Handheld intelligent blood vessel puncture system

Abstract

The invention discloses a handheld intelligent blood vessel puncture system which comprises an equipment physical layer, a signal processing layer and a software application layer. Wherein the equipment physical layer comprises an ultrasonic probe, a computing mainboard, a closed-loop driving system, a puncture mechanism, a power supply and a power supply management hardware module; the signal processing layer comprises a plurality of task modules for signal processing of image acquisition, motor control and data communication; the software application layer comprises a plurality of software function modules for image processing, AI algorithm, path planning and man-machine interaction; the system is used for obtaining a blood vessel image through real-time ultrasonic imaging after the ultrasonic probe is placed on a puncture part of a patient. The blood vessel image is analyzed through a built-in AI algorithm, and the blood vessel position and the blood vessel depth of the target blood vessel are automatically recognized and locked; planning an optimal puncture path according to the blood vessel position and the blood vessel depth; and the puncture mechanism is driven to complete puncture operation at an accurate angle and speed based on the optimal puncture path.

Description

Technical Field

[0001] This invention relates to the field of medical device optimization technology, and in particular to a handheld intelligent vascular puncture system. Background Technology

[0002] Vascular puncture is one of the most basic and crucial procedures in clinical medicine and emergency care, widely used in various scenarios such as intravenous infusion, blood collection, central venous catheter placement, and interventional therapy. However, the success rate and safety of this procedure are highly dependent on the operator's experience and skill. When dealing with special patients such as obese individuals, those with hypotension, shock, arteriosclerosis, edema, or infants, their blood vessels often exhibit characteristics such as being hidden, collapsed, thin, or highly mobile. This poses a significant challenge to traditional puncture methods that rely on visual observation and tactile sensation, resulting in a significantly higher failure rate on the first attempt. This not only increases patient suffering but may also lead to complications such as hematoma, nerve damage, and infection. Furthermore, in extreme conditions such as battlefield rescue, disaster relief, mobile clinics in remote areas, pre-hospital emergency care, and at night or in poor lighting conditions, insufficient ambient light, limited operating space, and a shortage of professional medical personnel further restrict the timeliness and reliability of traditional manual puncture, directly impacting patient safety.

[0003] To address these challenges, global efforts are underway to combine robotics with medical imaging to automate and automate vascular puncture procedures. Currently, both domestically and internationally, the field of robotic vascular puncture is in a phase of continuous development and exploration, but the overall technological level still lags significantly behind clinical goals.

[0004] I. Current Status of International Research and Analysis of Typical Products

[0005] The United States started earlier in this field, and relying on its advanced robotics and artificial intelligence foundation, it has produced several representative research prototypes and preliminary commercialized products. However, most of these systems still have significant limitations.

[0006] VascuLogic's Veebot venous blood collection robot is one of the early and well-known explorations. It integrates near-infrared 3D imaging technology and a 4-DOF mechanical control system, aiming to achieve the entire process from vein localization to automated puncture and blood collection. The system has shown potential in improving venous puncture success rates and reducing needlestick injuries, and has been validated in animal experiments. However, its core problem lies in its fixed design, resulting in a bulky system with poor portability, limiting its applicability to fixed blood collection rooms within hospitals. Furthermore, there is still no definitive public information regarding its approval by the U.S. Food and Drug Administration (FDA), thus limiting its clinical application and practicality.

[0007] Rivanna's Accuro 3S portable ultrasound system and its accompanying SpineNav-AI artificial intelligence image recognition software received FDA approval in July 2025, marking a significant advancement. This system focuses on the automatic identification and assisted localization of anatomical structures in axial nerve puncture scenarios (such as epidural anesthesia), and its lightweight hardware and advanced software algorithms embody some trends towards portability and intelligence. However, this product is essentially an "image navigation-assisted system" and does not integrate any automated puncture actuators; the puncture procedure still requires manual intervention by the physician, thus failing to achieve complete closed-loop automation from identification to execution.

[0008] Obvius Robotics' CERTA access system, developed for the high-risk procedure of central venous catheter (CVC) insertion, combines robotics and imaging technologies to improve the accuracy and safety of punctures. While the system has received FDA "Breakthrough Device" designation, its complex structure, typically including a handheld robotic unit, a separate imaging device, and an operating table, results in a large overall size and inconvenient deployment. Crucially, the puncture angle adjustment still requires manual intervention; the robot only handles the subsequent needle insertion. This is a typical "semi-automatic assisted" mode, failing to achieve fully automated path planning and posture adjustment, and thus falls short of truly "integrated intelligent puncture."

[0009] Mendaera's Focalist handheld ultrasound-guided interventional robot is another product that received FDA approval (July 2025), initially focusing on urological surgery. It integrates artificial intelligence, real-time imaging, and robotics, showcasing a highly integrated design. However, similar to the CERTA system, Focalist consists of multiple parts, including a handheld robotic unit, a separate imaging device, and an operating table. Its complex architecture and overall size and weight make it less than "portable," making it more suitable for fixed environments like operating rooms and less suitable for rapid deployment in emergency situations.

[0010] The AI-GUIDE system, jointly developed by MIT Lincoln Laboratory and Massachusetts General Hospital, represents cutting-edge scientific research. Designed for femoral artery vascular access puncture, the system interfaces with common portable ultrasound equipment, providing intelligent navigation from ultrasound probe positioning and puncture point determination to needle insertion. Animal experiments have shown its performance to be comparable to that of experienced healthcare professionals. However, to complete the entire procedure, AI-GUIDE requires additional portable ultrasound equipment and a commercial ultrasound panel. Its low system integration and fragmented components prevent it from being deployed as a standalone, integrated device for emergency use, limiting its potential application in extreme conditions.

[0011] II. Current Status and Challenges of Domestic Development

[0012] Existing achievements and products mainly focus on the relatively simple application scenario of peripheral venous blood collection and infusion.

[0013] A team from Harbin Institute of Technology, in collaboration with VascuLogic, has developed a prototype blood collection robot through Shanghai Bangcece Company. However, it remains in the prototype stage, and its technological approach and product form are heavily influenced by its foreign partner. Beijing Fuxi Nine Needles Company has completed the development of the FUXI intelligent infusion robot prototype and accumulated a large amount of ultrasound image data of the back of the hand's blood vessels for algorithm training, laying the foundation for intelligent recognition. However, no significant breakthroughs have been publicly disclosed in its productization process. Chengdu Kerry Medical Company's blood collection robot has obtained a type approval report and entered the clinical trial stage, making it one of the fastest-progressing teams. However, publicly available information indicates that the product also faces challenges related to size and ease of deployment.

[0014] Currently, the only intelligent blood collection robot to have obtained Class III medical device certification comes from Beijing Minas Surgical Robotics Co., Ltd. However, this product is primarily designed for fixed environments such as hospital laboratories and blood collection centers. Its complex structure, large size and weight make deployment and movement difficult, failing to meet the stringent requirements of rapid response, strong environmental adaptability, and extreme portability in high-risk, high-dynamic environments such as battlefields, disaster sites, and ambulances. Therefore, despite its certification, large-scale promotion and application outside of fixed hospital settings have not been observed.

[0015] III. Existing Technological Bottlenecks and Market Demand Gaps

[0016] A comprehensive analysis of the current situation both domestically and internationally reveals that current robotic vascular puncture technology generally faces the following core bottlenecks:

[0017] 1. Low system integration and lack of portability: Most systems consist of multiple modules such as separate robotic arms, imaging equipment, and control consoles. They are bulky and complex, making them unable to be quickly deployed and used on the go, and are seriously out of touch with the needs of emergency rescue, battlefield and other scenarios.

[0018] 2. Insufficient Level of Intelligence: Existing technologies mostly remain at the level of "image recognition + manual execution" or "semi-automatic assistance." Even those with image recognition capabilities often lack integrated intelligent path planning and fully automated closed-loop puncture control capabilities that can respond in real time to dynamic changes in blood vessels (such as displacement caused by respiratory movements, vascular collapse, etc.). In complex procedures like central venous puncture, which have extremely high requirements for depth, angle, and path, clinical products that achieve a fully automated closed-loop "recognition-planning-execution" process are almost non-existent.

[0019] 3. Limited application scenarios: Existing research and development and products are overly focused on peripheral venous blood collection, while there is a lack of mature and reliable automated solutions for deep vascular puncture scenarios such as central venous puncture, which are more technically challenging and riskier but crucial in critical care.

[0020] 4. Poor environmental adaptability: Most existing equipment is designed for the stable environment of hospitals and lacks the tolerance and adaptability to extreme external environments such as vibration, bumps, temperature changes, and poor lighting conditions.

[0021] Existing vascular puncture robot technologies worldwide have failed to effectively resolve the core contradiction of "lightweight, portable, intelligent, and integrated" solutions. There is an urgent market demand for a novel vascular puncture system that integrates high-precision medical image recognition, intelligent 3D path planning, and adaptive closed-loop puncture control, while also possessing excellent portability and environmental robustness, applicable to everything from routine wards to extreme emergency scenarios. Summary of the Invention

[0022] To address the problems existing in the prior art, the present invention provides the following technical solution: a handheld intelligent vascular puncture system that effectively solves the core contradiction of "lightweight, portable, intelligent, and integrated". It integrates high-precision medical image recognition, intelligent three-dimensional path planning, and adaptive closed-loop puncture control, while also possessing excellent portability and environmental robustness. It is a brand-new vascular puncture system applicable to everything from routine wards to extreme emergency scenarios.

[0023] This invention provides a handheld intelligent vascular puncture system, comprising: a device physical layer, a signal processing layer, and a software application layer; wherein:

[0024] The physical layer of the device includes an ultrasound probe, a computing motherboard, a closed-loop drive system, a puncture mechanism, a power supply, and a power management hardware module.

[0025] The signal processing layer includes multiple task modules for signal processing, such as image acquisition, motor control, and data communication.

[0026] The software application layer includes multiple software functional modules for image processing, AI algorithms, path planning, and human-computer interaction;

[0027] The system is used to acquire vascular images through real-time ultrasound imaging after the ultrasound probe is placed at the puncture site of the patient; the built-in AI algorithm analyzes the vascular images, automatically identifies and locks the location and depth of the target blood vessel; plans the optimal puncture path based on the location and depth of the blood vessel; and drives the puncture mechanism to complete the puncture operation at a precise angle and speed based on the optimal puncture path.

[0028] Preferably, the computing motherboard is used to perform image processing, AI algorithms, path planning, and human-computer interaction; the computing motherboard adopts a stacked configuration, with the top layer being the core board and the bottom layer being the computing motherboard baseboard. The core board serves as the core processing unit of the computing motherboard, and the computing motherboard baseboard provides display interfaces, communication interfaces, USB interfaces, clock interfaces, and fan interfaces; the computing motherboard employs a composite heat dissipation solution: a 1.5mm thick thermally conductive silicone pad is applied to the chip surface; an aluminum alloy heat sink is in close contact with the chip via four spring screws; the surface of the heat sink is anodized, with an emissivity of 0.8; four NTC thermistors are arranged below the NPU core area to monitor the temperature in real time;

[0029] The closed-loop drive system is used to precisely control the movement of the puncture mechanism. The puncture system uses a brushed DC motor as the puncture actuator, with position detection achieved through an encoder and a two-stage reducer. The drive board MCU of the closed-loop drive system drives four MOSFETs through two gate driver chips, collects phase current through sampling resistors, and collects phase voltage through voltage divider resistors. The drive board MCU performs drive control of the current loop, speed loop, and position loop based on the rotation angle feedback from the encoder, thereby realizing motor drive control. The drive board has a dual H-bridge architecture, including a main control MCU, gate drivers for dual-channel H-bridge drivers, power MOSFETs, current sampling components, and a buck converter providing 5V / 3A power. The drive board supports one or more control modes among position mode, speed mode, and torque mode. The closed-loop drive system controls the puncture mechanism based on a hierarchical control architecture control algorithm.

[0030] The power management hardware module manages the battery through a battery management scheme implemented by the power board. The battery is a medical-grade lithium-ion battery, connected in series and parallel. The power board collects the battery power in real time and collects the battery temperature via NTC to ensure battery safety. The power board has reverse connection protection, overcurrent protection, and overvoltage protection blocks, and provides power to the ultrasound probe and the computing motherboard to meet their power requirements. The battery management scheme is a PMIC power integrated management scheme, and uses the RK806-1 power management chip to achieve independent power supply for multiple voltage domains. The power management hardware module also includes a power tree with a hierarchical step-down architecture, including four power supply methods: core power supply, memory power supply, NPU power supply, and interface power supply.

[0031] Preferably, the puncture mechanism is a continuously adjustable robotic arm, employing a spatial orthogonal worm gear composite transmission structure, including: an angle adjustment mechanism based on worm gear transmission and a depth adjustment mechanism based on double-headed rectangular screw transmission; the ultrasound probe is a high-frequency linear array probe, including an ultrasound motherboard based on FPGA architecture, and performs vascular imaging based on parallel receiving beamforming technology.

[0032] Preferably, the signal processing layer includes:

[0033] The image acquisition and preprocessing module is implemented based on the Android Camera2 API. The preprocessing process includes: grayscale conversion for converting YUV to grayscale image, denoising for nonlocal mean denoising, enhancement for CLAHE contrast-limited histogram equalization, and normalization for normalizing the image to the 0-1 range.

[0034] The motor control module comprises four parts: a motor, sensors, a driver, and driver software. The sensors include an encoder, limit switches, and a reducer. The encoder is used for motor positioning, the limit switches are used to determine the extreme positions or origin of the movement, and the reducer is used to increase torque. The driver includes a central control unit, a drive gate, MOSFETs, and an EMC filter unit. The central control unit controls the MOSFETs through the drive gate, and outputs current through the MOSFETs to drive the motor. The driver software includes a parametric configuration engine, motion control algorithms, and real-time monitoring and diagnostic algorithms. The parametric configuration engine integrates a motor parameter database and a load characteristic model, supporting customized operating conditions. Parameters are input and dynamic characteristic curves are generated to assist in verifying the matching between the motor and the load. The motion control algorithm is used for high-precision position / velocity closed-loop control, including anti-disturbance algorithms for PID control, field-oriented control (FOC), and sliding mode control (SMC). The three-phase current is decomposed into torque and flux components through coordinate transformation to perform torque and flux decoupling control and precise motor control. The puncture control is implemented based on FreeRTOS real-time tasks, including: sending blood vessel parameters to the motor drive board, the motor drive board calculating the puncture depth and puncture angle, and reading displacement or status feedback in real time to determine whether the puncture process is successful. Thus, the status synchronization and error detection of the puncture process are achieved through event groups or semaphores.

[0035] Data communication module: The user obtains and provides feedback on the position, force, and angle information during the puncture process, and issues an alarm or determines that the puncture is complete based on the information.

[0036] Preferably, the software application layer includes software functional modules for image processing, AI algorithms, path planning, and human-computer interaction, wherein the AI ​​algorithm is a method for locating and classifying veins and arteries; the method for locating and classifying veins and arteries includes a vessel intelligent detection algorithm, a vein-artery classification algorithm, a high-precision dynamic vessel localization algorithm, and a target vessel determination and puncture path planning algorithm; wherein:

[0037] The intelligent vascular detection algorithm is used to automatically identify vascular structures in ultrasound images based on deep learning technology.

[0038] The vein-artery classification algorithm is used to accurately distinguish between veins and arteries, avoiding accidental punctures;

[0039] The high-precision dynamic blood vessel localization algorithm is used to achieve accurate segmentation of blood vessel boundaries based on detection. The high-precision dynamic blood vessel localization algorithm is implemented by a blood vessel accurate fitting method based on random walk and graph cut deformation model.

[0040] The software application layer is based on the localization and classification methods of veins and arteries. After ultrasound acquisition of vascular images, the intelligent vascular detection algorithm calls the trained vascular detection model to perform global detection on all blood vessels in the input image. The detection results of the global detection are deeply integrated with the arteriovenous classification method of color ultrasound. Through multimodal information interaction and verification, high-precision localization and type identification of blood vessels in the image are achieved, providing basic data for target vessel selection and puncture path calculation.

[0041] Preferably, the intelligent blood vessel detection algorithm is based on a lightweight target detection network for blood vessel detection and modifies the lightweight target detection network, including: using a C2f module for the backbone; using an anchor-free + decoupled-head detection head; using a combination of classification BCE Loss, regression CIoU, and VFL as the loss function; changing the box matching strategy from static matching to Task-Aligned Assigner matching; disabling Mosaic operation in the last 10 epochs; and increasing the total number of training epochs from 300 to 500; simultaneously, introducing the CBAM attention mechanism into the backbone of the lightweight target detection network's attention mechanism; the loss function of the lightweight target detection network includes: replacing CIoU loss with GIoU loss, where: GIoU is defined as: GIoU = IoU - |C\(A∪B)| / |C|; the GIoU loss function... The formula is: L_{GIoU}=1-GIoU; where: A and B are the predicted bounding box and the ground truth bounding box, C is the smallest closed rectangle surrounding A and B, IoU=|A∩B| / |A∪B|, the value range of IoU is [0,1], and the value range of GIoU is [-1,1]; The lightweight object detection network is subjected to lightweight processing, which includes: replacing standard convolution with depthwise separable convolution, channel pruning, 8-bit integer quantization, and compressing the model size of the lightweight processing from 240MB to 45MB.

[0042] Preferably, the arteriovenous vessel classification algorithm is implemented based on a multimodal fusion arteriovenous vessel classification framework, which includes:

[0043] (1) The first modal feature extraction unit is used to extract blood vessel regions based on B-ultrasound images. It uses YOLOv8 to detect blood vessel regions and outputs bounding boxes and confidence scores to obtain the first modal features based on the blood vessel region images detected by the target.

[0044] (2) The second modality feature extraction unit is used to generate preliminary classification templates for arteries and veins based on hemodynamic features and to obtain second modality features based on color Doppler spectrum analysis; wherein the generation of preliminary classification templates based on hemodynamic features includes: blood flow direction detection based on centripetal blood flow in veins and centrifugal blood flow in arteries; velocity feature extraction based on high arterial velocity and pulsation and low venous velocity and stability; and spectral waveform analysis based on pulsating waveforms in arteries and continuous waveforms in veins;

[0045] (3) Modal feature fusion and network structure setting unit, used to fuse the first modal feature and the second modal feature, set the network structure of the corresponding deep classification network, train the deep classification network to form EfficientViT as the base network, fuse the first modal feature and the second modal feature based on the target detection blood vessel region image into a multimodal input, perform high-precision classification of blood vessel type through the deep network, and output the confidence score of automatic classification;

[0046] (4) Pressure feedback decision correction unit, used to start the pressure feedback process when the confidence score of automatic classification is <0.85 or inconsistent with the spectrum classification result. The pressure feedback process includes: linearly increasing the pressure at a rate of 5 mmHg / s; monitoring vascular deformation by calculating the change in vascular cross-sectional area in real time; and calculating the deformation index: VDI=(ΔA / ΔP)×f_pulse, where: ΔA / ΔP represents the rate of change of area under unit pressure change; f_pulse represents the pulsation frequency, and the following discrimination rules are set: artery: VDI>threshold, and changes significantly with pressure; and / or vein: VDI≈0, uniform deformation;

[0047] The deep classification network based on Efficient ViT, trained from a deep classification network, includes: unsupervised training on a large open-source database followed by fine-tuning on a target fine-classification dataset; wherein the unsupervised training is performed on multiple large-scale visual image datasets including ImageNet, COCO, and proprietary datasets, and fine-tuning is performed on an interface target fine-classification dataset.

[0048] Preferably, the high-precision dynamic blood vessel localization algorithm includes: designing an energy function based on the fusion of random walk and graph cut algorithms to design feature information; preprocessing the blood vessel image based on K-means clustering and Hough transform; and verifying the blood vessel image segmentation results using the level set method.

[0049] Preferably, the target vessel identification and puncture path planning algorithm includes determining a target vessel selection strategy and planning a puncture path; wherein,

[0050] The strategy for determining target vessel selection includes:

[0051] Select the target vein or artery based on the working status: If the current working mode is venipuncture, select an available vein that is at an appropriate distance from the skin as the target vessel, and if there are multiple available vessels, they will form a candidate set; if the current working mode is arterial puncture, select the arterial vessel that is closest to the skin as the target vessel, and if there are multiple available vessels, they will form a candidate set.

[0052] The optimal target vessel is selected based on the location and correlation of the blood vessels in the diagram, including: selecting the thickest vessel for puncture based on geometric calculations; ensuring there are no other blood vessels between the punctured vessel and the skin; and selecting venous target vessels where there is no overlap of blood vessels; selecting vessels as close to the skin as possible while meeting the thickness requirements; and selecting arterial target vessels where there are no other blood vessels between the punctured vessel and the skin and no overlap of blood vessels.

[0053] The planning objectives for the puncture path include: the insertion point into the skin, the angle between the needle and the skin, and the insertion length.

[0054] For venipuncture path planning, the center of the target vessel and the center of the needle direction are connected by a line. It is then determined whether there are other veins in the path. If not, the planning is successful, and the needle angle and insertion length are calculated. If there are other veins, it is determined whether they meet the puncture requirements. If so, the target vessel is replaced, and the puncture path planning is restarted. If not, the process returns to selecting another vein as the target. If the requirements cannot be met, an "inappropriate position" message is displayed, and the operator makes adjustments.

[0055] For arterial puncture path planning, the process includes: connecting the target vessel center with the needle direction selection center, and determining whether there are other veins or arteries in the connecting path; if not, the planning is successful, and the needle angle and insertion length can be calculated; if there are, it indicates that safe puncture is not possible, and the fixed position of the needle is readjusted. If the needle has a lateral movement control mechanism, the needle position is adjusted laterally by a fixed step size for searching; if the needle does not have a lateral movement control mechanism, the position is not suitable, and the operator needs to make adjustments.

[0056] Preferably, the software application layer includes:

[0057] Image recognition and vessel identification module: Vessel recognition is implemented based on TensorFlow Lite, using deep learning algorithms to identify vascular structures in images, extracting key parameters such as vessel diameter, depth, and location, and determining whether the target vessel is located in the center of the image and whether it meets the puncture conditions. The image recognition process is combined with AI algorithm processing, and model inference relies on a lightweight AI framework deployed on an ARM embedded platform to achieve real-time recognition.

[0058] Human-computer interaction and decision support module: The interactive interface is implemented based on Android custom View. After identifying a puncturable blood vessel, the system prompts the user through the interface "There is a suitable blood vessel, do you want to puncture?" and controls the "puncture" button to become operable. With the graphical interface as the core, the interactive logic is built using an embedded GUI framework to assist the user in deciding whether to perform the puncture. After the user confirms, the puncture control stage begins.

[0059] Anomaly Handling and Reset Module: Anomaly handling is implemented using a finite state machine. When a puncture fails or is completed, the system guides the user into a reset process, controls the motor to return to its initial position, and waits for a reset completion signal. If the reset is successful, the system prompts "Searching for blood vessel" and re-enters the image recognition process, forming a closed loop.

[0060] The puncture system provided by this invention has the following beneficial effects:

[0061] 1. A more flexible handheld intelligent vascular puncture system was designed using a robotic arm based on a spatial orthogonal worm gear composite transmission architecture, which is reflected in:

[0062] (1) The puncture arm angle adjustment mechanism adopts a small volume high power density hollow cup motor, combined with a planetary reducer and a self-locking spatial orthogonal worm gear composite transmission architecture. Based on engineering principles, the angle adjustment mechanism is modeled in 3D using SolidWorks to demonstrate the geometric dimensions, manufacturability and functional requirements of the model. Static simulation (load ≥ 0.3 N·m) is performed using ANSYS to predict the strength and stability of the structure. The self-locking critical angle of the spatial orthogonal worm gear is calculated based on the Coulomb friction model to construct stable and reliable self-locking conditions and ensure self-locking performance. An incremental magnetic encoder is used to monitor the angle position and perform repeatability positioning accuracy tests. A miniature, mechanical position detection sensor is used to zero the position after each angle adjustment to avoid the accumulation of cumulative errors.

[0063] (2) The puncture arm mechanism adopts a small volume high power density hollow cup motor, which is combined with a planetary reducer and a double-headed rectangular screw (1.2mm pitch) to form a drive device. In addition, an integrated slider screw nut, combined with a linear rectangular multi-faceted auxiliary guide rail, jointly constructs the puncture system architecture. Based on engineering principles, a 3D model of the puncture mechanism is created using SolidWorks to demonstrate the model's geometric dimensions, manufacturability, and functional requirements. Static simulation is performed using ANSYS to predict the strength and stability of the structure. Utilizing the principle of converting the continuity of mechanical motion into the discreteness of electrical signals, an incremental magnetic encoder is used to monitor the puncture depth. A miniature, mechanical position detection sensor is used to reset the encoder's initial position after each puncture to avoid the accumulation of cumulative errors.

[0064] (3) The needle clamp and needle guide are made of medical polycarbonate (Sabic Lexan 943A) and injection molded. They have passed ISO 10993-5 biocompatibility certification. Based on engineering principles, the needle clamp and needle guide are modeled in 3D using SolidWorks to demonstrate the geometric dimensions, injection molding performance and functional requirements of the model. Static simulation is performed using ANSYS to predict the strength and stability of the structure.

[0065] (4) Continuous stability of the puncture mechanism: To ensure the continuous stability of the puncture mechanism, real-time image analysis technology can be used to identify its vibration state and accurately quantify the vibration amplitude. On this basis, the acceleration curve is planned by the fifth-order polynomial interpolation (Quintic Spline) algorithm to achieve smooth control of the motion process. At the same time, a risk warning mechanism is established to avoid operational risks caused by vibration in a timely manner.

[0066] 2. Vein and artery detection, segmentation, and classification techniques based on deep learning algorithms

[0067] Breaking through the bottleneck of traditional puncture techniques' heavy reliance on medical experience, a complete technical system has been built around high-precision, low-latency intelligent puncture algorithms, forming a multi-dimensional collaborative innovation. In vascular detection, taking into account the computing power limitations of domestic embedded platforms, a lightweight architecture and optimization strategies are innovatively adopted. Through specific backbone networks, detection head design, and loss function combinations, accuracy is improved while computational overhead is reduced, achieving efficient real-time detection. Facing the challenge of vascular localization in ultrasound images, a multi-level localization mechanism is constructed by integrating random walk algorithms and graph cut deformation models. Adaptive threshold segmentation addresses noise sensitivity issues, and shape constraint terms and dynamic weighting mechanisms are introduced to effectively handle problems caused by blurred vascular boundaries and deformation, improving segmentation accuracy. For arteriovenous classification, a two-layer system of "multimodal fusion + pressure feedback" is innovatively constructed. Initial classification is achieved through lightweight networks and spectral feature fusion. When confidence is insufficient, decision correction is achieved through linear pressure detection and related dynamic calculations to ensure classification accuracy. For puncture path planning, a target vessel evaluation model integrating multiple factors is innovatively established. A path search algorithm is developed based on puncture needle characteristics to achieve optimal planning in complex scenarios, supporting autonomous high-precision robot operation. Through in-depth algorithm optimization, a fully intelligent solution is formed, reducing reliance on medical experience, adapting to domestic hardware platforms, and providing core technical support for robot autonomous puncture.

[0068] 3. Dynamic vessel segmentation technology based on a three-level collaborative framework of "probabilistic modeling - deformation optimization - closed-loop verification"

[0069] Addressing the three core challenges of intraoperative ultrasound vascular segmentation: blurred vessel boundaries due to grayscale noise; contour breakage caused by dynamic displacement; and missegmentation of complex morphologies such as bifurcation structures, this study breaks through the traditional single-stage segmentation paradigm and pioneers a three-level collaborative framework of "probabilistic modeling-deformation optimization-closed-loop verification," achieving a progressive improvement from coarse segmentation to sub-pixel accuracy. Random walk-driven initial contour generation: An adaptive brightness threshold segmentation method based on ultrasound features is proposed. By constructing a vessel probability map (introducing multi-scale gradient weights), noise interference is suppressed, significantly improving the confidence of the initial contour compared to the traditional thresholding method. Dual-temporal map cut deformation optimization: An innovative dynamic weight energy function is designed, where the shape constraint term uses radial distance mapping to calculate pixel offsets in real time. The dynamic weight coefficient γ is exponentially related to the displacement amplitude, ensuring topological continuity during vessel deformation. Expert-guided closed-loop verification: A dual-index feedback mechanism of level set algorithm and Dice-Hausdorff is established, and online learning is carried out through expert-annotated datasets (multiple actual ultrasound images), which significantly improves the segmentation accuracy of bifurcation parts.

[0070] 4. Localized deployment acceleration technology for algorithms based on multi-core scheduling and cache optimization

[0071] Regarding deployment on domestically produced platforms, this algorithm system emphasizes innovative localized design based on domestically produced computing motherboards. Targeting the computing power characteristics of domestically produced embedded computing motherboards, it achieves deep adaptation between the algorithm and domestic hardware through lightweight network architecture (such as specific backbone networks and detection head design) and targeted optimization strategies (loss function combination, improved box matching methods, parallel computing, etc.). This reduces computational overhead while ensuring detection accuracy and real-time response performance, ensuring that the entire process of algorithm modules, including blood vessel detection, dynamic localization, arteriovenous classification, and puncture path planning, can run efficiently on domestically produced computing motherboards. This forms an intelligent puncture solution adapted to the domestic hardware ecosystem, providing core support for the localization of robotic autonomous puncture technology.

[0072] 5. System Integration Design

[0073] (1) Multi-module interface compatibility: Differences in interface protocols and data formats between different hardware modules may lead to data transmission delays or losses. Key technologies: Adopting OPC UA based on the IEC 62541 standard as the software interface standard to solve the "information silo" problem; designing an adaptive interface conversion module to be compatible with different levels through signal conditioning circuits, and designing redundant communication links in accordance with ISO11898-2; based on the principle of "interface abstraction", encapsulating hardware interfaces through an intermediate layer to achieve unified API calls. Solution: Formulating test specifications based on the "equivalence class partitioning method" to cover all interface combination scenarios and ensure that the data transmission packet loss rate is ≤0.01%.

[0074] (2) Real-time data interaction and processing: The algorithm recognition results need to be converted into robotic arm motion commands in real time. A delay of more than 50ms may lead to puncture deviation. Key technologies: A task scheduling mechanism is built based on RTOS (such as VxWorks). The RMS algorithm is used to sort high-frequency tasks according to the principle of "the shorter the task cycle, the higher the priority". TSN technology is introduced to achieve μs-level time synchronization in accordance with IEEE 802.1AS. The critical data channel is guaranteed by the flow shaping algorithm. Solution: Build a HIL simulation platform to simulate 1000 puncture scenarios and ensure that the core data transmission delay is ≤20ms.

[0075] (3) Overall Electromagnetic Compatibility (EMC) Optimization: Electromagnetic interference from internal components may affect the accuracy of ultrasound images and must comply with IEC 60601-1-2 standards. Key Technologies: Based on the EMC principles of "shielding, filtering, and grounding," Faraday cage shielding (attenuation ≥40dB) is used for sensitive components, a common-mode choke is connected in series at the motor power supply, and a star grounding network is used to avoid ground loop interference. Solution: Pre-compatibility testing is conducted according to IEC 60601-1-2 to ensure that radiated emissions in the 30MHz~1GHz band are ≤54dBμV / m.

[0076] (4) Motion control and algorithm coordination: The puncture point coordinates output by the algorithm need to be accurately converted into the robot arm trajectory. Mechanical errors and vibrations may cause positioning deviations. Key technologies: Establish a kinematic model based on the DH parameter method to ensure that the coordinate transformation error is ≤0.1mm; use a cubic polynomial interpolation algorithm to plan the trajectory and introduce the Jacobian matrix to realize spatial transformation; combine MPC to handle nonlinear errors and improve the closed-loop accuracy based on the "feedback control principle". Solution: Test with 100 sets of virtual blood vessel models to ensure that the three-dimensional deviation between the actual and target puncture points is ≤0.5mm.

[0077] (5) Modular integration of software system: Software modules need to work together and support functional expansion.

[0078] Key technologies: Employing a microservice architecture to decompose modules into independent units; introducing DDS middleware conforming to the OMG DDS standard to build a real-time data bus (transmission latency ≤1ms); using the OSGi framework to achieve dynamic module deployment and support on-the-fly updates. Solution: Automated interface testing to simulate 1000 module start-ups and shutdowns, ensuring an MTBF ≥ 1000 hours.

[0079] (6) Safety Redundancy and Fault Tolerance: Sensor failure or algorithm misjudgment may lead to vascular damage, requiring a highly reliable emergency mechanism. Key Technology: SIL 3 level (probability of dangerous failure per hour) is set according to IEC 61508 standard. The system employs a "two-out-of-two" voting mechanism to process critical data, triggering an alarm when the deviation exceeds 0.3mm; it also builds an FTA model to optimize emergency logic (shutdown within 0.5 seconds). Solution: Conduct 1000 fault injection tests to ensure emergency response time ≤ 0.3 seconds and shutdown deviation ≤ 0.2mm.

[0080] (7) Optimization of domestic computing motherboard resources: Differences in computing power distribution and memory bandwidth of domestic motherboards may lead to low algorithm efficiency. Key technologies: Quantization perception training based on NPU characteristics is adopted to quantize the model into 8-bit integers (accuracy loss ≤1%); task parallelism is optimized according to Amdahl's Law to make computing power utilization ≥85%; dynamic resource scheduling algorithm is developed to avoid "computing power bottleneck". Solution: Test 1000 clinical data on domestic motherboards to ensure that the accuracy of blood vessel recognition is ≥98% and the processing time per case is ≤500ms. Attached Figure Description

[0081] Figure 1 This is a schematic diagram illustrating the principle and structure of the handheld intelligent vascular puncture system described in this invention.

[0082] Figure 2This is a schematic diagram of the computing motherboard structure described in this invention.

[0083] Figure 3 This is a schematic diagram of the drive board structure of the closed-loop drive system described in this invention.

[0084] Figure 4 This is a schematic diagram of the method for locating and classifying veins and arteries according to the present invention.

[0085] Figure 5 This is a schematic diagram of the YOLO v8 network structure described in this invention.

[0086] Figure 6 This is a schematic diagram of the CBAM attention mechanism described in this invention.

[0087] Figure 7 This is a schematic diagram illustrating the use of labeled blood vessel images for blood vessel detection training annotation as described in this invention.

[0088] Figure 8 This is a schematic diagram illustrating the classification results of veins and arteries according to the present invention.

[0089] Figure 9 This is a schematic diagram of the overall structure of the ViT algorithm described in this invention.

[0090] Figure 10 This is a schematic diagram of the DeiT network structure described in this invention. Detailed Implementation

[0091] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0092] Establishing vascular access promptly is crucial for trauma care, ensuring the timely infusion of emergency medications, analgesics, and fluids. While peripheral intravenous catheters play a key role in fluid resuscitation, central venous catheters not only provide stable and reliable vascular access but also enable large-volume fluid resuscitation and facilitate invasive endovascular treatments. Whether at the scene of the injury, during medical transport, in the emergency room, or at a military frontline surgical base, the ability to establish vascular access promptly directly determines the timing of emergency care. For cases with the longest transport times, timely establishment of central venous access often yields the greatest results. However, such procedures typically require experienced medical personnel. In a hospital setting, intensive care physicians are often far from the scene of the injury. Major obstacles to obtaining central venous access on-site include a lack of advanced ultrasound image interpretation skills and insufficient proficiency in image-guided vascular puncture needle insertion. The handheld intelligent vascular puncture system described in this embodiment enables medical personnel with limited vascular access skills to master the establishment of central venous access in pre-hospital emergency scenarios. Establishing central venous access is more challenging than peripheral venous access because the vessels are located deeper, making direct observation and manual palpation difficult. Major arterial structures are often adjacent to the target vein, increasing the risk of accidental arterial injury. Although ultrasound-guided central venous placement has significant value, it is rarely performed in pre-hospital emergency settings due to the extensive training and practice required for both ultrasound image interpretation and catheter placement. Standard Serdinger technique catheter placement involves several steps: ① puncture of the target vessel, ② placement of a coaxial guidewire, ③ needle removal, ④ repeated dilation of the puncture site as needed, ⑤ catheter placement via the guidewire, and ⑥ guidewire removal. The initial puncture is the most skill-intensive and prone to failure step. Furthermore, successful guidewire insertion and needle withdrawal almost guarantee the success of subsequent steps and eliminate the risk of needle injury. This embodiment of the handheld intelligent vascular puncture system helps medical personnel with varying levels of training to obtain central venous access. This handheld robot is, firstly, a portable device using a domestically developed computing platform. Through AI-assisted ultrasound image interpretation, it locates the optimal puncture position, allowing medical personnel to insert a precise and automated puncture needle after pressing the puncture button. Secondly, phantom and clinical trial results demonstrate that this handheld intelligent vascular puncture system enables both expert and novice medical personnel to quickly locate blood vessels without manually interpreting ultrasound images. Thirdly, clinical trials show that after less than 30 minutes of training and trial use, medical personnel operating this handheld intelligent vascular puncture system achieve the same accuracy in performing the Seldinger technique as experienced medical personnel using manual puncture, with improved speed. The core innovation of this product lies in its ability to achieve full-process AI navigation and puncture of ultrasound-guided interventional needle puncture without requiring manual interpretation of ultrasound images or advanced puncture skills.

[0093] like Figure 1 As shown, this invention provides a handheld intelligent vascular puncture system, a portable medical device integrating medical imaging, artificial intelligence, and robotics technologies. Figure 1 As shown, the system adopts an integrated design, integrating functions such as ultrasound imaging, intelligent recognition, path planning and automatic puncture, which can assist medical staff to quickly and accurately complete vascular puncture operations under various conditions.

[0094] The handheld intelligent vascular puncture system is technically divided into three layers: the physical layer of the device, the signal processing layer, and the software application layer; among which:

[0095] The physical layer of the device includes hardware modules such as an ultrasonic probe, a computing motherboard, a puncture mechanism, and a power management system.

[0096] The signal processing layer is used for signal processing tasks such as image acquisition, motor control, and data communication.

[0097] The software application layer includes software functional modules such as image processing, AI algorithms, path planning, and human-computer interaction.

[0098] The system's workflow is as follows: Medical staff place the device's ultrasound probe at the patient's puncture site, and the system acquires vascular images through real-time ultrasound imaging; the built-in AI algorithm analyzes the images, automatically identifies and locks onto the target blood vessel; the system plans the optimal puncture path based on the blood vessel's location and depth; after the operator presses the puncture button, the system drives the puncture mechanism to complete the puncture operation at a precise angle and speed.

[0099] I. The physical layer of the handheld intelligent vascular puncture system includes:

[0100] (a) such as Figure 2 As shown, the computing motherboard is the core processing unit of the system, used to perform image processing, AI algorithm calculations, and device control tasks. Based on the reliability and safety requirements of medical devices, this invention adopts a domestically developed and controllable computing platform. The computing motherboard uses a domestically designed CPU. Besides handling the device's business logic, the CPU's primary function is the processing of venous blood vessel algorithms. It must meet the demands of powerful AI computation while minimizing power consumption to ensure performance and achieve a handheld miniaturization solution. Simultaneously, the computing motherboard must support various peripherals such as UART, RS485, SPI, eDP, LVDS, and RTC.

[0101] 1. Hardware Architecture of the Computing Motherboard: The computing motherboard adopts a stacked configuration, with the core board at the top and the computing motherboard baseboard at the bottom. The core board integrates an RK3588 CPU, 8GB LPDDR4X memory, 128GB eMMC flash memory, and SPIFlash, serving as the core processing unit of the computing motherboard. The computing motherboard baseboard provides abundant interface resources, including display interfaces, communication interfaces, USB interfaces, clock interfaces, and fan interfaces.

[0102] As a preferred implementation, the display interface supports three interface standards: MIPI, eDP, and LVDS, depending on the selected screen interface. The communication interface provides multiple industrial communication interfaces such as RS232, RS485, and CAN to ensure reliable connection with various peripheral modules. The RTC clock is used to record the device time, ensuring that the system maintains correct time information even when power is off.

[0103] 2. Calculate the motherboard's power management scheme.

[0104] Battery management comprises a power board and batteries, using medical-grade lithium-ion batteries connected in series and parallel. The power board implements the battery management solution, monitoring battery charge in real time and battery temperature via NTC to ensure safe operation. The power board features reverse connection protection, overcurrent protection, and overvoltage protection, and supplies power to the ultrasound, computing motherboard, and driver board, meeting their power requirements. The computing motherboard uses a PMIC power integrated management solution, employing the RK806-1 power management chip to achieve independent power supply for multiple voltage domains. The power tree design adopts a hierarchical buck architecture, including the following key components:

[0105] (1) Core power supply: A 4-phase Buck converter is used to provide a stable power supply of 0.9V / 12A to the CPU core. Each phase is equipped with a 3.3μH power inductor and a 470μF ceramic capacitor, and the ripple is controlled within ±2%.

[0106] (2) Memory power supply: A 2-phase Buck converter is used to provide 1.1V / 8A power to LPDDR4X, equipped with a 2.2μH power inductor and a 220μF ceramic capacitor;

[0107] (3) NPU power supply: Independent 1.2V / 6A power supply circuit, equipped with low ESR tantalum capacitor array to ensure AI operation stability;

[0108] (4) Interface power supply: Provides multiple power outputs of 3.3V, 5V and 12V, each equipped with overcurrent protection and soft start circuit.

[0109] 3. Calculate the signal integrity design scheme for the motherboard.

[0110] To ensure high-speed signal transmission quality, the PCB design adopts an 8-layer board structure, with the following layer stack-up scheme:

[0111] Top layer: Signal layer, impedance controlled 50Ω single-ended / 100Ω differential; Layer 2: GND layer, complete ground plane; Layer 3: Signal layer, high-speed signal routing; Layer 4: Power layer, core power partition; Layer 5: Signal layer, medium and low speed signals; Layer 6: GND layer; Layer 7: Power layer, interface power; and Bottom layer: Signal layer, peripheral interfaces.

[0112] High-speed interfaces such as eDP and MIPI DSI use stripline wiring, strictly maintaining equal length (±5mil) differential pairs and impedance matching. Clock signals are grounded to avoid crosstalk.

[0113] 4. Calculate the motherboard's thermal management scheme.

[0114] To address the high power consumption of the RK3588 chip, the computing motherboard employs a composite heat dissipation solution: a 1.5mm thick thermally conductive silicone pad is applied to the chip surface; an aluminum alloy heat sink is tightly connected to the chip via four spring screws; the heat sink surface is anodized with an emissivity of 0.8; and four 0402 NTC thermistors are placed below the NPU core area to monitor the temperature in real time.

[0115] (ii) Closed-loop drive system

[0116] The closed-loop drive system is responsible for precisely controlling the movement of the puncture mechanism, ensuring the accuracy and safety of the puncture operation. The palm vein puncture system uses a brushed DC motor as the puncture actuator, with position detection achieved through an encoder. A two-stage reducer is used to ensure efficiency while also increasing torque. The drive board MCU drives four MOSFETs through two gate drive chips, collects phase current through sampling resistors, and collects phase voltage through voltage divider resistors. Based on the rotation angle feedback from the encoder, the MCU realizes drive control of the current loop, speed loop, and position loop, achieving motor drive control and enabling puncture accuracy to reach sub-millimeter level. Finally, the drive board communicates with the mainboard via RS485 communication.

[0117] 1. The closed-loop drive system uses a brushed DC motor as the drive element, with the following specific parameters: Rated voltage: 12VDC; No-load speed: 8500±15% RPM; Rated torque: 35.3 mN·m; Stall torque: 272 mN·m; Rated current: 1.08A; No-load current: 0.12A; Motor constant: 25.3mN·m / √W; Rotor inertia: 11.3 g·cm².

[0118] The motor is equipped with a 1024-line incremental magnetic encoder, which achieves a resolution of 4096 pulses per revolution through 4x frequency technology, and the angle detection accuracy reaches 0.088 degrees.

[0119] 2. Drive circuit scheme for closed-loop drive system

[0120] like Figure 3 As shown, the driver board adopts a dual H-bridge architecture, and the specific circuit composition is as follows: Main control MCU: STM32F405RGT6, operating frequency 168MHz; Gate drive: DRV8702 dual-channel H-bridge driver; Power MOSFET: AON7400, withstand voltage 30V, on-resistance 6.5mΩ; Current sampling: 0.1Ω / 1% precision resistor with INA240 current sense amplifier; Power management: TPS54331 buck converter provides 5V / 3A power supply.

[0121] The driver board supports multiple control modes, including: position mode: PID position control based on encoder feedback; speed mode: FOC field orientation control with speed accuracy ±1%; torque mode: torque control based on current loop with resolution of 0.1 mN·m.

[0122] 3. Control algorithm of closed-loop drive system

[0123] The motion control algorithm adopts a hierarchical control architecture:

[0124] (1) Position loop PID controller:

[0125] P[k]=Kp*e[k]+Ki*∑e[i]+Kd*(e[k]-e[k-1])

[0126] Among them, e[k]=θ_target-θ_actual

[0127] Kp=2.5, Ki=0.8, Kd=0.3

[0128] (2) Speed ​​loop FOC control: The three-phase current is decomposed into torque component (Iq) and flux linkage component (Id) through Clarke / Park transformation:

[0129] Iα=Iu

[0130] Iβ=(Iu+2*Iv) / √3

[0131] Id=Iα*cosθ+Iβ*sinθ

[0132] Iq=-Iα*sinθ+Iβ*cosθ

[0133] It adopts SVPWM modulation, with a switching frequency of 20kHz and a dead time of 200ns.

[0134] (3) Anti-disturbance algorithm: Integrating sliding mode control (SMC) to handle sudden load changes:

[0135] s = e + λ·∫e·dt

[0136] u = u_eq + K·sat(s / Φ)

[0137] Where, λ = 15, K = 8, Φ = 0.1

[0138] (III) Power supply system design

[0139] The power supply system provides stable and reliable power supply for the entire device, meeting the strict requirements of medical equipment for safety and reliability.

[0140] 1. Battery pack: Four 18650 lithium-ion batteries are connected in series, with a built-in protection IC (DW01 + in cooperation with 8205A MOSFET) and thermal protection through a 60°C temperature control switch.

[0141] The functions of the battery management system (BMS) include: overcharge protection (4.28 ± 0.025V), over-discharge protection: (2.90 ± 0.08V), over-current protection (adjustable from 8 - 15A), short-circuit protection (response time < 1ms), and battery charge measurement (coulomb meter IC MAX17048, accuracy ±1%).

[0142] 2. The power distribution network provides input over-voltage protection, reverse connection protection, soft start circuit, and EMC filtering (common mode choke + π-type filter). At the same time, to extend the battery life, the system adopts multi-level power consumption management.

[0143] (IV) Continuously adjustable robotic arm

[0144] The continuously adjustable robotic arm is the core actuator for precise puncture. It adopts a spatial orthogonal worm and gear compound transmission structure, including: an angle adjustment mechanism using worm and gear transmission and a depth adjustment mechanism using a double-headed rectangular lead screw transmission. Among them: the worm is single-headed, with a module of 0.5 and a lead angle of 3°; the worm gear has 30 teeth, with a reduction ratio of 30:1. The worm is made of 45 steel after quenching, and the worm gear is made of PA66 + 30% GF, with a friction coefficient μ = 0.12, satisfying γ < arctan(μ) = 6.8° to form self-locking. The drive assembly includes: a coreless DC motor, a planetary reducer, an encoder, and a limit switch. To ensure the smoothness of the puncture process, a fifth-order polynomial interpolation is used to plan the acceleration curve, as shown in the following formula:

[0145]

[0146] Boundary conditions: The velocities and accelerations at the starting and ending moments are both zero.

[0147] (v) Ultrasound module

[0148] The ultrasound module is responsible for vascular imaging, providing real-time image data to the system. The ultrasound module uses a high-frequency linear array probe with 128 array elements. The ultrasound motherboard is based on an FPGA architecture, with 128 channels of parallel digital beamforming. The ultrasound module employs parallel receiving beamforming technology for imaging, achieving depth adaptation through dynamic focusing, suppressing fundamental noise through pulse inversion technology, and improving contrast through spatial recombination.

[0149] (vi) The equipment's appearance adopts an ergonomic design to ensure ease of operation and user comfort. It features a geometric interlacing style, with functional areas divided into a display and control area and a puncture execution area. The outer shell is made of ABS+PC alloy, possessing excellent mechanical strength and biocompatibility. The puncture area uses a replaceable sterile sheath and medical-grade TPU, 0.1mm thick, with sound wave attenuation <0.5dB. It is fixed with a snap-on installation, with a replacement time of <30 seconds.

[0150] II. The signal processing layer is used for signal processing tasks such as image acquisition, motor control, and data communication, including:

[0151] (I) Image Acquisition and Preprocessing Module: After system startup, the system acquires real-time images of the patient's vascular region by connecting to ultrasound equipment. The key to this stage lies in image quality preprocessing, such as noise reduction, contrast enhancement, and image size standardization, to ensure the image meets the input requirements of subsequent recognition algorithms. High frame rate and low latency data transmission are achieved through the embedded platform's image interface.

[0152] Image acquisition is implemented based on the Android Camera2 API; the image preprocessing process includes: grayscale conversion (converting YUV to grayscale image), denoising (non-local mean denoising, h=10), enhancement (CLAHE contrast-limited histogram equalization) and normalization (normalizing the image to the 0-1 range).

[0153] (II) Motor Control Module: This module mainly consists of four parts: motor, sensors, drivers, and driver software. The motor is a brushed DC motor, which allows for stable operation at low speeds and strong anti-interference capabilities. Sensors include encoders, limit switches, and reducers. The encoder is used for motor positioning, the limit switches are used to determine the extreme positions or origin of the movement, and the reducer is used to increase torque for smoother movement. The driver includes a central control unit, drive gate, MOSFETs, and an EMC filter unit. The central control unit controls the MOSFETs through the drive gate, and the MOSFETs output current to drive the motor. The driver software includes a parametric configuration engine, motion control algorithms, real-time monitoring and diagnostics, and other functions. The parametric configuration engine integrates a motor parameter database (rated power, torque curve) and a load characteristic model, supports custom operating condition parameter input (speed, acceleration, inertia ratio), and generates dynamic characteristic curves (such as speed-torque relationships) to assist in verifying the matching between the motor and the load. The motion control algorithm achieves high-precision position / velocity closed-loop control, supports point-to-point positioning, trajectory interpolation, and force-position hybrid control, and integrates anti-disturbance algorithms such as PID control, field-oriented control (FOC), and sliding mode control (SMC). Through coordinate transformation (Clarke / Park), the three-phase current is decomposed into torque components (IQ) and flux linkage components (ID), achieving decoupled control of torque and flux linkage for precise motor control. The puncture control is based on a FreeRTOS real-time task. The system sends vascular parameters to the motor drive board, which calculates the puncture depth and angle, and reads displacement or status feedback in real time to determine whether the puncture process is successful. The core technologies of this module are motor motion control and FreeRTOS task scheduling mechanism, which realizes status synchronization and error detection of the puncture process through event groups or semaphores.

[0154] (III) Data Communication Module: The system needs to achieve high-speed and stable communication with the puncture execution device, provide real-time feedback on key information such as position, force, and angle during the puncture process, and have corresponding alarm and puncture completion determination mechanisms. The core technology of this module is the design of a serial communication protocol.

[0155] III. The software application layer includes software functional modules such as image processing, AI algorithms, path planning, and human-computer interaction. This software application layer is implemented based on methods for locating and classifying veins and arteries, and includes:

[0156] The flowchart of the method for locating and classifying veins and arteries implemented in the software application layer is as follows: Figure 4As shown, the implementation principle includes: after ultrasound acquisition of vascular images, the intelligent vascular detection algorithm first calls the trained vascular detection model to perform global detection of all blood vessels in the input image; then, the detection results are deeply integrated with the color ultrasound arteriovenous classification method, and through multimodal information interaction verification, high-precision localization and type identification of blood vessels in the image are achieved, providing accurate basic data for subsequent target vessel selection and puncture path calculation. In actual clinical applications, the system innovatively introduces an operator-assisted pressure-based vein confirmation step, further avoiding the risk of misjudgment through a human-machine collaborative dual verification mechanism, and comprehensively ensuring the safety and reliability of the puncture operation.

[0157] (a) Intelligent vascular detection algorithm: used to automatically identify vascular structures in ultrasound images based on deep learning technology.

[0158] 1. Framework selection for intelligent blood vessel detection algorithm: Blood vessel detection based on lightweight target detection network.

[0159] To enable the deployment of portable devices and ensure puncture efficiency, a lightweight deep network is used for blood vessel detection and localization. For example, YOLOv8, YOLOv11, NanoDet, and NanoDe-Plus detection algorithms can be used for blood vessel detection. This embodiment uses YOLOv8 as the basic detection framework, and its network structure is as follows: Figure 5 As shown, improvements and upgrades have been made based on the first 7 versions of the YOLO detection algorithm. The improvements include: using the C2f module for the backbone; using anchor-free + decoupled-head for the detection head; using a combination of classification BCE Loss, regression CIoU and VFL for the loss function; changing the box matching strategy from static matching to Task-Aligned Assigner matching; disabling the Mosaic operation in the last 10 epochs; and increasing the total number of training epochs from 300 to 500.

[0160] 2. The training data preprocessing workflow includes:

[0161] Image normalization: Normalizing pixel values ​​to the range [0,1];

[0162] Data augmentation includes random rotation (±15°), scaling (0.8-1.2x), and brightness adjustment (±20%).

[0163] And Mosaic enhancement: turn it off in the last 10 epochs of training to avoid over-enhancement.

[0164] 3. Improvements to the attention mechanism include: introducing the CBAM attention mechanism in Backbone.

[0165] Because a lightweight backbone feature extraction network is used, the number of parameters is significantly reduced. However, the algorithm's ability to extract contextual and spatial information is somewhat diminished. Therefore, an attention mechanism can be introduced to optimize detection accuracy. For example, CBAM (convolutional block attention module) can be used. It is a lightweight convolutional neural network attention mechanism designed to improve the network's representational power by focusing on important features in both channel and spatial dimensions. CBAM consists of two modules, with the structure as follows: Figure 6 As shown: Channel Attention Module (CAM) and Spatial Attention Module (SAM). CAM enhances the network's feature representation by learning the importance weights for each channel. SAM enhances the network's feature representation by learning the importance weights for each spatial location. CBAM concatenates CAM and SAM, processing the input feature map sequentially. CBAM can focus on important features in both channel and spatial dimensions, thereby improving the model's representational power and performance. This study introduces the CBAM attention mechanism at different layers, enhancing the model's feature extraction capability at multiple scales, and effectively improving the algorithm's detection accuracy with only a small increase in the number of parameters.

[0166] Channel attention module: Mc(F) = σ(MLP(AvgPool(F)) + MLP(MaxPool(F))) = σ(W1(W0(F_avg)) + W1(W0(F_max)))

[0167] Spatial attention module: Ms(F)=σ(f ^{7×7}([AvgPool(F);MaxPool(F)]))=σ(Conv^{7×7}([F_avg;F_max]))

[0168] Final output:

[0169] 4. Loss function optimization includes: replacing CIoU loss with GIoU loss to address the issue of unclear aspect ratio description. Specifically:

[0170] Definition of GIoU: GIoU = IoU - |C\(A∪B)| / |C|;

[0171] The GIoU loss function is: L_{GIoU} = 1 - GIoU;

[0172] Where: A and B are the predicted bounding box and the ground truth bounding box, C is the smallest closed rectangle enclosing A and B, IoU=|A∩B| / |A∪B|, the value range of IoU is [0,1], and the value range of GIoU is [-1,1].

[0173] Taking YOLOv8 as an example, the bounding box loss functions used are DFL Loss (distribution focal loss) and CIoU Loss (complete intersection over union loss). CIoU requires significant computational overhead and is unclear in its description of the intersection-over-union ratio (IoU) and does not consider the balance between the learning difficulty of samples. This embodiment uses a more updated and suitable loss function for the blood vessel detection task, such as GIoU instead of CIoU. GIoU loss introduces the minimum closed rectangle enclosing the two bounding boxes on top of IoU loss, solving the problem of zero gradient when two targets (A and B) have no intersection. After using the GIoU loss function, the number of algorithm parameters is slightly reduced, the loss function is more stable, and the accuracy of flame and smoke detection is also improved.

[0174] 5. Lightweight Model

[0175] To adapt to embedded platforms, the model was lightweighted, including replacing standard convolution with depthwise separable convolution, channel pruning, reducing the number of parameters by 30%, 8-bit integer quantization, accuracy loss of <1%, and compressing the model size from 240MB to 45MB.

[0176] (ii) Vein and artery classification algorithm, used to accurately distinguish between veins and arteries and avoid accidental puncture.

[0177] 1. The multimodal fusion framework for classifying arterial and venous vessels consists of three main parts:

[0178] (1) Extraction of blood vessel regions based on B-ultrasound images, using YOLOv8 to detect blood vessel regions, and output bounding boxes and confidence scores;

[0179] (2) Color Doppler spectrum analysis, including:

[0180] Blood flow direction detection: venous blood flow towards the heart, arterial blood flow towards the heart;

[0181] Flow velocity feature extraction: Arterial flow velocity is high and pulsating, while venous flow velocity is low and stable;

[0182] Spectral waveform analysis: Arteries show pulsating waveforms, while veins show continuous waveforms;

[0183] (3) Deep classification network

[0184] The network structure, which integrates features from two modalities, is as follows:

[0185] Input: [224×224×3 image, 32-dimensional spectral features] -> Convolutional layer: Conv7×7, 64, stride=2 -> Max pooling: 3×3, stride=2 -> Residual block ×3: [Conv1×1, 64Conv3×3, 64Conv1×1, 256]×3 -> Feature fusion: Connect the spectral features to the fully connected layer -> Classification head: FC(512) → FC(256) → FC(2) -> Output: Probability of vein / artery.

[0186] (4) Stress feedback decision correction module, used to start the stress feedback process when the automatic classification confidence level is <0.85:

[0187] (A) Linear pressurization: Pressure increases at a rate of 5 mmHg / s

[0188] (B) Vascular deformation monitoring: Real-time calculation of changes in vascular cross-sectional area

[0189] (C) Deformation index calculation: VDI=(ΔA / ΔP)×f_pulse

[0190] Where: ΔA / ΔP: rate of change of area under unit pressure change; f_pulse: pulsation frequency;

[0191] Discrimination rules: Arteries: VDI > threshold and changes significantly with pressure; and / or veins: VDI ≈ 0 and deformation is uniform.

[0192] In this embodiment, the multimodal fusion-based venous and arterial vessel classification framework is adopted because venous and arterial punctures are critical operations in clinical medicine, and mispunctures can lead to serious complications. Through the intelligent positioning framework of multimodal fusion and the pressure feedback decision correction module, high-precision classification of vessel types is achieved, ensuring the correctness of puncture target selection, and achieving the following three technical effects:

[0193] Improved classification accuracy: The accuracy of blood vessel type classification is improved by fusing a lightweight target detection network with color Doppler spectral analysis;

[0194] Conflict Decision Optimization: Develop a decision correction module based on pressure feedback to resolve the conflict between automatic classification and spectrum analysis;

[0195] Safe puncture guarantee: By quantifying the vascular deformation index (VDI), accurate identification of arteries and veins is achieved, minimizing the risk of accidental puncture.

[0196] If a vascular classification method based on a single ultrasound image is used, it is necessary to acquire ultrasound images of multiple body parts that may be subject to puncture, such as the back of the hand, forearm, and upper arm. This requires acquiring images from multiple individuals and adjusting the acquisition process multiple times. The locations of blood vessels in the acquired images must be manually labeled, marking veins and arteries as shown. Figure 7 As shown, distinguishing between veins and arteries in a single static short-axis ultrasound image frame is quite difficult, therefore it needs to be combined with high-precision vein and artery classification methods based on color Doppler or spectral Doppler.

[0197] Blood vessel classification based on color Doppler or spectral Doppler ultrasound is widely used and relatively mature, providing high-precision differentiation between veins and arteries. In color Doppler ultrasound images, veins are generally represented in blue and arteries in red, such as... Figure 8 As shown. However, the accuracy of the blood vessel range is relatively poor. The blue and red areas generally have irregular edges, and some non-vascular areas are detected in addition to real blood vessels. Combining the results of deep learning for blood vessel localization can screen out real veins and arteries.

[0198] Therefore, this invention employs a scheme that integrates ultrasound images and Doppler spectral analysis results, followed by classification using a depth model. Pressure feedback is then used to further determine the vessel type. The multimodal fusion classification architecture includes:

[0199] (1) Extraction of blood vessel regions based on B-ultrasound images, using YOLOv8 to detect blood vessel regions, and output bounding boxes and confidence scores;

[0200] (2) Color Doppler spectrum analysis: Combine hemodynamic characteristics (such as flow velocity and direction) to generate a preliminary classification template to distinguish between veins (low flow velocity, unidirectional flow) and arteries (high flow velocity, pulsatile flow);

[0201] (3) Deep classification network fusion: The image of the blood vessel region detected by the target is fused with the spectral analysis features into a multimodal input, and the blood vessel type is classified with high accuracy through a deep network, and the confidence score is output.

[0202] (4) Pressure feedback decision correction module: When the automatic classification confidence is lower than the threshold (e.g., <0.85) or inconsistent with the spectrum classification result, the pressure feedback process is triggered.

[0203] For deep classification networks, to achieve high-precision classification of veins and arteries and to identify the specific states they represent, this embodiment uses various fine-classification network models and selects a suitable model for vessel classification. These mainly include the following types of fine-classification templates.

[0204] The fine-grained classification model ResNet is an integrated image classification ResNet model. Image classification based on MovileNet has wide applications and mature deployment schemes. The lightweight ResNet can be used for various target classification tasks in the embodiments of this invention.

[0205] MovileNet, a fine-grained classification model, is an integrated image classification model based on MobileNetV2. Image classification based on MovileNet has wide applications and mature deployment schemes. The lightweight MovileNet can be used for various target classification tasks in the embodiments of this invention.

[0206] The ViT fine-grained classification model directly applies the standard Transformer structure to images with minimal modifications to the overall image classification process. Specifically, the ViT algorithm divides the entire image into small image patches, then feeds the linear embedding sequences of these patches as input to the Transformer network, and finally trains the image classification system using supervised learning. The overall structure of the ViT algorithm is as follows. Figure 9 As shown, the model is first pre-trained on a large-scale dataset such as JFT-300 or ImageNet-21K, and then transferred to other medium or smaller-scale datasets. This embodiment uses some lightweight ViT network structures for integration to achieve efficient computation. Specifically, it uses algorithms such as Microsoft's EfficientVit (which simplifies the model through group attention) and MIT's EfficientVit (which simplifies the model using depthwise convolution and linear attention) for integration.

[0207] The fine-grained classification model DeiT learns on a general-sized dataset and introduces distillation methods into the training of VIT. It adopts a teacher-student training strategy, proposing token-based distillation, and uses a convolutional network as the teacher network for distillation, achieving better results than using a transformer architecture as the teacher. The DeiT network structure is as follows: Figure 10 As shown, the introduced distillation token is mainly used for distillation learning during network training. In the self-attention layers, it continuously interacts with the class token and image patches, allowing the model to learn from the output of the teacher network and also serving as a supplement to the class token. Regarding the loss function, a distillation loss is added to the cross-entropy loss.

[0208] Ultimately, Efficient ViT was chosen as the base network, achieving a balance between accuracy and speed.

[0209] 2. Self-supervised learning methods address the problem of insufficient sample size.

[0210] Because the number of samples in each class is limited, especially when adding new types, the number of images directly usable for training is generally not large. Collecting and labeling a large number of images requires a significant amount of manpower and time. Training deep neural networks is more difficult when there are few training samples. Therefore, this invention proposes first developing an unsupervised network based on a large open-source database, and then fine-tuning the training on a target fine-classification database to ensure the network's robustness and generalization performance.

[0211] By drawing inspiration from the supervised learning method MoCo, we achieve unsupervised pre-training of the target fine-classification network and supervised fine-tuning of the target fine-classification tool.

[0212] During the unsupervised training phase, a training image is used... Perform two types of random data augmentation to obtain and They consist of two encoders and Encode the vector to output the vector. , and , The loss function is then:

[0213]

[0214] in, The InfoNCE loss function is used, namely:

[0215]

[0216] in, The same image of q after passing through The output of is a sample of the same type as q. For q, it is a non-class sample. For query encoder, It is a momentum encoder. After the update, momentum was updated. The method is .

[0217] Unsupervised training was performed on multiple large-scale visual image datasets such as ImageNet, COCO, and our own dataset, and fine-tuning was performed on a fine-classification dataset of interface objects.

[0218] (iii) A high-precision dynamic blood vessel localization algorithm is used to achieve accurate segmentation of blood vessel boundaries based on detection. The high-precision dynamic blood vessel localization algorithm is implemented by a blood vessel accurate fitting method based on random walk and graph cut deformation model.

[0219] After blood vessel detection and vein / artery classification, high-precision blood vessel localization, i.e., blood vessel segmentation, can be performed. This embodiment employs a blood vessel precise fitting method and system based on random walk and graph cut deformation models. It integrates random walk and graph cut algorithms, designs energy functions based on feature information, preprocesses blood vessel images based on K-means clustering and Hough transform, and uses the level set method to verify the blood vessel image segmentation results, eliminating missegmentation phenomena, thereby achieving precise blood vessel segmentation and precise localization.

[0220] Image segmentation is the process and technique of dividing an image into non-overlapping regions with distinct characteristics and extracting the target of interest. It is a crucial step from image processing to image analysis, and its segmentation quality directly affects the quality of subsequent image analysis, understanding, and recognition. Based on whether human intervention is involved, image segmentation algorithms can be divided into unsupervised segmentation algorithms and supervised segmentation algorithms. The graph cut algorithm, proposed by Boykov et al., is one of the most widely used supervised segmentation methods (an interactive segmentation method). In the field of image segmentation, graph cut algorithms have many excellent characteristics, such as allowing user intervention, high efficiency, strong robustness, high segmentation accuracy, and N-dimensional segmentation capability. It is widely used in computer vision for foreground / background segmentation and stereo vision. Fully automatic segmentation algorithms often yield unsatisfactory results in complex images, while semi-automatic segmentation algorithms, by providing user intervention opportunities, improve segmentation accuracy and are therefore widely used. The graph cut algorithm can be described as a combined optimization process of binary (foreground and background) labels for each pixel in the image. First, the user selects a subset of pixels from the target and background regions to be segmented as target seed points and background seed points, respectively. These labeled points become crucial information guiding the computer in segmenting the target and background regions; this information is called hard constraints. Besides hard constraints, another important constraint in graph cut algorithms is soft constraints, which include regional information and boundary information of the image. Hard and soft constraints together form the energy function in graph cut algorithms. The core theory of graph cut algorithms is to find the optimal segmentation result by minimizing this energy function. The energy optimization process utilizes weighted graph mapping and network flow theory, transforming the energy minimization problem into solving the minimum cut problem of a weighted graph. The goal of graph cut is to find the most suitable label for all pixels in the image (e.g., labeling the foreground as 1 and the background as 0), which can be achieved by minimizing the energy function. First, an undirected weighted graph is constructed, consisting of vertices and weighted edges. Vertices are divided into two categories: pixel vertices of the image and two terminal vertices. Based on the connections between different vertices, the edges of a weighted graph can be divided into two categories: connections between terminal vertices and pixels, and connections between adjacent pixels. The goal of image segmentation is to find a line that cuts the connection between the background and the target in the image; this line is called a "cut". If the sum of the weights of all the cut edges is minimized, then this cut is called a minimum cut, meaning the energy function is minimized. The Ford-Funken theory shows that the maximum flow problem in a network is equivalent to the minimum cut problem.Therefore, the minimum cut of the weighted graph, i.e. the final segmentation result, can be obtained through the maximum flow algorithm.

[0221] In a preferred embodiment, the high-precision dynamic blood vessel localization algorithm is implemented using a blood vessel accurate fitting method based on random walk and graph cut deformation models, including:

[0222] 1. Based on the initial segmentation using random walks, the pixel classification probability is obtained by solving the Laplace equation:

[0223] L·x=-r

[0224] Where: L is the graph Laplacian matrix, x is the probability vector of unknown nodes, and r is the boundary condition vector.

[0225] Specific steps:

[0226] (1) Constructing a graph structure: Each pixel is a node, and adjacent pixels are connected by edges;

[0227] (2) Setting seed points: the inside of the blood vessel is the foreground seed, and the surrounding area is the background seed;

[0228] (3) Calculate weights: Set edge weights based on image gradient.

[0229] w_{ij}=exp(-β·(I_i-I_j)²)

[0230] The conjugate gradient method is used to solve the linear system, thereby obtaining the pixel probability distribution.

[0231] 2. Based on the random walk results, a graph cut algorithm is used for fine-grained segmentation. The energy function is defined as follows:

[0232]

[0233] Specific forms of each energy term:

[0234] (1) Data item energy:

[0235] R_u(1)=-ln(P(u∈O))

[0236] R_u(0)=-ln(P(u∈B))

[0237] The foreground probability P(u∈O) is calculated based on a Gaussian mixture model.

[0238] (2) Boundary term energy:

[0239] B_{uv}=exp(-|I_u-I_v|² / 2σ_g²) / dist(u,v)

[0240] (3) Shape constraint energy:

[0241] S_u(1)=d(u,x_O) / r_O

[0242] S_u(0)=1-d(u,x_O) / r_O

[0243] (3) The segmentation results are verified using the level set method:

[0244]

[0245] Through iterative evolution, the contour becomes smoother and more continuous.

[0246] (iv) Algorithm for target vessel identification and puncture path planning

[0247] 1. Determine the target vessel selection strategy

[0248] Typically, multiple blood vessels of different types may be detected in an image. It is necessary to first determine the extent of the target blood vessel, and then select the target vessel.

[0249] The first step is to select the arterial and venous target based on the work status:

[0250] If the current working mode is venipuncture, select an available vein (displayed as a circle with the largest radius in the image) that is appropriately distanced from the skin as the target vessel. If multiple available vessels are available, they form a candidate set. Appropriate distance from the skin here mainly depends on the type of vessel to be punctured. Currently, venipuncture is commonly performed on the basilic vein. The great saphenous vein or small saphenous vein, which are generally not punctured very close to the skin.

[0251] Since the basilic vein and the great and small saphenous veins have significantly different distribution locations on ultrasound images, a simple threshold range can be used. For example, if the coordinates on the image are greater than a certain threshold (this threshold is related to the puncture site, the characteristics of the color Doppler ultrasound equipment, and parameter adjustments, and can be preset after determining the equipment and site), the great and small saphenous veins can be excluded.

[0252] If the current working mode is arterial puncture, select the artery closest to the skin (the image shows the artery closest to the top as the target vessel) as the target vessel. If there are multiple available vessels, they will form a candidate set.

[0253] The second step is to select the optimal target vessel based on the location and correlation of the blood vessels in the image.

[0254] Target vein selection for vein puncture generally involves choosing the thickest vessel for the procedure; ensuring there are no other vessels between the punctured vessel and the skin; and avoiding any overlap of vessels. These selection strategies can be determined through geometric calculations and comparisons.

[0255] Artery target vessel selection: with the required diameter, it should be as close to the skin as possible; there should be no other vessels between the punctured vessel and the skin; and there should be no overlapping of vessels.

[0256] In this embodiment, different selection criteria are used depending on the type of puncture:

[0257] (1) Venous puncture: The basilic vein should be selected first, with a diameter >2.5mm and a depth in the range of 3-15mm. There should be no other blood vessels between the vein and the skin, and the bifurcation of the blood vessels should be avoided.

[0258] (2) Arterial puncture: Select a superficial artery with a diameter >2.0 mm, obvious pulsation, easy to identify, and avoid important nerves.

[0259] 2. Plan the puncture path

[0260] The steel needle is a rigid object and is considered as a line for planning. The objectives of puncture needle path planning include: the insertion point into the skin, the angle between the steel needle and the skin, and the insertion length (to determine the depth of penetration).

[0261] First, planning the venous puncture route.

[0262] Connect the center of the target vessel to the center of the needle direction and determine if there are other veins in the path of the connection.

[0263] If not, the planning is successful; simply calculate the needle angle and insertion length.

[0264] If so, determine whether the vein in the path meets the puncture requirements (the vein and its thickness meet the puncture requirements) and replace it with the target vein to replan the puncture path. If not, return to select other veins as the target vein.

[0265] If the requirements cannot be met, a message will be displayed indicating that the location is unsuitable, and the operator will need to make adjustments.

[0266] Second, arterial puncture path planning

[0267] Because arteries are located deeper in the body, they often have veins between them and the skin, and these veins must be avoided during puncture. Generally, the length of the needle passing through the body should be minimized during the puncture process. A planning strategy was designed to search for feasible paths from the vertical direction to both sides:

[0268] Connect the center of the target blood vessel to the center of the needle direction and determine if there are other veins or arteries in the path of the connection.

[0269] If not, the planning is successful; simply calculate the needle angle and insertion length.

[0270] If so, a message should be displayed indicating that safe puncture is not possible, and the needle's fixed position needs to be readjusted. For needles with a lateral movement control mechanism, the needle position should be adjusted laterally in fixed steps during the search. For needles without a lateral movement control mechanism, the position is considered unsuitable, and the operator needs to make adjustments.

[0271] In a preferred embodiment, the puncture path planning considers multiple constraints:

[0272] (1) Geometric constraints: Puncture angle: 15-45° range; Distance between the needle insertion point and the blood vessel: Minimize the puncture path length; Avoid important tissues such as bones and nerves.

[0273] (2) Safety constraints: There must be no other blood vessels in the path; it must not pass through tendons or nerve bundles; avoid areas with arterial pulsation.

[0274] (3) Mechanical constraints: Limit the range of motion of the robotic arm; avoid singular configurations; meet the torque output requirements.

[0275] (iv) Application software functional modules

[0276] The system software adopts a modular design and mainly includes three functional modules.

[0277] 1. Image Recognition and Vessel Identification Module: Vessel recognition is implemented based on TensorFlow Lite. Deep learning algorithms are used to identify vascular structures in images, extracting key parameters such as vessel diameter, depth, and location to determine if the target vessel is located in the center of the image and whether it meets the puncture requirements. The image recognition process incorporates AI algorithm processing, while model inference relies on a lightweight AI framework deployed on an ARM embedded platform for real-time recognition.

[0278] 2. Human-Computer Interaction and Decision Support Module: The interactive interface is implemented based on a custom Android View. After identifying a puncturable blood vessel, the system prompts the user via the interface, "A suitable blood vessel is available, do you wish to puncture?" and makes the "puncture" button operable. This stage is centered on a graphical interface, using an embedded GUI framework to construct the interactive logic and assist the user in deciding whether to perform the puncture. After user confirmation, the puncture control stage begins.

[0279] 3. Anomaly Handling and Reset Module: Anomaly handling is implemented using a finite state machine. When a puncture fails or is completed, the system guides the user into a reset process, controlling the motor to return to its initial position and waiting for a reset completion signal. If the reset is successful, the system prompts "Searching for blood vessel" and re-enters the image recognition process, forming a closed loop. This part focuses on ensuring system state consistency and avoiding abnormal actions caused by motor position deviations.

[0280] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A handheld intelligent vascular puncture system, characterized in that, include: The device consists of a physical layer, a signal processing layer, and a software application layer; among which: The physical layer of the device includes an ultrasound probe, a computing motherboard, a closed-loop drive system, a puncture mechanism, a power supply, and a power management hardware module. The signal processing layer includes multiple task modules for signal processing, such as image acquisition, motor control, and data communication. The software application layer includes multiple software functional modules for image processing, AI algorithms, path planning, and human-computer interaction; The system is used to acquire vascular images through real-time ultrasound imaging after the ultrasound probe is placed at the puncture site of the patient; the built-in AI algorithm analyzes the vascular images, automatically identifies and locks the location and depth of the target blood vessel; plans the optimal puncture path based on the location and depth of the blood vessel; and drives the puncture mechanism to complete the puncture operation at a precise angle and speed based on the optimal puncture path.

2. The handheld intelligent vascular puncture system according to claim 1, characterized in that, The computing motherboard is used to perform image processing, AI algorithms, path planning, and human-computer interaction. The computing motherboard adopts a stacked configuration, with the top layer being the core board and the bottom layer being the computing motherboard baseboard. The core board serves as the core processing unit of the computing motherboard, and the computing motherboard baseboard provides display interfaces, communication interfaces, USB interfaces, clock interfaces, and fan interfaces. The computing motherboard employs a composite heat dissipation solution: a 1.5mm thick thermally conductive silicone pad is applied to the chip surface. The aluminum alloy heat sink is in close contact with the chip via four spring screws; the surface of the heat sink is anodized and has an emissivity of 0.8; four NTC thermistors are placed below the NPU core area to monitor the temperature in real time. The closed-loop drive system is used to precisely control the movement of the puncture mechanism. The puncture system uses a brushed DC motor as the puncture actuator, with position detection achieved through an encoder and a two-stage reducer. The drive board MCU of the closed-loop drive system drives four MOSFETs through two gate driver chips, collects phase current through sampling resistors, and collects phase voltage through voltage divider resistors. The drive board MCU performs drive control of the current loop, speed loop, and position loop based on the rotation angle feedback from the encoder, thereby realizing motor drive control. The drive board has a dual H-bridge architecture, including a main control MCU, gate drivers for dual-channel H-bridge drivers, power MOSFETs, current sampling components, and a buck converter providing 5V / 3A power. The drive board supports one or more control modes among position mode, speed mode, and torque mode. The closed-loop drive system controls the puncture mechanism based on a hierarchical control architecture control algorithm. The power management hardware module manages the battery through a battery management scheme implemented by the power board. The battery is a medical-grade lithium-ion battery, connected in series and parallel. The power board collects the battery power in real time and collects the battery temperature via NTC to ensure battery safety. The power board has reverse connection protection, overcurrent protection, and overvoltage protection blocks, and provides power to the ultrasound probe and the computing motherboard to meet their power requirements. The battery management scheme is a PMIC power integrated management scheme, and uses the RK806-1 power management chip to achieve independent power supply for multiple voltage domains. The power management hardware module also includes a power tree with a hierarchical step-down architecture, including four power supply methods: core power supply, memory power supply, NPU power supply, and interface power supply.

3. The handheld intelligent vascular puncture system according to claim 2, characterized in that, The puncture mechanism is a continuously adjustable robotic arm, employing a spatial orthogonal worm gear composite transmission structure, including: an angle adjustment mechanism based on worm gear transmission and a depth adjustment mechanism based on double-headed rectangular screw transmission; the ultrasound probe is a high-frequency linear array probe, including an ultrasound motherboard based on FPGA architecture, and performs vascular imaging based on parallel receiving beamforming technology.

4. The handheld intelligent vascular puncture system according to claim 3, characterized in that, The signal processing layer includes: The image acquisition and preprocessing module is implemented based on the Android Camera2 API. The preprocessing process includes: grayscale conversion for converting YUV to grayscale image, denoising for nonlocal mean denoising, enhancement for CLAHE contrast-limited histogram equalization, and normalization for normalizing the image to the 0-1 range. The motor control module comprises four parts: a motor, sensors, a driver, and driver software. The sensors include an encoder, limit switches, and a reducer. The encoder is used for motor positioning, the limit switches are used to determine the extreme positions or origin of the movement, and the reducer is used to increase torque. The driver includes a central control unit, a drive gate, MOSFETs, and an EMC filter unit. The central control unit controls the MOSFETs through the drive gate, and outputs current through the MOSFETs to drive the motor. The driver software includes a parametric configuration engine, motion control algorithms, and real-time monitoring and diagnostic algorithms. The parametric configuration engine integrates a motor parameter database and a load characteristic model, supporting customized operating conditions. Parameters are input and dynamic characteristic curves are generated to assist in verifying the matching between the motor and the load. The motion control algorithm is used for high-precision position / velocity closed-loop control, including anti-disturbance algorithms for PID control, field-oriented control (FOC), and sliding mode control (SMC). The three-phase current is decomposed into torque and flux components through coordinate transformation to perform torque and flux decoupling control and precise motor control. The puncture control is implemented based on FreeRTOS real-time tasks, including: sending blood vessel parameters to the motor drive board, the motor drive board calculating the puncture depth and puncture angle, and reading displacement or status feedback in real time to determine whether the puncture process is successful. Thus, the status synchronization and error detection of the puncture process are achieved through event groups or semaphores. Data communication module: The user obtains and provides feedback on the position, force, and angle information during the puncture process, and issues an alarm or determines that the puncture is complete based on the information.

5. The handheld intelligent vascular puncture system according to claim 4, characterized in that, The software application layer includes software functional modules for image processing, AI algorithms, path planning, and human-computer interaction. The AI ​​algorithms are methods for locating and classifying veins and arteries. These methods include intelligent vessel detection algorithms, vein-arterial vessel classification algorithms, high-precision dynamic vessel localization algorithms, and target vessel identification and puncture path planning algorithms. The intelligent vascular detection algorithm is used to automatically identify vascular structures in ultrasound images based on deep learning technology. The vein-artery classification algorithm is used to accurately distinguish between veins and arteries, avoiding accidental punctures; The high-precision dynamic blood vessel localization algorithm is used to achieve accurate segmentation of blood vessel boundaries based on detection. The high-precision dynamic blood vessel localization algorithm is implemented by a blood vessel accurate fitting method based on random walk and graph cut deformation model. The software application layer is based on the localization and classification methods of veins and arteries. After ultrasound acquisition of vascular images, the intelligent vascular detection algorithm calls the trained vascular detection model to perform global detection on all blood vessels in the input image. The detection results of the global detection are deeply integrated with the arteriovenous classification method of color ultrasound. Through multimodal information interaction and verification, high-precision localization and type identification of blood vessels in the image are achieved, providing basic data for target vessel selection and puncture path calculation.

6. The handheld intelligent vascular puncture system according to claim 5, characterized in that, The proposed intelligent blood vessel detection algorithm is based on a lightweight object detection network and modifies that network, including: using a C2f module for the backbone; employing an anchor-free + decoupled-head detection head; using a combination of classification BCE Loss, regression CIoU, and VFL as the loss function; changing the bounding box matching strategy from static matching to Task-AlignedAssigner matching; disabling Mosaic in the last 10 epochs; and increasing the total number of training epochs from 300 to 500. Furthermore, a CBAM attention mechanism is introduced into the backbone of the lightweight object detection network. The loss function of the lightweight object detection network includes replacing CIoU loss with GIoU loss, where GIoU is defined as: GIoU = IoU - |C\(A∪B)|. The GIoU loss function is: L_{GIoU}=1-GIoU; where A and B are the predicted and ground truth boxes, C is the smallest closed rectangle enclosing A and B, IoU=|A∩B| / |A∪B|, the range of IoU is [0,1], and the range of GIoU is [-1,1]; the lightweight object detection network is subjected to lightweight processing, which includes: replacing standard convolution with depthwise separable convolution, channel pruning, 8-bit integer quantization, and compressing the model size from 240MB to 45MB.

7. A handheld intelligent vascular puncture system according to claim 6, characterized in that, The arteriovenous vessel classification algorithm is implemented based on a multimodal fusion arteriovenous vessel classification framework, which includes: (1) The first modal feature extraction unit is used to extract blood vessel regions based on B-ultrasound images. It uses YOLOv8 to detect blood vessel regions and outputs bounding boxes and confidence scores to obtain the first modal features based on the blood vessel region images detected by the target. (2) The second modality feature extraction unit is used to generate preliminary classification templates for arteries and veins based on hemodynamic features and to obtain second modality features based on color Doppler spectrum analysis; wherein the generation of preliminary classification templates based on hemodynamic features includes: blood flow direction detection based on centripetal blood flow in veins and centrifugal blood flow in arteries; velocity feature extraction based on high arterial velocity and pulsation and low venous velocity and stability; and spectral waveform analysis based on pulsating waveforms in arteries and continuous waveforms in veins; (3) Modal feature fusion and network structure setting unit, used to fuse the first modal feature and the second modal feature, set the network structure of the corresponding deep classification network, train the deep classification network based on the deep classification network to form Efficient ViT as the base network, fuse the first modal feature and the second modal feature based on the target detection blood vessel region image into a multimodal input, perform high-precision classification of blood vessel type through the deep network, and output the confidence score of automatic classification; (4) Pressure feedback decision correction unit, used to start the pressure feedback process when the confidence score of automatic classification is <0.85 or inconsistent with the spectrum classification result. The pressure feedback process includes: linearly increasing the pressure at a rate of 5 mmHg / s; monitoring vascular deformation by calculating the change in vascular cross-sectional area in real time; and calculating the deformation index: VDI=(ΔA / ΔP)×f_pulse, where: ΔA / ΔP represents the rate of change of area under unit pressure change; f_pulse represents the pulsation frequency, and the following discrimination rules are set: artery: VDI>threshold, and changes significantly with pressure; and / or vein: VDI≈0, uniform deformation; The deep classification network based on Efficient ViT, trained from a deep classification network, includes: unsupervised training on a large open-source database followed by fine-tuning on a fine-classification dataset of targets; wherein the unsupervised training is performed on multiple large-scale visual image datasets including ImageNet, COCO, and proprietary datasets, and fine-tuning is performed on a fine-classification dataset of interface targets.

8. The handheld intelligent vascular puncture system according to claim 7, characterized in that, The high-precision dynamic blood vessel localization algorithm includes: designing an energy function based on the fusion of random walk and graph cut algorithms to design feature information; preprocessing blood vessel images based on K-means clustering and Hough transform; and verifying the blood vessel image segmentation results using the level set method.

9. A handheld intelligent vascular puncture system according to claim 8, characterized in that, The target vessel identification and puncture path planning algorithm includes determining the target vessel selection strategy and planning the puncture path; wherein... The strategy for determining target vessel selection includes: Select the target vein or artery based on the working status: If the current working mode is venipuncture, select an available vein that is at an appropriate distance from the skin as the target vessel, and if there are multiple available vessels, they will form a candidate set; if the current working mode is arterial puncture, select the arterial vessel that is closest to the skin as the target vessel, and if there are multiple available vessels, they will form a candidate set. The optimal target vessel is selected based on the location and correlation of the blood vessels in the diagram, including: selecting the thickest vessel for puncture based on geometric calculations and comparisons; ensuring there are no other blood vessels between the punctured vessel and the skin; and selecting venous target vessels where there is no overlap of blood vessels; selecting vessels as close to the skin as possible while meeting the thickness requirements; and selecting arterial target vessels where there are no other blood vessels between the punctured vessel and the skin and no overlap of blood vessels. The planning objectives for the puncture path include: the insertion point into the skin, the angle between the needle and the skin, and the insertion length. For planning the venipuncture path, the center of the target vessel and the center of the needle direction are connected by a line. It is then determined whether there are other veins in the path. If not, the planning is successful, and the needle angle and insertion length are calculated. If there are other veins in the path, it is determined whether they meet the puncture requirements. If so, the target vessel is replaced and the puncture path is replanned. If not, the process returns to selecting other veins as the target vessel. If the requirements cannot be met, an "inappropriate position" message is displayed, and the operator makes adjustments. For arterial puncture path planning, the process includes: connecting the target vessel center with the needle direction selection center, and determining whether there are other veins or arteries in the connecting path; if not, the planning is successful, and the needle angle and insertion length can be calculated; if there are, it indicates that safe puncture is not possible, and the fixed position of the needle is readjusted. If the needle has a lateral movement control mechanism, the needle position is adjusted laterally by a fixed step size for searching; if the needle does not have a lateral movement control mechanism, the position is not suitable, and the operator needs to make adjustments.

10. A handheld intelligent vascular puncture system according to claim 9, characterized in that, The software application layer includes: Image recognition and vessel identification module: Vessel recognition is implemented based on TensorFlow Lite, using deep learning algorithms to identify vascular structures in images, extracting key parameters such as vessel diameter, depth, and location, and determining whether the target vessel is located in the center of the image and whether it meets the puncture conditions. The image recognition process is combined with AI algorithm processing, and model inference relies on a lightweight AI framework deployed on an ARM embedded platform to achieve real-time recognition. Human-computer interaction and decision support module: The interactive interface is implemented based on Android custom View. After identifying a puncturable blood vessel, the system prompts the user "There is a suitable blood vessel, do you want to puncture?" through the interface and controls the "puncture" button to become operable. With the graphical interface as the core, the interactive logic is built using an embedded GUI framework to assist the user in deciding whether to perform the puncture. After the user confirms, the puncture control stage is entered. Anomaly Handling and Reset Module: Anomaly handling is implemented using a finite state machine. When a puncture fails or is completed, the system guides the user into a reset process, controls the motor to return to its initial position, and waits for a reset completion signal. If the reset is successful, the system prompts "Searching for blood vessel" and re-enters the image recognition process, forming a closed loop.

Images