An arterial blood sampling intelligent guiding method and device based on near-infrared images

Abstract

The application discloses an arterial blood sampling intelligent guiding method and device based on near-infrared images, and the method comprises the following steps: synchronously collecting multispectral near-infrared images of three characteristic wave bands of 760 nm, 850 nm and 940 nm, and constructing a multispectral feature tensor; calculating a difference index based on a pixel-level oxygenation index and a deoxygenation index, combining a graph cut rough segmentation and a MobileU-Net deep learning fine segmentation to generate an arterial candidate mask; collecting time sequence images to extract a heart rate matched pulsation energy, generating a confirmed arterial mask through double verification of spectral features and pulsation features, and marking suspected arteries for those who do not pass the pulsation verification; multi-objective optimization is performed on the confirmed artery to select an optimal blood sampling point; three-point contact positioning and binocular vision fusion are adopted to estimate the subcutaneous depth of the artery; after the artery, vein, suspected artery, optimal blood sampling point, needle insertion direction and estimated depth are registered on the body surface, the micro laser projector is directly projected onto the patient's skin surface to guide the medical staff to manually implement puncture. The application combines spectral physiological recognition and pulsation function verification, and cooperates with AR body surface projection guidance, so that the one-time success rate and safety of arterial blood sampling are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of clinical auxiliary medical technology, and in particular to an intelligent guided method and device for arterial blood sampling based on near-infrared imaging. Background Technology

[0002] Arterial blood gas analysis is a core diagnostic tool for clinically assessing acid-base imbalance, degree of hypoxia, and respiratory function. The radial artery is the preferred puncture site for arterial blood sampling due to its superficial location and ease of compression for hemostasis. However, because arteries are small in diameter, deep in location, and exhibit significant individual variations, and because arterial palpation is challenging in obese, edematous, hypotensive, and elderly patients, repeated punctures, hematomas, and accidental venous punctures are common problems in clinical practice.

[0003] Existing technologies include near-infrared imaging-based blood vessel identification solutions, such as: CN115040079B selects blood collection points by sorting vein segment widths and comparing the overlap rate of photos, but it is designed for vein blood collection and does not have the ability to distinguish between arteries and veins. CN114677713A determines the needle insertion point by traversing pixels to find the maximum boundary point and midpoint, but it relies solely on single-band grayscale morphological features and cannot distinguish between arteries and veins. CN104799815B uses three near-infrared cameras to acquire images from different angles for 3D reconstruction. Although its manual mentions that arteries and veins can be distinguished based on the different absorption rates of near-infrared light by hemoglobin, this distinction method is only an incidental post-processing step in its 3D reconstruction process. It does not provide any specific distinction algorithm, feature calculation method, or verification data. Technically, it is only a simple heuristic rule of "coarse blood vessel + slightly different gray level = artery" rather than a quantitative spectral analysis method. Furthermore, it does not utilize the physiological characteristics of arterial pulsatile blood flow for functional verification, resulting in a high rate of misidentification of arteries and veins.

[0004] Furthermore, CN106456056B discloses a near-infrared arterial visualization device that uses a near-infrared light source to illuminate the wrist and utilizes the absorption characteristics of arterial blood to near-infrared light to create an image, assisting the operator in identifying the location of the radial artery. This solution belongs to the category of auxiliary guidance devices, but it has the following inherent drawbacks: (1) Using only single-band near-infrared imaging and relying on grayscale morphological features to identify blood vessels, it is impossible to distinguish between arteries and veins at the spectral level, and misjudgment is likely to occur in the area where veins and arteries run together. (2) Based solely on static image recognition, without utilizing the physiological characteristics of arterial pulsatile blood flow for functional verification, the recognition results lack physiological confirmation. (3) The recognition results are only displayed on the device screen. The operator needs to perform coordinate transformation between the screen image and the patient's body surface, which is not intuitive enough. (4) It does not provide multi-dimensional guidance information such as blood vessel depth estimation and needle insertion angle suggestions, and the guidance dimension is singular.

[0005] Overall, existing technologies have the following fundamental drawbacks when applied to arterial blood sampling guidance: Insufficient ability to distinguish between arteries and veins: Relying solely on single-band grayscale or morphological features to identify blood vessels without utilizing the spectral absorption differences between oxyhemoglobin and deoxyhemoglobin, it is unable to distinguish between arteries and veins from a physiological perspective, resulting in a high risk of accidental vein puncture.

[0006] Lack of arterial function verification mechanism: all methods use static image analysis and do not utilize the core physiological feature of arterial pulsatile blood flow. The identification remains at the "morphological appearance" level rather than the "functional confirmation" level, resulting in a high false positive rate.

[0007] The guidance method is not intuitive: the recognition results are only displayed on the screen, and the operator needs to switch between the "screen coordinates" and the "patient's body surface coordinates" in their mind, which is a heavy cognitive burden and has low guidance efficiency.

[0008] Key guidance dimensions are missing: Most solutions only output two-dimensional planar information, lacking subcutaneous vascular depth prediction, making it difficult to provide operators with a reference for needle insertion depth; and the stability of depth measurement is insufficient in handheld scenarios.

[0009] Equipment positioning deviates from clinical needs: fully automated puncture protocols have high compliance costs and low acceptance; existing auxiliary protocols lack precision and functionality, failing to meet the safety requirements of arterial blood collection. Summary of the Invention

[0010] This invention addresses the problems in the prior art by providing an intelligent guided method and device for arterial blood sampling based on near-infrared imaging.

[0011] The technical solution adopted by this invention to solve the above-mentioned technical problems is as follows: an intelligent guided method for arterial blood sampling based on near-infrared images, comprising the following steps: S1. Synchronous acquisition and preprocessing of multispectral near-infrared images S1.1 Simultaneously acquires near-infrared images of the target puncture area in three characteristic bands through the imaging unit. The three characteristic bands are λ1 = 760 nm, λ2 = 850 nm, and λ3 = 940 nm. Among them, 760 nm is the characteristic absorption peak of deoxyhemoglobin, 850 nm is the isoabsorption point of oxyhemoglobin and deoxyhemoglobin, and 940 nm is the characteristic absorption peak of oxyhemoglobin.

[0012] S1.2 performs joint preprocessing on the three-band images: distortion correction is completed based on camera intrinsic parameters, adaptive illumination equalization is achieved using the multi-scale Retinex algorithm, and edge-preserving denoising is completed using bilateral filtering.

[0013] S1.3 Constructing Multispectral Feature Tensors The feature vector of each pixel is .

[0014] S2. Intelligent arteriovenous segmentation based on hemoglobin spectral characteristics S2.1 Calculate pixel-level spectral feature indices: Oxygenation index: OI reflects the relative concentration of oxyhemoglobin. In arterial regions, the proportion of oxyhemoglobin is high, and the OI value is significantly higher. Deoxygenation index: , is the reciprocal of the oxygenation index, which helps characterize vascular blood properties from the perspective of the relative concentration of deoxyhemoglobin; Oxygenation-deoxygenation difference index: The difference between oxygenation and deoxygenation characteristics is used as the core spectral feature for arteriovenous classification; the ODI in the arterial region is positive and has a large amplitude, while the ODI in the venous region is close to zero or negative.

[0015] S2.2 Coarse Segmentation Based on Spectral Features: Using the ODI feature map as the main classification basis and combining it with local spatial consistency constraints, an improved graph cut algorithm is used for coarse segmentation to generate a mask for artery candidate regions. With vein area mask .

[0016] S2.3 Deep Learning Fine Segmentation: A lightweight MobileU-Net network is constructed, employing an encoder-decoder architecture. The encoder uses depthwise separable convolutions instead of standard convolutions, with four downsampling stages to progressively extract multi-scale spectral features. The decoder uses bilinear upsampling, fusing features from corresponding encoder layers via skip connections. The final output layer uses a sigmoid activation function to output an artery probability map. .

[0017] The network was trained using a combination of clinically collected radial artery multispectral image datasets and publicly available vascular datasets, with a training sample size of ≥2000 cases. Annotation was completed jointly by senior nurses and radiologists. The Adam optimizer was used during training, and the loss function was a weighted combination of Dice loss and cross-entropy loss. The total loss function was:

[0018] in , The weighting ratios were determined based on ablation experiments: weighting ratios (1.0 / 0.0, 0.7 / 0.3, 0.5 / 0.5, 0.3 / 0.7, 0.0 / 1.0) were tested on the validation set. The Dice coefficient reached its optimum at 0.7 / 0.3, indicating that the combined strategy of using Dice loss to optimize the overlap of blood vessel region segmentation and cross-entropy loss to assist in optimizing pixel-level classification accuracy is effective. The network inference frame rate is ≥15fps, meeting the real-time operation requirements of portable devices.

[0019] S2.4 Multi-feature fusion confirmation: The arterial probability map and the coarse spectral segmentation results are weighted and fused to generate the final arterial candidate mask.

[0020] in The adaptive threshold is determined as follows: The ODI values ​​of all pixels within the artery candidate region output by the coarse spectral segmentation are taken, and their mean is calculated. with standard deviation ,Pick ,coefficient The value is set to 1.0~2.0; the false positive rate is reduced by adapting the threshold to the differences in skin optical properties of different patients and combining cross-validation of spectral features and deep learning features.

[0021] S3. Verification of arterial pulsation function based on time-series images S3.1 Time-series image acquisition: Continuously acquire near-infrared image sequences of artery candidate regions at a frame rate of ≥30fps. The data collection time covers at least three complete cardiac cycles.

[0022] S3.2 Motion Interference Detection and Confidence Weighting: The angular velocity and acceleration data of the probe are acquired using the IMU inertial sensor built into the imaging unit to identify time periods of probe jitter and rapid displacement. For the light intensity time-series signal within the time window affected by motion interference, corresponding confidence weights are assigned in the subsequent pulsation energy calculation. The weight values ​​range from [value missing]. When the IMU detects that the angular velocity amplitude exceeds the preset severe jitter threshold, the weights are reset to 0, meaning that the signal for that period is completely discarded to avoid strong noise interfering with the extraction of pulsation features, thus obtaining the corrected light intensity timing signal. .

[0023] S3.3 Time-Frequency Feature Extraction: The corrected light intensity time-series signal of each pixel within the artery candidate mask is detrended, and short-time Fourier transform is used to extract time-frequency features, which are then calculated in relation to the patient's real-time heart rate. Normalized pulsatile energy within the matched frequency range: ,in This refers to the permissible range of heart rate fluctuations. This represents the Fourier transform.

[0024] S3.4 Artery Authenticity Dual Verification: A pixel is determined to be a real artery pixel when it simultaneously meets the following two conditions: (1) Pulsating energy That is, the pulsatile component that matches the heart rate is significant in the light intensity signal; (2) This pixel belongs to the artery candidate mask. The spectral characteristics support the conclusion that it is an artery.

[0025] Among them, pulsation energy threshold An adaptive threshold, determined during the preoperative individualized calibration phase: The average pulsation energy of background skin pixels outside the target region is calculated. with standard deviation ,Pick The coefficient k is set to 3-5 to ensure that the false positive rate is less than 5%.

[0026] The artery mask is confirmed by a set of pixels verified through dual authentication. Areas that fail pulsation verification based solely on spectral characteristics are marked as suspected arteries. Suspected arteries do not participate in subsequent blood collection point optimization calculations; they are only marked with a yellow dashed outline on the projection and display interfaces, with the message "Suspected artery, palpation confirmation recommended"—that is, prompting the operator to manually palpate the area and confirm it a second time by sensing pulsation through touch, thus fully leveraging the advantages of human-machine collaboration.

[0027] Boundary condition degradation handling: If the ODI feature discrimination is insufficient due to extremely low patient blood oxygen or abnormal skin optical properties, resulting in an empty coarse segmentation output or an area smaller than the preset threshold (e.g., 5mm²), the adaptive threshold calculation automatically switches to full-image ODI statistics, using 0.8 times the threshold as a temporary threshold, and simultaneously marks "Abnormal spectral signal, please check probe fit or consider palpation." If the IMU continuously detects severe jitter causing signal weights to be 0 for more than 30 consecutive frames, pulsation verification cannot be completed. The system automatically extends the acquisition time to 6 seconds and attempts recalibration. If this is still ineffective, the spectral recognition result is directly marked as "suspected artery," strongly prompting the operator to manually palpate, ensuring safe and conservative guidance even under non-ideal conditions.

[0028] S3.5 Generates an arterial pulsation distribution map: Within the identified arterial region, a pulsation function image is generated using pulsation amplitude as the brightness and pulsation phase as the hue, visually displaying the distribution of arterial pulsation strength and the direction of pulsation wave propagation.

[0029] S4. Intelligent blood collection point selection driven by multi-objective optimization S4.1 Candidate Point Generation: On the confirmed arterial mask corresponding to the vessel segment, a set of candidate blood collection points is generated along the arterial centerline at 0.5mm intervals. No candidate blood collection points are generated in suspected arterial areas.

[0030] S4.2 Multidimensional Feature Extraction: Extract the following features for each candidate point p: Blood vessel diameter : Obtained through arterial boundary detection; Pulse amplitude : Reflects the intensity of the arterial pulsation at that location; Blood vessel straightness : through the curvature of the center line calculate, The higher the straightness, the easier it is to puncture; Distance from adjacent veins The greater the distance, the lower the risk of accidental vein puncture; Anatomical location adaptation: A lightweight semantic segmentation network (such as U-Net) is pre-trained, taking near-infrared grayscale images as input, and outputting region probability maps of skeletal protrusions (such as the radial styloid process) and tendons (such as the palmaris longus tendon). Regions with a probability > 0.7 are marked as "non-puncture areas," and candidate points falling into these areas are scored directly to zero, ensuring that recommended points avoid these high-risk anatomical structures. This network can share feature extraction layers with the main segmentation network without adding additional computational burden.

[0031] S4.3 Multi-objective Integrated Scoring: Constructing a Weighted Scoring Function:

[0032] in Min-Max normalization; weight vector The weighting of pulsation amplitude is determined based on the experience of clinical experts. The highest value among all weights is used to select arterial segments with obvious pulsation.

[0033] S4.4 Outputs optimal blood collection point information: Selects the candidate point with the highest score as the recommended blood collection point. Output its two-dimensional coordinates and the angle of blood vessel orientation. Pulsation intensity level, vessel diameter.

[0034] S5. Subcutaneous vessel depth prediction based on contact localization and binocular vision fusion The S5.1 probe has three miniature contact positioning pins arranged in a triangle at its front end. When in use, the probe is gently placed on the skin surface, and the three positioning pins simultaneously contact the skin to determine the probe's reference plane. The miniature pressure sensor built into the positioning pin confirms that the probe is stably attached to the skin, and the angle between the reference plane and the skin surface is ≤15°.

[0035] S5.2 uses a binocular near-infrared camera to simultaneously acquire left and right viewpoint images. After epipolar correction, an improved SGM stereo matching algorithm is used to calculate the disparity map. The matching cost function of the improved SGM algorithm is defined as:

[0036] in, Let $D$ be the cost of the sum of absolute differences of pixel $p$ under parallax $d$. To confirm the value of the artery mask at this pixel (1 for the artery region, 0 for the background). The weighting coefficient (values ​​from 0.5 to 1.5) is used to improve the discriminative power of the matching cost by weighting the arterial region, thus prioritizing the accuracy of disparity calculation in the arterial region.

[0037] S5.3 Using the skin reference plane determined by the contact positioning pin as the zero depth point, and combining the camera calibration parameters and disparity map, the original depth estimate of the artery at the optimal blood collection point is obtained through pinhole imaging model inversion. .

[0038] When the probe tilts at an angle When the angle between the probe's optical axis and the skin's normal exceeds a 5° threshold, a triangular geometric model is used for depth compensation.

[0039] ,in The result is obtained by combining the relative displacement difference between the tips of the three contact positioning pins with the attitude calculation of the binocular camera. The estimated subcutaneous depth of the artery after compensation.

[0040] Handling Depth Prediction Failure: If the parallax map hole rate exceeds 30% (due to insufficient skin texture or reflection), the depth prediction module will output a "Depth estimation unreliable" message, without displaying specific values, and will suggest that the operator refer to ultrasound or palpation to estimate the depth, ensuring that unreliable quantitative data is not provided.

[0041] The depth result in S5.4 is only used as a reference for puncture and is not used for automatic control of puncture procedures.

[0042] S6. Real-time body surface guidance based on AR projection S6.1 Real-time registration of the body surface: Based on the skin texture and blood vessel contour features, the patient's body surface is registered in real time. When the patient's wrist moves slightly, the projection coordinates are corrected in real time to ensure that the guidance information is accurately aligned with the body surface.

[0043] S6.2 The following guidance information is projected directly onto the patient's skin surface using a miniature laser projector: Blood vessel distribution: Arteries are marked with red lines, veins with blue lines, and suspected arteries with yellow dashed lines; Optimal blood collection point: The recommended blood collection point location is marked with a bright gold circle, and the vessel diameter, pulsation intensity level, and estimated depth are labeled on the side; Needle insertion direction: The recommended needle insertion direction (along the direction of the blood vessel) is marked with a green arrow. Puncture angle: The recommended needle insertion angle is indicated by the angle marking (30°~45° with the skin surface).

[0044] The S6.3 projected content is updated in real time at a frequency of ≥15Hz and simultaneously displayed on the interactive display unit for the operator to refer to from multiple angles.

[0045] S7. Assisted monitoring and feedback during the puncture process S7.1 Needle Tip Recognition and Tracking: The needle tip position is identified in real time through a near-infrared imaging unit, employing a dual-mode switching tracking mechanism. (1) Priority mode: When the infrared reflective markers attached to the tip of the blood collection needle are detected, the marker tracking method is used to determine the position of the needle tip based on the pre-calibrated relative offset between the needle tip and the markers. The positioning accuracy is high and it is suitable for routine puncture scenarios. (2) Degradation mode: When no effective reflective marker is detected (such as no marker is pasted or the marker is blocked), it automatically switches to the needle body contour segmentation and tracking method. The needle body contour is segmented by deep learning and the needle tip coordinates are fitted. It is suitable for emergency operation scenarios without markers.

[0046] When the needle tip approaches the skin surface (distance <3cm), it automatically switches to a higher frame rate acquisition mode to improve the stability and accuracy of close-range needle tracking.

[0047] If both modes fail (e.g., the marker is completely obscured and the confidence level of needle contour segmentation is below the threshold), the device will issue an audible and visual alarm for "tracking loss," prompting the operator to pause the puncture procedure and readjust the probe's field of view to prevent blind needle insertion without tracking assistance.

[0048] It displays the positional deviation of the needle tip relative to the recommended blood collection point in real time, assisting the operator in adjusting the needle insertion position.

[0049] S7.2 Puncture Depth Indication: Combining the three-dimensional positioning of the needle tip with the marked position on the skin surface, the current needle tip depth is estimated, and the operator is prompted with the relative relationship between the current depth and the estimated arterial depth.

[0050] S7.3 Pulse Waveform Reference: Real-time display of arterial pulsation waveforms in the area near the needle tip, assisting the operator in determining the relative distance between the needle tip and the artery.

[0051] S7.4 Assisted verification of puncture results: After puncture, the blood return characteristics and pulsation waveform changes are observed through near-infrared imaging to assist the operator in judging whether the puncture was successful.

[0052] S8. Preoperative individualized calibration and continuous optimization S8.1 Preoperative data acquisition: Acquires approximately 3 seconds of pre-data, automatically obtains the patient's real-time heart rate, skin optical characteristics, and baseline level of pulsation signal, and adaptively adjusts image enhancement parameters and pulsation energy determination threshold.

[0053] S8.2 Data Recording and Optimization: With the operator's consent, patient characteristics, identification results, deviation between recommended and actual puncture points, and success rate for each operation are anonymously recorded for subsequent iterations to optimize algorithm parameters and scoring weights.

[0054] A smart arterial blood sampling guidance device based on near-infrared imaging The apparatus for implementing the above method includes: Handheld imaging probe: integrates a multispectral near-infrared imaging unit, a binocular near-infrared stereo vision unit, an IMU inertial sensor, and three miniature contact positioning pins arranged in a triangle; the multispectral near-infrared imaging unit is equipped with an LED light source array with three characteristic bands of 760nm, 850nm, and 940nm and a corresponding near-infrared camera; the overall weight of the probe is ≤300g, which meets the ergonomic requirements for single-handed operation.

[0055] Miniature laser projection unit: preferably integrated inside the probe (the projection optical axis and imaging optical axis are pre-calibrated in the factory and do not require on-site calibration after leaving the factory), or can be installed on a bedside bracket as a standalone implementation and driven by the host processing unit via wired or wireless means.

[0056] The host processing unit includes a GPU-accelerated computing module and has built-in algorithms for artery recognition, pulsation verification, depth estimation, AR registration and projection control, needle tip tracking, and puncture-assisted monitoring.

[0057] Interactive display unit: High-resolution touch screen that simultaneously displays raw near-infrared images, artery segmentation results, pulsation function images, recommended blood collection point information, and puncture-assisted monitoring information, and supports interactive operations such as zooming and annotation.

[0058] Control switch unit: Foot pedal control switch enables non-contact image acquisition triggering and screen freezing, conforming to aseptic operation specifications.

[0059] As a preferred embodiment, the handheld imaging probe can also integrate a laser speckle contrast imaging (LSCI) module, which includes a 785nm laser diode and a high-speed CMOS camera with a frame rate of ≥100fps. Based on the temporal changes in speckle contrast, it extracts blood flow pulsation signals, replacing ordinary near-infrared temporal images to complete arterial pulsation verification, and further improves the signal-to-noise ratio of pulsation detection.

[0060] As a preferred embodiment, the device is also equipped with a disposable sterile protective film sleeve that is adapted to the shape of the handheld imaging probe. This sleeve is used to isolate the probe body from the surgical area during puncture operations, thus meeting clinical sterility requirements.

[0061] Compared with existing technologies, this invention has the following advantages: Based on multispectral near-infrared imaging with three characteristic bands (760nm, 850nm, and 940nm), this invention quantitatively distinguishes arteries and veins from the physiological level of hemoglobin oxygenation status. Combining coarse segmentation with MobileU-Net deep learning for fine segmentation significantly improves the accuracy of arterial and vein differentiation and effectively reduces the false positive rate of arterial identification. Furthermore, it introduces a dual verification mechanism based on temporal image-based pulsation and heart rate matching to confirm the authenticity of arteries from a functional perspective, fundamentally reducing the risk of misidentifying veins as arteries. It employs three-point contact positioning to determine the reference plane and combines... The device combines binocular vision with tilt angle compensation for depth estimation, providing operators with a stable reference for subcutaneous depth. Simultaneously, AR projection directly projects guidance information such as arteries, veins, suspected arteries, optimal blood collection points, needle insertion direction, and estimated depth onto the patient's skin, eliminating the cognitive burden of switching between screen and skin coordinates and providing intuitive guidance. The entire device is handheld, positioned as an auxiliary tool rather than a fully automated puncture device, preserving the final decision-making authority of medical personnel. Combined with a foot switch and disposable sterile protective film, it fully complies with clinical aseptic operation standards, effectively improving the success rate and safety of arterial blood collection on the first attempt. Attached Figure Description

[0062] The present invention will be further described in detail below with reference to the accompanying drawings and preferred embodiments. However, those skilled in the art will understand that these drawings are drawn only for the purpose of explaining the preferred embodiments and therefore should not be regarded as a limitation on the scope of the present invention. In addition, unless otherwise specified, the drawings are not necessarily drawn to scale.

[0063] Figure 1 This is a flowchart of the method of the present invention.

[0064] Figure 2 This is a block diagram of the intelligent arterial blood collection guidance device of the present invention. Detailed Implementation

[0065] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings. Those skilled in the art will appreciate that these descriptions are merely descriptive and exemplary and should not be construed as limiting the scope of protection of the present invention.

[0066] It should be noted that similar labels in the following figures indicate similar items; therefore, once an item is defined in one figure, it may not be further defined and explained in subsequent figures. Example

[0067] The following two embodiments mainly illustrate in detail the specific implementation of an intelligent guidance method and device for arterial blood sampling based on near-infrared images, such as... Figure 1 and 2 As shown, the details are as follows: Example

[0068] In this embodiment, the operator first removes the handheld imaging probe (weighing no more than 300 grams) from its sterile packaging and checks the three miniature contact positioning pins (10 mm on each side) arranged in an equilateral triangle at the front of the probe to ensure they are intact, confirming that the initial reading of the pressure sensor is zero. Then, a disposable sterile protective sleeve (transparent material with a near-infrared transmittance of no less than 90%) is tightly fitted over the probe, ensuring that the sleeve does not obstruct the imaging window, positioning pins, or laser projection outlet. The probe data cable is then connected to the host processing unit (with a built-in NVIDIA Jetson AGX Orin GPU module), the 10.1-inch high-resolution touchscreen is activated, the main control software is opened, and the foot pedal control switch (optional) is placed near the operator's feet as needed. The operator supports the patient's forearm with their left hand, allowing the wrist to extend naturally with the palm facing upwards, placing a sandbag under the wrist to stabilize it. The right hand holds the probe, aiming it at the radial artery area of ​​the wrist (approximately 1-2 cm proximal to the wrist crease), and slowly moves it downwards until all three positioning pins lightly touch the skin. By observing the pressure sensor readings on the display screen, when the three pressure values ​​are all stable between 0.1 and 0.3 Newtons and the difference between them does not exceed 0.05 Newtons, it is considered that the probe reference plane is in good contact with the skin surface. At this time, the angle between the probe optical axis and the skin normal (calculated from the relative displacement difference of the positioning pin tip) is usually less than 10°, and this angle will be displayed on the screen in real time.

[0069] After the equipment is in place, the operator presses the foot switch (or lightly touches the acquisition button on the probe) to trigger the pre-acquisition mode. The equipment continuously acquires near-infrared images of the target area in three characteristic bands (760 nm, 850 nm, and 940 nm) for 3 seconds (90 frames in total) at a frame rate of 30 frames per second for individualized calibration. During the calibration process, the skin's basic reflectivity is first calculated based on the average of multiple frames: a 10×10 pixel window is selected at the four corners of the image, avoiding blood vessels, and the average gray value of each band is calculated. Based on this, the illumination equalization parameters of the multi-scale Retinex algorithm are set—the gain coefficient for areas with gray values ​​below 30 is set to 1.8, the gain coefficient for areas above 200 is set to 0.6, and linear interpolation is used in the intermediate areas to stretch the overall gray-level histogram to the full range of 0–255. Then, the patient's real-time heart rate is extracted. The global mean of the light intensity time-series signals of all pixels is taken, and after bandpass filtering from 0.5 to 4.0 Hz, the dominant frequency is extracted using the autocorrelation method. If the confidence level of the dominant frequency is greater than 0.8, it is confirmed as heart rate; otherwise, the operator is prompted to manually input the heart rate value obtained from the monitor or palpation. In this example, the measured heart rate... Blasts per minute (i.e., 1.2 Hz). Next, the background pulsation energy baseline is calculated: avascular skin regions are selected (automatically identified by the preliminary vessel detection algorithm), and the power spectral density of the temporal signal of the light intensity for each pixel is calculated, and the frequency... Integrating within the Hertz range yields the pulsation energy of each background pixel. Then, its mean can be calculated. and standard deviation Set coefficients (This value applies to the radial artery in the general population), then the pulsation energy threshold. .like If the value exceeds 0.05, it indicates excessive ambient light interference. The device will automatically increase the LED light source power and recalibrate. After calibration, the device will emit a short beep, and the screen will display "Calibration complete, recognition can begin".

[0070] During the formal identification phase, the device synchronously triggers the three-band light source arrays to illuminate sequentially in a time-multiplexed order (10 milliseconds exposure time per frame, 3.3 milliseconds interval between frames). The corresponding near-infrared camera (1280×1024 resolution, 5.2 micrometer pixel size) simultaneously acquires the data. The total time to complete one frame of a multispectral image is approximately 40 milliseconds, meaning the multispectral tensor update rate is approximately 25 frames per second. For each frame of the three-channel raw image, the calibrated camera intrinsic parameters (including focal length, principal point coordinates, and radial distortion coefficients) are first used. Distortion correction is performed by reconstructing the distortion-free image through inverse mapping and bilinear interpolation; then, the multi-scale Retinex algorithm (Gaussian wrap-around scale) is used. The illumination and reflection components were separated by taking 15, 80, and 200 pixels respectively, retaining only the reflection component to eliminate uneven illumination; finally, a bilateral filter (spatial domain) was applied. Pixels, grayscale Edge-preserving denoising is performed at grayscale levels to smooth the texture while preserving blood vessel boundaries. A multispectral feature tensor is then constructed after preprocessing. (in , The feature vector of each pixel is And store it in the GPU memory.

[0071] In the arteriovenous spectral segmentation stage, for each pixel Calculate the oxygenation index sequentially Deoxygenation index and oxygenation-deoxygenation difference index All computations were accelerated using CUDA parallel processing, taking less than 5 milliseconds. Subsequently, based primarily on the ODI feature map, an improved graph cut algorithm was used for coarse segmentation: the data terms employed a Gaussian mixture model of ODI values, and the smoothing terms used a four-neighbor Potts model. The top 5% of high-value pixels in the entire image's ODI value are selected as artery seed points, the bottom 5% as vein seed points, and the rest are treated as unknown points. After iterative minimum cut, the artery candidate mask is obtained. and vein mask The coarse segmentation results serve as the basis for subsequent adaptive threshold calculations and also provide auxiliary information for the deep learning network. On the other hand, a lightweight MobileU-Net network is constructed for fine segmentation. Its encoder employs depthwise separable convolutions (3×3 kernels, dilation factor 2), with four downsampling stages. The feature map sizes are 512×640, 256×320, 128×160, and 64×80, with 32, 64, 128, and 256 channels respectively. The decoder uses bilinear upsampling, with each layer concatenating with the corresponding encoder feature map via skip connections. Finally, it undergoes a 1×1 convolution and sigmoid activation to output an artery probability map. The network weights were pre-trained jointly on a self-built radial artery multispectral dataset (2500 cases, covering different ages, skin colors, and BMIs) and a publicly available vascular dataset, using the Adam optimizer with an initial learning rate of [missing information]. Batch size 8, training 100 epochs, total loss function (in For the Sørensen-Dice coefficient loss, (This is a binary cross-entropy loss). This weighting ratio was confirmed as optimal through ablation experiments, achieving a Dice coefficient of 0.892 on the independent validation set (500 cases). The network inference time per frame is approximately 55 milliseconds, meeting real-time requirements. Subsequently, the coarse segmentation and deep learning results are fused, and the average ODI value of all pixels within the coarse segmented arterial region is taken. and standard deviation In this example , Set coefficients Then the adaptive threshold The final artery candidate mask is defined as... In other words, an artery is only considered a candidate artery when the product of the artery probability and the ODI feature exceeds a threshold. This cross-validation strategy effectively reduces the false detection rate of a single method. The segmentation results are displayed in real time as a red pseudo-color overlay on a grayscale image.

[0072] To functionally confirm the authenticity of the artery, the device continuously acquired at least 3.5 seconds (105 frames in total) of 850 nm single-band near-infrared temporal images within the artery candidate mask area at a frame rate of 30 frames / second (this band is an isoabsorption point, less affected by blood oxygen saturation, and suitable for pulsation monitoring). Simultaneously, the angular velocity of the probe's built-in IMU was read at a sampling rate of 200 Hz. And acceleration data. After aligning the IMU data with the image frame timestamps, the angular velocity amplitude within the corresponding time window (15 milliseconds before and after) for each frame is calculated. If the amplitude exceeds 5 radians / second (the threshold for severe jitter), the confidence weight of the corresponding frame is reset to 0 and it is completely discarded; if it is between 2 and 5 radians / second, the weight is linearly mapped to 0.3 to 0.8; if it does not exceed 2 radians / second, the weight is 1.0. The same rule applies to weight reduction when the acceleration change exceeds 1.5g. In this example, of the 105 frames acquired, 8 frames had their weight reduced to 0.5 due to slight jitter, 2 frames had their weight reduced to 0 due to severe movement, and the remaining 95 frames had a weight of 1.0. For each pixel in a valid frame, its light intensity time-series signal is detrended, then subjected to a short-time Fourier transform using a Hanning window (window length 128 points, overlap 64 points, frequency resolution approximately 0.23 Hz) to extract the heart rate frequency band. The power spectral density within the Hertz range was calculated, and the total power across the entire frequency band (0–15 Hz) was normalized to obtain the normalized pulsatile energy. ,in hertz, Hertz. Then perform double verification: simultaneously satisfying (This example is calibrated to obtain) , Therefore And this pixel belongs to Only those that are confirmed are identified as confirming arteries and constitute a confirming artery mask. Pixels with insufficient pulsation energy but only spectral characteristics are marked as "suspected arteries." This area is not included in the subsequent blood sampling point selection; it is only marked with a yellow dashed outline on the projection and display interface, prompting the operator to manually confirm by palpation. In addition, within the confirmed artery area, a pulsation distribution map is generated using pulsation amplitude as the brightness and pulsation phase as the hue. The phase gradient direction can indicate the direction of blood flow conduction, providing the operator with more references.

[0073] After confirming the arterial mask, the device performs morphological skeletonization (using the Zhang-Suen thinning algorithm) to obtain a centerline with a single pixel width, and removes short branches with a length of less than 5 mm, retaining the main branches. A set of candidate blood sampling points is generated along the centerline of the main branches with a sampling step size of 0.5 mm (approximately 0.08 mm / pixel, or about 6 pixels). In this case, the main branch of the radial artery is approximately 25 mm long, and a total of 50 candidate points were generated. For each candidate point... Four main features were extracted: blood vessel diameter (Search for the arterial boundary along the vertical centerline and take the median distance between the two sides), pulsation amplitude (Brightness values ​​taken from pulsation distribution map), vessel straightness (in (This refers to the local curvature, obtained through least-squares circle fitting), and the distance to adjacent veins. (i.e., to the vein mask) The minimum Euclidean distance is calculated. Simultaneously, a pre-trained lightweight semantic segmentation network identifies skeletal protrusions (such as the radial styloid process) and tendon regions (such as the palmaris longus tendon). Regions with a probability > 0.7 are marked as "non-puncture areas," and if a candidate point falls into these areas, the score is directly set to zero. The above four features are then Min-Max normalized to... Intervals, and using a weighted scoring function. The weights were determined based on feedback from 10 senior nurses: (Blood vessel diameter) (Pulsation amplitude, highest weight). (Straightness) (Distance from vein) satisfies and After calculating the scores of all candidate points, the one with the highest score is selected as the recommended blood collection point. In this example, the 28th candidate point (approximately 14 mm from the wrist crease) scored 0.92, corresponding to a vessel diameter of 2.3 mm, a pulsation amplitude of 0.85 (high), a straightness of 0.91, and a distance of 3.2 mm from the vein. The angle of the vessel's course at this point was also recorded. (Angle between the centerline tangent and the horizontal axis) and pulsation intensity level (strong).

[0074] Next, subcutaneous vascular depth estimation was performed. The probe, equipped with a built-in binocular near-infrared camera (baseline distance 40 mm, focal length 12 mm), simultaneously acquired 850 nm images from both left and right viewpoints, which were then polarimetrically corrected (using a calibrated rotation matrix). Translation vector and essential matrix After eliminating vertical disparity, an improved semi-global matching (SGM) algorithm is used to calculate the disparity map. The matching cost function is defined as follows: ,in The sum of absolute differences within a 5x5 window. To confirm the artery mask values ​​(artery is 1, background is 0). Cost differentiation is enhanced for arterial regions. Cost aggregation is performed using dynamic programming along four paths (horizontal left, horizontal right, vertical up, and vertical down), with smoothed penalty coefficients. , The disparity search range is 0–80 pixels (corresponding to a depth of 0–60 mm). After left-right consistency detection (threshold 1 pixel) and subpixel interpolation, the optimal blood sampling point is obtained by backprojecting the skin plane defined by the three contact positioning pins as the zero depth point, based on the extrinsic parameters of the binocular camera. raw depth estimate In this example, the measurement was... Millimeters. Simultaneously, the angle between the probe's optical axis and the skin's normal is calculated in real-time based on the difference in the positioning pin's projected coordinates and IMU attitude data. In this example Since the angle exceeds the 5° threshold, tilt compensation is performed. Millimeters, the final output is the compensated estimated depth. The depth is measured in millimeters (with an error range of ±0.3 mm, derived from the standard deviation of previous validation experiments). This depth is for operator reference only and is not intended for automatic control. If the disparity map hole rate exceeds 30%, the depth module will output a "depth estimation unreliable" message. In this example, the disparity map hole rate is less than 5%, so the output is normal.

[0075] After completing artery identification, blood collection point selection, and depth estimation, the device provides intuitive guidance to the operator through real-time registration on the body surface and AR projection. The registration process uses skin texture (such as fine wrinkles and pores) and blood vessel contours in the current frame of near-infrared image as feature points (SURF descriptors, matching distance ratio threshold 0.8), and performs affine transformation matching with the initial calibration frame. If the patient's wrist displacement exceeds 0.5 mm, the projection coordinates are updated immediately. The miniature laser projector (50 mW optical power, 532 nm green light wavelength, Class 1M safety rating, projection and imaging optical axes pre-calibrated at the factory and maintained in a fixed relative position after shipment) is driven by the host and projects guidance information onto the patient's skin surface at a refresh rate of 20 Hz: confirmed arteries are displayed as solid red lines (1.5 mm wide), accompanying veins as solid blue lines, and suspected arteries as dashed yellow lines (2 mm long, 1 mm interval); recommended blood sampling points are highlighted with gold rings (3 mm outer diameter, 1 mm inner diameter) and marked with "Ø2.3 mm, strong pulsation depth 2.5 mm"; from this point, follow the direction of the blood vessel... A green arrow (10 mm long, tip pointing to the distal end) is projected to indicate the direction of needle insertion; a fan-shaped angle indicator (range 30°~45°) is projected next to the arrow, with the label "Recommended 35° needle insertion". The color, brightness and size of all projected elements are automatically adjusted based on feedback from the ambient light sensor, while the interactive display screen synchronously displays the complete image, segmentation results and projection preview.

[0076] During the puncture, the device activates its auxiliary monitoring function. The operator uses a disposable arterial blood collection needle (22G, 25 mm in length), with an infrared reflective marker (a 1 mm diameter microbead reflective film) attached to its end. The device defaults to "marker-priority tracking mode," which extracts bright areas with grayscale values ​​exceeding 230 in each near-infrared image frame using morphological opening operations, calculates their geometric center, and then calculates the needle tip coordinates based on the pre-calibrated relative offset between the needle tip and the marker. If the marker is obscured or not attached, the device automatically switches to "needle contour segmentation tracking mode" within 0.2 seconds. This mode uses a lightweight U-Net to segment the needle mask, fits a straight line of the needle body using Hough transform, and takes the end point as the needle tip. When the needle tip is less than 3 cm from the skin surface, the device automatically increases the acquisition frame rate from 30 frames / second to 60 frames / second and reduces the search window to enhance stability. If both modes fail (e.g., the marker is completely obscured and the needle segmentation confidence level is below 0.6), an audible and visual alarm will be issued (a buzzer sound and a red flashing message "Tracking lost, please pause" on the screen), prompting the operator to stop the puncture and adjust the probe's field of view. Simultaneously, the screen will display the current needle tip position and the recommended blood collection point in real time. Horizontal and vertical deviations (in millimeters), and the current estimated depth of the needle tip and its relationship to the artery depth. The difference (e.g., "distance from artery: +0.5mm" indicates the needle is still above the artery, "-0.3mm" indicates the needle may have passed through the posterior wall). The average pulsation waveform of the area near the needle tip (within a 3mm radius circle) is also plotted at the bottom of the screen. Its amplitude gradually increases as the needle tip approaches the artery, assisting the operator in judging the needle depth. After puncture, blood return is observed. The device automatically captures a 3-second image sequence after blood return, analyzes the increase in ODI value and changes in pulsation waveform in the blood return area (the pulsation amplitude decreases as the needle occupies the blood vessel after entering the artery), and comprehensively judges whether the puncture was successful. If successful, a green "✔ Puncture successful" is displayed; otherwise, an orange "⚠ Please check blood return" is displayed, and the operation data is saved.

[0077] Finally, with the consent of both the patient and the operator, the device anonymously records information such as the patient's age, BMI, heart rate, mean skin grayscale value, recognition result, deviation between the actual puncture point and the recommended point, success rate, and puncture time for each procedure. This information is encrypted and stored locally, and periodically exported by hospital administrators for subsequent algorithm iterations (such as adjusting the weight vector). Adaptive threshold coefficient and (Or, incremental training of the deep learning network). After the operation in this example is completed, all data is automatically packaged and archived, and the device enters standby mode. Example

[0078] Unlike Embodiment 1, this preferred embodiment integrates a laser speckle contrast imaging (LSCI) module within the handheld imaging probe. This module includes a 785nm laser diode (5mW output power, 20mm x 20mm spot diameter) and a high-speed CMOS camera with a frame rate of at least 100 frames per second (150 frames per second in this example). During pulsation verification, the operator can choose to enable LSCI mode instead of ordinary near-infrared time-series imaging. Specifically, the laser diode emits coherent light to illuminate the candidate artery region, and the high-speed camera continuously acquires speckle images (each frame exposure time is 2ms). For each frame of speckle image, the spatial contrast of each pixel is calculated. ,in The standard deviation of strength within a 5×5 neighborhood. The contrast map is obtained by averaging the intensity of the neighborhood. Then, a Fourier analysis, identical to that in Example 1, is performed on the contrast variation sequence of the same pixel over time (150 data points / second) to extract the pulsation energy within the heart rate frequency band as a pulsation index. Since speckle contrast is extremely sensitive to changes in blood flow velocity, its pulsation signal-to-noise ratio can be improved by approximately 8–12 dB compared to ordinary light intensity signals, making it particularly suitable for patients with poor peripheral circulation, hypotension, or obesity. All other steps—including spectral segmentation, blood sampling point selection, depth prediction, AR projection, and puncture monitoring—are consistent with Example 1, except for the pulsation energy threshold. It is necessary to recalibrate based on the LSCI baseline (usually about 1.5 times that of the normal mode). This calibration is automatically completed by the preoperative pre-acquisition program without manual intervention.

[0079] This invention adopts the design concept of "automatic device processing, with the operator only needing to focus on the results." During clinical operation, the operator only needs to complete three core actions: gently placing the probe against the skin, triggering the sampling (by foot or hand), and performing the puncture guided by AR projection on the body surface. From the moment the probe contacts the skin to the output of the projection guidance information, the device automatically completes all recognition and calculation, with a total processing time of approximately 3.5 seconds (including preoperative calibration). The operator requires no additional intervention, significantly lower than the average operation time of traditional arterial blood sampling. The device's internal algorithm is fully automated; the operator does not need to understand the technical details of spectral analysis, image segmentation, stereo matching, etc., to obtain intuitive references for vascular distribution, optimal sampling point, needle insertion direction, and depth, greatly shortening the learning curve for novice nurses.

Claims

1. A smart guided method for arterial blood sampling based on near-infrared imaging, characterized in that, Includes the following steps: S1. Synchronous acquisition and preprocessing of multispectral near-infrared images: Synchronously acquire near-infrared images of the target puncture area in three characteristic bands of 760nm, 850nm and 940nm, and construct multispectral feature tensors after distortion correction, illumination equalization and edge denoising preprocessing. S2. Intelligent segmentation of arteries and veins based on hemoglobin spectral features: Pixel-level oxygenation index, deoxygenation index and oxygenation-deoxygenation difference index are calculated based on the multispectral feature tensor. Combined with coarse spectral segmentation and fine segmentation by deep learning, an artery candidate mask is generated. S3. Arterial pulsation function verification based on time-series images: Time-series near-infrared images of candidate arterial regions are acquired. Combined with motion interference suppression and time-frequency analysis, pulsation energy matching the patient's real-time heart rate is extracted. Through dual verification of spectral features and pulsation features, a confirmed arterial mask is generated. Regions that pass only spectral features but fail pulsation verification are marked as suspected arteries and do not participate in subsequent blood sampling point selection, for operator palpation confirmation. S4. Intelligent selection of blood collection points driven by multi-objective optimization: Generate candidate blood collection points on the blood vessel segment corresponding to the confirmed arterial mask, extract multi-dimensional features and weight scores, and output the optimal blood collection point and corresponding blood vessel direction information; S5. Subcutaneous depth estimation of blood vessels based on contact positioning and binocular vision fusion: Three miniature contact positioning needles arranged in a triangle at the front end of the probe simultaneously contact the skin to determine the probe reference plane. An improved SGM stereo matching algorithm is used to calculate the disparity map. After tilt angle compensation, the estimated subcutaneous depth of the artery at the optimal blood collection point is output. This depth is only used as a reference for medical staff to perform puncture. S6. Real-time surface guidance based on AR projection: The guidance information for identifying arteries, veins, suspected arteries, optimal blood collection points, needle insertion direction, puncture angle, and estimated depth is directly projected onto the patient's skin surface after real-time registration based on skin texture and vascular contour features. This allows medical staff to manually perform arterial blood collection punctures based on the guidance information.

2. The intelligent guided method for arterial blood collection according to claim 1, characterized in that, Step S2 specifically includes: Calculate the pixel-level oxygenation index Deoxygenation index and oxygenation-deoxygenation difference index ,in A small constant is preset to prevent division by zero; Based on the ODI feature map, an improved graph cut algorithm is used to complete the coarse spectral segmentation, generate artery candidate regions, and then output the artery probability map through the MobileU-Net deep learning network. The artery probability map and the ODI feature map are weighted and fused to generate an artery candidate mask, and a threshold is determined. As an adaptive threshold, the mean ODI value within the coarsely segmented arterial region is used. with standard deviation ,according to Calculate the coefficients It is a preset positive number, with a value range of 1.0 to 2.

0.

3. The intelligent guided method for arterial blood collection according to claim 2, characterized in that, The MobileU-Net network adopts an encoder-decoder architecture. The encoder uses depthwise separable convolution, the decoder uses bilinear upsampling and fuses corresponding layer features through skip connections, and the output layer uses the sigmoid activation function. The network training uses a weighted combination of Dice loss and cross-entropy loss, with Dice loss having a weight of 0.7 and cross-entropy loss having a weight of 0.

3.

4. The intelligent guided method for arterial blood collection according to claim 1, characterized in that, Step S3 specifically includes: Time-series images are acquired at a preset frame rate of not less than 30fps. Probe motion data is acquired through an IMU inertial sensor, and confidence weights are assigned to signals within the time window affected by motion interference. When the angular velocity exceeds a preset severe jitter threshold of 5 rad / s, the signal weight for the corresponding time period is reset to 0 and completely discarded. Time-frequency analysis was performed on the corrected light intensity time-series signal to calculate the normalized beat energy within the frequency range matching the patient's real-time heart rate; The pulsation energy threshold is an adaptive threshold, taking the average pulsation energy of the background skin pixels. with standard deviation ,according to The calculation is performed, where the coefficient k is a preset positive number with a value range of 3 to 5; Pixels that simultaneously satisfy the condition of having pulsation energy greater than the threshold and belonging to the artery candidate mask are identified as confirmed arteries; regions that fail the pulsation verification based solely on spectral features are marked as suspected arteries, prompting the operator to confirm by palpation.

5. The intelligent guided method for arterial blood collection according to claim 1, characterized in that, In step S4, the multidimensional features of candidate blood collection points include vessel diameter, pulsation amplitude, vessel straightness, distance from adjacent veins, and anatomical location fit. The anatomical location fit is determined by identifying bony prominences and tendon regions, excluding candidate points falling into these regions. The weighted scoring function is... ,in This indicates Min-Max normalization, where the weights satisfy... ,and .

6. The intelligent guided method for arterial blood collection according to claim 1, characterized in that, In step S5, the matching cost function of the improved SGM stereo matching algorithm is: in Let $D$ be the cost of the sum of absolute differences of pixel $p$ under parallax $d$. To confirm the arterial mask value, This is the weighting coefficient, and its value ranges from 0.5 to 1.5; When the probe tilt angle exceeds the 5° threshold, depth compensation is performed using a triangular geometric model. ,in This is the original depth estimate. The angle between the probe's optical axis and the skin's normal. For the compensated estimated depth, the It is obtained by combining the relative displacement difference of the tips of the three contact positioning pins with the attitude calculation of the binocular camera.

7. The intelligent guided method for arterial blood collection according to claim 1, characterized in that, It also includes auxiliary monitoring steps during the puncture process: A dual-mode switching mechanism is used to track the needle tip of the blood collection needle: infrared reflective marker tracking is used first, and the needle body contour segmentation tracking method is automatically switched when no effective marker is detected; when the needle tip is less than 3cm away from the skin, the acquisition frame rate is automatically increased; when both tracking modes fail, an audible and visual alarm for tracking loss is issued. It displays the positional deviation between the needle tip and the recommended blood collection point, the current estimated depth, and the arterial pulsation waveform near the needle tip in real time, assisting the operator in adjusting the puncture procedure.

8. A near-infrared imaging-based intelligent arterial blood collection guidance device, used to implement the method described in any one of claims 1-7, characterized in that, include: The handheld imaging probe integrates a multispectral near-infrared imaging unit, a binocular near-infrared stereo vision unit, an IMU inertial sensor, and three miniature contact positioning pins arranged in a triangle; the multispectral near-infrared imaging unit is equipped with an LED light source array with three characteristic bands of 760nm, 850nm, and 940nm and a corresponding near-infrared camera. A miniature laser projection unit, integrated inside the probe or mounted on a bedside bracket, is used to project guidance information directly onto the patient's skin surface after registration with the body surface. The host processing unit includes a GPU-accelerated computing module and has built-in algorithms for artery recognition, pulsation verification, depth estimation, AR registration and projection control, needle tip tracking and puncture-assisted monitoring. An interactive display unit is used to simultaneously display images, recognition results, and puncture guidance information.

9. The intelligent arterial blood collection guidance device according to claim 8, characterized in that, The handheld imaging probe also integrates a laser speckle contrast imaging module, which includes a 785nm laser diode and a high-speed CMOS camera with a frame rate of ≥100fps, for extracting blood flow pulsation signals based on temporal changes in speckle contrast.

10. The intelligent arterial blood collection guidance device according to claim 8, characterized in that, The device is also equipped with a disposable sterile protective film cover adapted to the shape of a handheld imaging probe; the projection optical axis and imaging optical axis of the micro laser projection unit are pre-calibrated in the factory to maintain a fixed relative position after leaving the factory.

Images