Abstract | ||
---|---|---|
Accessing the venous bloodstream to obtain a blood sample is the most common clinical routine. Nevertheless, due to the reliance of venipuncture on manual technique, first-stick accuracy of venipuncture falls below 50% in difficult cases. A surge of research on robotic guidance for autonomous vascular access have been conducted. With regard to robotic venipuncture, efficiency and accuracy of vein segmentation is of much importance. This paper describes a method to accurately and efficiently segment, localize and track the topology of human veins from near-infrared (NIR) images. Both spatial and color augmentation are implemented on the dataset at first. Next, Mixer-UNet is used for identifying veins that would be hard to find in clinical visual assessment. The Mixer-UNet is developed on the basis of UNet and MLP-Mixer. Through the flexible information exchange through Token-mixing layer and Channel-mixing layer, Mixer-UNet can extract features from NIR images accurately. The performance of Mixer-UNet is validated on 270 NIR images, which are collected from 30 volunteers. Mixer-UNet reaches 93.07% on Accuracy indicator. Compared with the best-performing baseline, the F1-score indicator increases by 2.82%, reaching 78.37% in testing sample. The high accuracy and robustness of Mixer-UNet is expected to improve the vein segmentation of NIR images, and further contributes to the goal of an improved automated venipuncture robot. |
Year | DOI | Venue |
---|---|---|
2022 | 10.1007/978-3-031-13841-6_6 | INTELLIGENT ROBOTICS AND APPLICATIONS (ICIRA 2022), PT IV |
Keywords | DocType | Volume |
Medical robotics, Venipuncture robot, Vein segmentation model, MLP-Mixer, UNet | Conference | 13458 |
ISSN | Citations | PageRank |
0302-9743 | 0 | 0.34 |
References | Authors | |
0 | 5 |