmentation, and model performance was evaluated using the Dice coefficient and Intersection over Union (IoU) via 5-fold cross-validation.
Results
The study cohort (median age, 74.5 years [IQR, 64.3–78.0]; 86% male; median BMI, 24.3 kg/m² [IQR, 21.7–26.3]) yielded a total of 1,100 images. The MedSAM model demonstrated high accuracy in detecting femoral vessels, achieving a Dice coefficient of 0.82 and an IoU of 0.70. Furthermore, 3D vascular structures were successfully reconstructed from the predicted contours.
Conclusion
We demonstrated a proof of concept for generating 3D vascular models from hand-scanned femoral ultrasound images via semantic segmentation, providing a technical foundation for future MR-guided navigation.