Portable Venipuncture Robot
for Autonomous Vascular Access

We present a portable robotic device capable of performing autonomous vascular access with minimal supervision. Our 9-DOF robot combines NIR imaging and ultrasound technology to accurately locate and access veins, addressing challenges in traditional venipuncture procedures.

The robot features a compact design with positioning, support, information processing, and puncturing units. It utilizes advanced algorithms for vein segmentation and real-time needle guidance, ensuring high precision and safety during the procedure.

Our experimental results demonstrate a success rate of 98.24% in simulated tests, with a maximum puncture error of 1.79mm. This innovation has the potential to significantly improve patient care, reduce the workload on medical staff, and minimize the risk of complications in venipuncture procedures.

Our venipuncture robot features a 6-DOF compact mechanism designed for precise vascular access. The robot's structure is built around a reliable short transmission chain, primarily consisting of precision ball screws and Maxon motors, ensuring high mechanical efficiency and accuracy.

The motor system comprises six key components:

• Motors 1, 2, and 3 operate jointly for x-y plane movement
• Motor 4 determines the puncture angle
• Motor 5 controls needle movement along the y-direction
• Motor 6 advances the needle to the vein puncture point

A key feature is the 3-DOF Needle Module, which includes:

• A linear stage driven by a lead screw spindle
• A worm drive system for independent insertion angle control
• A self-locking rotational joint for safety
• Two ball screw mechanisms for precise needle movement

This design allows for accurate adjustment of puncture speed and dramatically improves puncture stabilityand needle tip positioning accuracy.

Near Infrared Imaging

The robot employs Near Infrared (NIR) imaging technology for accurate vein detection. This non-invasive method allows for clear visualization of superficial veins, enhancing the robot's ability to identify suitable puncture sites.

We randomly select six NIR images as examples and show the vein segmentation and suitable puncture areas in Fig. 8. We propose a Dual-In-Dual-Out network with two-step learning and two-task learning to determine the suitable puncture area and angle from the NIR image inputs. A visual illustration of the proposed network is shown in Fig. 9. It contains two steps of training: first, it trains a Single-In-Single-Out network to segment the vein from the NIR image; second, it inputs both the NIR image and vein segmentation from the first step training into the Dual-In-Dual-Out network to regress the suitable puncture area and angle.

Five examples of the suitable puncture area regression by the four methods are shown in Fig. 10. We can visually see that both the Dual-In-Single-Out and Dual-In-Dual-Out network can distinguish between the suitable and non-suitable puncture area better than the Single-In-Single-Out and Single-In-Dual-Out network, indicating the importance and value of bringing the vein segmentation into the network's input. For the regression of suitable puncture areas and angle, the mean and std DSC are shown in Table Ⅰ and Table Ⅱ.

Near Infrared Imaging

Complementing the NIR system, ultrasound technology is used for precise depth estimation and longitudinal vein imaging. The ultrasound device (st-1c transducer, frequency 7.5MHz) provides high-resolution images of vein cross-sections, enabling accurate calculation of puncture depth.

An illustration of the dataset is shown in Fig. 11, where green boxes illustrate the veins, yellow boxes illustrate the vein shadow and red boxes represent the vessel edges. In previous studies, UNet, FPN, and other models have shown good performance in image segmentation. To merge the advantages of these deep-learning neural networks, our model integrates multiple image segmentation networks. Through stacking methods and feature image coding, we propose the Integrated Segmentation Model (ISM), offering high precision for vein segmentation from ultrasonic images. As shown in Fig. 12, the overall structure of ISM includes two layers. The first layer is composed of three sub-models (FPN, PSPNet, and UNet, numbered as models 1, 2, and 3). The output result diagram of the first layer was used as the input training picture of the second layer. The second layer of the sub-model is composed of LinkNet, which takes the label picture (GT) of the original data set as the recognition target.

As can be seen in Table II, the proposed ISM demonstrates its superiority over traditional models in terms of the three indicators, namely Dice-Similarity-Coefficient (DSC), Hausdorff-Distance (HD95) and Intersection-Over-Union (IOU). The ISM model achieved significant improvements in multiple indicators: its DSC value increased by about 6 %, reaching 94.62 %, and the IOU value increased by about 11%, in complex samples. Finally, a clear puncture point is calculated through the connection domain algorithm. Experimentally, the success rate of selecting the suitable vein for puncture is 99.21%.

Training procedure of the FPN, UNet, PSPNet and ISM: (a) Dice Loss curve of FPN, UNet, PSPNet, ISM. (b) ROC-AUC curve of FPN, UNet, PSPNet, ISM of testing picture. (c) The DSC curve of vein segmentation results using different models: FPN, UNet, PSPNet, LinkNet, ISM.

The robot's control system integrates data from both NIR and ultrasound imaging to guide the needle with high precision. Advanced algorithms process the imaging data in real-time, determining the optimal puncture site, angle, and depth.The system allows for dynamic adjustment during the procedure, compensating for any patient movement or vein deformation. This ensures a high success rate and minimizes the risk of complications.