A Near-Infrared Imaging System for Robotic Venous Blood Collection.

Area of Science:

• Medical Robotics
• Medical Imaging
• Computer Vision

Background:

• Robotic venous blood collection offers potential over manual methods.
• High-quality vein imaging is crucial for successful robotic blood collection.
• Accurate vein segmentation is essential for precise robotic navigation.

Purpose of the Study:

• To develop an advanced vein imaging device and a robust vein segmentation algorithm.
• To improve the accuracy and reliability of vein imaging for robotic applications.
• To enhance the performance of vein image segmentation using deep learning and optimized stereo matching.

Main Methods:

• Developed a vein imaging device based on simulation analysis of imaging parameters.
• Proposed a U-Net+ResNet18 neural network integrating ResNet18 residual blocks for vein segmentation.
• Utilized Bootstrap Your Own Latent (BYOL) pre-training for ResNet18 encoder, transferring parameters to enhance segmentation with limited data.
• Optimized the AD-Census stereo matching algorithm with a variable-weight approach for improved adaptability.

Main Results:

• The BYOL+U-Net+ResNet18 method demonstrated significant improvements over standard U-Net, including an 8.31% reduction in Binary Cross-Entropy (BCE) and a 15.95% increase in Intersection over Union (IoU).
• Hausdorff Distance (HD) was reduced by 5.50%, and the Dice coefficient (Dice) increased by 9.20%, indicating superior segmentation quality.
• The optimized AD-Census stereo matching algorithm achieved a 25.69% reduction in average error, showing enhanced stereo matching performance.

Conclusions:

• The developed vein imaging system and U-Net+ResNet18 network significantly improve vein image segmentation accuracy.
• Optimized stereo matching enhances the reliability of the vein imaging system.
• The system shows promise for real-time puncture guidance in future robotic venous blood collection applications.