An improved YOLO Nano model for dorsal hand vein detection system

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• Venipuncture is essential but challenging due to vein variability.
• Near-infrared (NIR) imaging aids venipuncture, but current systems lack embedded vein identification.
• Existing research focuses on vein segmentation, not suitability for puncture in embedded systems.

Purpose of the Study:

• To develop a practical system for detecting and locating optimal veins for venipuncture using deep learning.
• To create an efficient embedded system for real-time vein analysis.
• To improve the accuracy and speed of vein detection for medical applications.

Main Methods:

• Designed a dedicated NIR-based system for dorsal hand vein image acquisition and puncture site localization.
• Developed YOLO Nano-Vein, an optimized deep learning network based on YOLO Nano with architectural modifications (trimmed architecture, reduced output scales, ASPP).
• Evaluated the system's performance using average precision (AP), detection time, and network parameters.

Main Results:

• The YOLO Nano-Vein system achieved an increased average precision (AP) from 91.68% to 93.23%.
• Detection time and network parameters were reduced by 22% and 17.5%, respectively, compared to YOLO Nano.
• The proposed network demonstrated higher accuracy and reduced detection time versus YOLO Nano and YOLOv3.

Conclusions:

• The YOLO Nano-Vein system offers a practical and accurate solution for identifying optimal puncture veins.
• The optimized deep learning approach enhances performance on embedded devices with limited computational resources.
• This technology has strong applicability for improving venipuncture procedures in clinical settings.