An imaging system and processing algorithm for the extraction and three-dimensional imaging of subcutaneous blood vessels is proposed to overcome the difficulty of vascular recognition in thick parts of surface tissue. Vascular visualization technology is used in the medical field to treat scenarios such as venipuncture and interventional therapy to reduce the additional trauma to the patient. Since hemoglobin in the blood has a higher absorption rate of light in the near-infrared (NIR) band (700-1000 nm) than lipids, proteins, and water, vascular tissue appears as a dark shadow area projected on the surface of the skin in images taken in the NIR band, and the position of the shadow area changes with the viewing perspective. According to the above principles, some researchers use multi-view imaging technology to perform three-dimensional reconstruction of subcutaneous blood vessels. This technique consists of two main steps: the first one is the vascular segmentation on the grayscale image, and the second one is the stereo matching on multi-view images to reconstruct the three-dimensional information of blood vessels. However, in the available literature, the applicable body parts of the equipment are limited due to the light source and camera arrangement. Other drawbacks include the noise line segment in the extraction result and the lack of algorithm efficiency optimization. Therefore, we hope to design a vascular recognition module for the automatic puncture robot from the aspects of the light source and camera arrangement design and the improvement in the vascular skeleton extraction algorithm.

The optimization of the vascular segmentation effect includes the optimization of the original image quality and that of the image processing algorithm. Some studies have shown that improving the irradiance uniformity of the light source on the body surface can make the vascular region more distinctive in the grayscale image. Given such knowledge, our imaging system design uses a convergent binocular NIR-enhanced camera kit and a NIR LED array. We calculate the radiation of the LED bead according to the irradiance distribution formula of the approximate Lambertian source and use MATLAB software to simulate the total irradiance distribution of the LED array on a cylindrical surface (Fig. 2) and make a symmetrical two-board LED array light source according to the optimal design parameters (Fig. 3). The subsequent research on vascular skeleton extraction is carried out on the images taken with the designed imaging system. It includes seven steps: 1) selecting the region of interest (ROI); 2) weakening the image background; 3) performing contrast-limited adaptive histogram equalization (CLAHE); 4) performing two-dimensional Frangi filtering of multi-scale images; 5) performing Otsu's adaptive-threshold image binarization; 6) extracting the vascular skeleton by Zhang's thinning method; 7) performing skeleton branch pruning to remove noise line segments. The vascular skeleton in the left image is extracted by the above algorithm, and then the depth of the vascular skeleton is calculated by an improved sliding window algorithm with the information on the corresponding right image.

First, the designed imaging system is used to take NIR images of different parts of the body surface, including the back of the hands, forearms, and neck. The intermediate results of the vascular skeleton extraction algorithm (Fig. 5) and the three-dimensional reconstruction results of those body parts (Fig. 12) are analyzed. In addition, a bionic model is built with defibrinated sheep blood, beef slices, and pig skin (Fig. 8) to evaluate the consistency between real blood vessels and the vascular skeleton obtained by this system. The image processing results verify that the central line of the vascular skeleton extracted by this system can be consistent with the real blood vessel (Fig. 10), and the three-dimensional information on the obtained blood vessel is accurate (Fig. 11). For a higher processing speed of the vascular skeleton extraction algorithm, we rewrite the aforementioned algorithm to a parallel mode for GPU acceleration, then shoot 45 sets of left and right image pairs of different body parts, and record the processing speed of the original CPU algorithm and the GPU algorithm for a single image frame. The statistical results show that the GPU algorithm after acceleration takes an average of 64.40 ms per frame, which is 64% less than the original CPU algorithm. The improved sliding window matching algorithm takes an average of 105.32 ms per frame, and hence, the whole three-dimensional reconstruction process with GPU acceleration takes about 170 ms per frame.

The proposed three-dimensional imaging system for NIR subcutaneous blood vessels can effectively generate accurate three-dimensional images of subcutaneous blood vessels, which is suitable for various body parts such as the neck, forearm, and back of the hands and can also achieve good performance in thicker parts of surface tissue. The experimental results show that the extracted blood vessels are consistent with the real blood vessels, and the designed image processing algorithm takes an average total processing time of about 170 ms per frame. Hence, the expected reconstruction frame rate can reach 5 frame/s, which meets the requirements of intraoperative real-time modeling. To make this imaging system a module of the automatic puncture robot in the future, follow-up studies should include two aspects. The first one is to collect subcutaneous vascular patterns of people with different skin colors and different body fat content for the research on the adaptive adjustment method of skeleton extraction algorithm parameters. The second one is to build a theoretical model of light propagation in superficial biological tissues to correct the error of vascular depth estimation caused by light scattering.