Anesthesia puncture auxiliary positioning method based on artificial intelligence

Abstract

This invention relates to the field of ultrasound image processing technology, specifically to an artificial intelligence-based method for assisted localization during anesthesia puncture. The invention first extracts the resistance factor of acoustic wave transmission at each pixel; then, based on an optical flow algorithm, it captures the cumulative deflection angle of each historical frame relative to the current frame, and constructs a spatiotemporally aligned resistance data volume; further, it obtains a corrected resistance map, extracts the cumulative resistance map of acoustic wave arrival and the minimum resistance trajectory; further, it projects the minimum resistance trajectory back into each time slice within the resistance data volume to obtain the instantaneous total resistance sequence for each deep pixel; further, based on the data fluctuations within the instantaneous total resistance sequence and combined with the distribution of the cumulative deflection angle, it obtains the resistance-angle change gradient; finally, based on the cumulative resistance map of acoustic wave arrival, the resistance-angle change gradient, and the instantaneous total resistance sequence, it distinguishes between the real intervertebral space and artifacts in low signal-to-noise ratio ultrasound images, generating reliable localization guidance information.

Description

Technical Field

[0001] This invention relates to the field of ultrasound image processing technology, and more specifically to an artificial intelligence-based method for assisted localization during anesthesia puncture. Background Technology

[0002] In anesthesiology practice, spinal puncture relies on accurate localization of the intervertebral space. For obese patients, due to the significant thickening of the subcutaneous fat layer, bony landmarks are difficult to palpate, usually requiring ultrasound guidance. However, the strong scattering and attenuation of ultrasound waves by adipose tissue results in extremely low signal-to-noise ratios in imaging deep anatomical structures, generally producing "foggy" and blurry images. The actual intervertebral space often appears as a weak hypoechoic channel, while acoustic shadows and multiple reflection artifacts caused by bone edges, calcification, or fibrous septa also appear as similarly shaped dark areas, making reliable differentiation difficult based on a single static image.

[0003] Clinical experience shows that dynamic sector scanning can distinguish between real anatomical structures and artifacts by utilizing the difference in their response to changes in probe angle—the real gap remains relatively stable under fine angle adjustments, while artifacts tend to flicker violently or disappear with changes in angle. However, this method is highly dependent on the operator's experience and hand-eye coordination, and the optimal imaging angle is often fleeting, making it difficult to accurately memorize and reproduce while keeping the probe stable. Summary of the Invention

[0004] To address the technical problem of inaccurate localization of the actual intervertebral space due to static feature confusion in low signal-to-noise ratio ultrasound images, the present invention aims to provide an anesthesia-assisted puncture localization method based on artificial intelligence. The specific technical solution adopted is as follows: The process involves: acquiring the current image sequence to be analyzed; obtaining the acoustic wave transmission drag factor of each pixel based on its grayscale value; capturing pixel displacement features between adjacent frames using an optical flow algorithm, and combining these with the pixel grayscale features to obtain the cumulative deflection angle of each historical frame relative to the current frame; and constructing a drag data volume based on the cumulative deflection angle and the drag factor, using the current frame as a reference for spatiotemporal alignment. Based on the distribution of the drag factor along the time axis within the drag data body, and combined with the spatial distribution of pixels, a corrected drag map is obtained. The minimum drag trajectory of each deep pixel in the cumulative drag map and the preset image depth is extracted from the corrected drag map. The minimum drag trajectory is projected back into each time slice within the drag data body to obtain the instantaneous total drag sequence for each deep pixel. Based on the data fluctuations within the instantaneous total drag sequence, and combined with the distribution of the cumulative deflection angle, the drag-angle change gradient is obtained. Based on the cumulative resistance map of the sound wave arrival and the resistance-angle change gradient, a puncture target probability distribution map is generated; based on the puncture target probability distribution map and the instantaneous total resistance sequence, and combined with the cumulative deflection angle, positioning guidance prompt information is generated.

[0005] Furthermore, the method for obtaining the cumulative deflection angle includes: The sequence number in the image sequence decreases over time; the pixel-level displacement field of each frame relative to the previous frame is obtained using a dense optical flow algorithm; and the corner weighting weight of each pixel is obtained based on the depth coordinates of each pixel. Based on the horizontal displacement component of each pixel in each frame of the image, the rotation angle weighting and the preset conversion coefficient, the instantaneous rotation angle of each frame of the image is obtained; based on all the instantaneous rotation angles, the cumulative deflection angle of each historical frame relative to the current frame is obtained.

[0006] Furthermore, the method for acquiring the resistance data volume includes: For each historical frame, based on the cumulative deflection angle corresponding to the historical frame, and with a preset rotation axis as a reference, the drag factor of the pixels in the historical frame is subjected to inverse rotation transformation to obtain an aligned drag slice that is aligned with the spatial coordinate system of the current frame; the drag slice formed by the drag factor of the current frame and the aligned drag slices corresponding to all historical frames are constructed into a three-dimensional drag data volume in time index order.

[0007] Furthermore, based on the modified resistance map, a fast travel algorithm is used to obtain a cumulative resistance map of sound wave arrival; for each deep pixel, gradient descent backtracking is performed on the cumulative resistance map of sound wave arrival to obtain the minimum resistance trajectory.

[0008] Furthermore, the method for obtaining the drag-angle change gradient includes: Obtain the rotation range of the cumulative deflection angle within the image sequence. When the rotation range is less than a preset rotation threshold, use the resistance-angle change gradient from the previous moment. When the turning angle range is greater than or equal to the preset turning angle threshold, the resistance-angle change gradient is obtained based on the difference between two adjacent data in the instantaneous total resistance sequence of each deep pixel and the difference between two adjacent cumulative deflection angles.

[0009] Furthermore, the method for obtaining the probability distribution map of the puncture target points includes: The statistical mean values ​​of the cumulative resistance map of the sound wave arrival and the resistance-angle change gradient are obtained separately. The accessibility probability is obtained based on the cumulative resistance map of the sound wave arrival and its statistical mean value, and the stability probability is obtained based on the resistance-angle change gradient and its statistical mean value. The accessibility probability, the stability probability, and the preset depth mask are fused and calculated to generate the puncture target point probability distribution map.

[0010] Furthermore, the method for generating location guidance prompts includes: A target guidance point is determined in the probability distribution map of the puncture target point. The moment of minimum resistance is determined according to the instantaneous total resistance sequence corresponding to the target guidance point. Positioning guidance prompt information is generated based on the cumulative deflection angle corresponding to the moment of minimum resistance.

[0011] Further, the correlation coefficient between a frame image at the moment of minimum resistance and the current frame image is obtained; when the correlation coefficient is greater than or equal to a preset recommended threshold, the difference between the cumulative deflection angle corresponding to the moment of minimum resistance and the cumulative deflection angle of the current frame is calculated as the reset guide deviation angle; Based on the distribution of the reset guide deviation angle relative to the preset dead zone range, positioning guide prompt information is generated.

[0012] Furthermore, the method for obtaining the modified resistance map includes: For each slice in the resistance data body, an arithmetic mean projection is performed along the time axis to generate an average acoustic resistance distribution map; based on the average acoustic resistance of each pixel in the average acoustic resistance distribution map, combined with the lateral deviation of each pixel from the preset central axis, a corrected resistance map is obtained.

[0013] Furthermore, the method for obtaining the resistance factor includes: When the gray value of a pixel is greater than a preset first gray value threshold, the resistance factor is set to a preset maximum constant; when the gray value of a pixel is less than or equal to the preset first gray value threshold, the resistance factor is obtained by positively mapping the gray values.

[0014] The present invention has the following beneficial effects: This invention first extracts the acoustic wave transmission resistance factor for each pixel, characterizing the acoustic wave transmission resistance at each pixel. It then captures the cumulative deflection angle of each historical frame relative to the current frame using an optical flow algorithm, constructing a spatiotemporally aligned resistance data volume to eliminate spatial misalignment caused by probe movement. A corrected resistance map is then obtained to provide a foundational field for subsequent path search. The cumulative resistance map and minimum resistance trajectory are extracted, representing the theoretically optimal puncture path from the body surface to each deep candidate point, thus identifying a clear analysis object for subsequent dynamic stability analysis. The minimum resistance trajectory is then projected back into each time slice within the resistance data volume to obtain the instantaneous total resistance sequence for each deep pixel, capturing the response characteristics of each channel under probe angle changes. Based on the data fluctuations within the instantaneous total resistance sequence and the distribution of the cumulative deflection angle, the resistance-angle change gradient is obtained, representing the sensitivity of the channel corresponding to the deep candidate point to probe attitude disturbances, providing a basis for distinguishing artifacts and generating accurate positioning guidance prompts. Finally, positioning guidance prompts are generated based on the cumulative resistance map, resistance-angle change gradient, and instantaneous total resistance sequence, combined with the cumulative deflection angle. This solution constructs a resistance-angle dynamic stability model to distinguish between the real intervertebral space and artifacts in low signal-to-noise ratio ultrasound images, generating reliable positioning guidance information and solving the technical problem of difficult puncture positioning due to poor image quality. Attached Figure Description

[0015] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating an artificial intelligence-based anesthetic puncture-assisted localization method provided in one embodiment of the present invention; Figure 2 This is a flowchart of a method for generating location guidance prompts according to an embodiment of the present invention. Detailed Implementation

[0017] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of an anesthesia puncture-assisted positioning method based on artificial intelligence proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0019] The following description, in conjunction with the accompanying drawings, details a specific scheme for an anesthesia puncture-assisted positioning method based on artificial intelligence provided by the present invention.

[0020] Please see Figure 1 The diagram illustrates a flowchart of an artificial intelligence-based anesthesia puncture-assisted localization method according to an embodiment of the present invention, specifically including: Step S1: Obtain the current image sequence to be analyzed; obtain the acoustic wave transmission drag factor of each pixel based on the gray value of each pixel in the image; capture the pixel displacement features between adjacent frames based on the optical flow algorithm, and obtain the cumulative deflection angle of each historical frame relative to the current frame by combining the gray value features of the pixels; based on the cumulative deflection angle and drag factor, perform spatiotemporal alignment with the current frame as the reference to construct the drag data volume.

[0021] In one embodiment of the present invention, in order to achieve continuous analysis of the dynamic scanning process, the system allocates a fixed-length frame queue space in memory to store ultrasound image data within a recent period. In this example, the length K of the image sequence is 20, corresponding to a scanning duration of approximately 0.67 seconds (30 frames of images are acquired), which can both cover the doctor's action cycle of fine-tuning the probe and ensure the real-time performance of the system's calculations.

[0022] The system uses a rolling update mechanism to maintain the cached image sequence. The time index k in the image queue ranges from 0 to K-1, with the latest frame corresponding to k=0. Each time a new image is added, it is inserted at the front of the existing image queue. The frame with the original index k=0 becomes k=1, the frame with the original index k=1 becomes k=2, and so on, until the last frame is removed. When the number of images collected is less than K, the system remains silent to accumulate data.

[0023] Under this setting, any first image within the image sequence Frame data All represent a lag relative to the current time. Historical data from each sampling period. Among them... For pixel coordinates, Corresponding to the horizontal scan line of the probe. Corresponding to vertical depth.

[0024] It should be noted that the sequence length K can be adjusted automatically according to the sampling frequency of the ultrasound images and the required scan duration.

[0025] In the principle of ultrasound imaging, the grayscale value (brightness) of an image reflects the echo intensity of a tissue. High-echo (bright) areas usually correspond to bone surfaces, dense fascia, or calcifications. These tissues have a strong reflection and attenuation effect on sound waves, meaning that sound waves are difficult to penetrate. Low-echo (dark) areas usually correspond to uniform soft tissue, fluid-filled dark areas, or anatomical gaps. Sound waves experience less transmission loss in these media, meaning that sound waves can easily penetrate.

[0026] To facilitate the subsequent search for the optimal acoustic channel (i.e., the low-echo concave space), the resistance factor of sound wave transmission at each pixel is first obtained based on the gray value of each pixel in the image, which characterizes the sound wave transmission resistance at each pixel. Because when doctors perform sector scans with a handheld probe, images acquired at different times correspond to different physical cross-sectional angles. To analyze tissue characteristics within a unified anatomical spatial coordinate system, this spatial misalignment caused by probe movement must be eliminated. Therefore, an optical flow algorithm is used to capture pixel displacement features between adjacent frames. Combined with the grayscale features of pixels, the cumulative deflection angle of each historical frame relative to the current frame is obtained. Based on the cumulative deflection angle and drag factor, spatiotemporal alignment is performed with the current frame as a reference to construct a drag data volume. By analyzing the relative motion of the probe, a three-dimensional data volume with strictly aligned spatial coordinates is constructed, providing a foundation for subsequent analysis.

[0027] Preferably, in one embodiment of the present invention, a highly reflective interface (such as a bone surface) in the image is first identified. The system sets a preset first grayscale threshold (a value of 220 is recommended in an 8-bit grayscale image). When the grayscale value of a pixel is greater than the preset first grayscale threshold, it indicates that it is likely hard tissue that is impermeable by sound waves, and the drag factor is set to a preset maximum constant (in this example, ). This simulates the physical phenomenon of ultrasound waves reflecting off bone and creating a sound shadow behind them, forcing subsequent path planning to avoid the bone occlusion area.

[0028] When the gray value of a pixel is less than or equal to a preset first gray value threshold, the resistance factor is obtained by positively mapping the gray value.

[0029] As an example, a non-linear positive correlation function is used to map the original grayscale value of a pixel to obtain the drag factor. The formula for calculating the drag factor includes: ; in, Indicates a lag relative to the current time. For each sampling period, the coordinates are The grayscale value of the pixel; Indicates a lag relative to the current time. For each sampling period, the coordinates are The drag factor of the pixel; Indicates the drag gain coefficient. This value is used to amplify the resistance differences between different tissues; in this example, it is set to 100. Indicates the nonlinear adjustment index. In this example, we take 1.5; The base medium resistance represents the cost per unit distance for sound waves to propagate in an ideal homogeneous medium; in this example, it is set to 1.

[0030] In the formula, since The exponential term suppresses the background noise influence of low grayscale areas (dark areas), while nonlinearly increasing the resistance value of high grayscale areas (bright areas), further highlighting the difference in sound wave transmission cost between bright and dark areas; the larger the grayscale value of a pixel, the larger the calculated resistance factor, which indicates that the sound wave at the corresponding point is more difficult to penetrate, providing a basis for finding the optimal path with the minimum "penetration resistance".

[0031] It should be noted that the ultrasound image analyzed in this embodiment is a grayscale image. The method of converting color images to grayscale is a well-known technology and will not be described in detail here. The calculation process of the drag factor for each pixel of each frame image is the same. Only one example is described here and will not be repeated. By default, the origin of the coordinate system of each frame image is the lower left corner, and a two-dimensional coordinate system is established. In other embodiments of the present invention, the implementer can adjust the coordinate system settings and the parameter values ​​in the drag factor calculation formula, which must satisfy the positive physical mapping relationship of "higher grayscale → greater drag".

[0032] Preferably, in one embodiment of the present invention, the method for obtaining the cumulative deflection angle includes: the sequence number in the image sequence decreases over time, that is, the current frame image corresponds to k=0, and the earliest historical frame image corresponds to k=K-1; The dense optical flow algorithm is used to obtain the pixel-level displacement field of each frame relative to the previous frame (e.g., calculating the k-th frame relative to the (k-1)-th frame). The pixel-level displacement field contains the displacement field of each pixel. Displacement components in the horizontal direction and vertical displacement components This exhibits pixel displacement characteristics between adjacent frames; During the tilting process, the probe's physical rotation manifests as a rigid body motion on the ultrasound image plane, with the probe's center as an approximate axis of rotation. Under small-angle approximation conditions, this rotation is primarily characterized by horizontal pixel displacement, and the displacement amplitude is directly proportional to the pixel depth—that is, the deeper the tissue, the greater the horizontal displacement. Therefore, by analyzing the horizontal displacement field between adjacent frames, the instantaneous physical rotation angle of the probe at that moment can be deduced.

[0033] In ultrasound images of the back of obese patients, the superficial fat layer is usually physically compressed by the probe, appearing relatively static or slightly deformed in the image; while the deep anatomical structures sway significantly with the movement of the probe. If motion calculations are performed directly on the entire image, the large static areas on the surface will "dilute" the true motion signal from the deeper tissues, leading to an underestimation of the angles. Therefore, based on the depth coordinates of each pixel, a rotational weighting weight is obtained for each pixel, introducing a depth weighting mechanism to extract the motion information of the deep tissues.

[0034] As an example, normalizing the y-coordinate linearly (e.g., dividing by 255) yields the normalized depth coordinate. ,Will With preset depth gain coefficient The product of these factors, plus a constant 1, is used as the corner weighting factor. The constant 1 is used to adjust the value range to ensure the gain effect. Furthermore, based on the horizontal displacement component, rotation weight, and preset conversion coefficient of each pixel in each frame image, the instantaneous rotation angle of each frame image is obtained; based on all instantaneous rotation angles, the cumulative deflection angle of each historical frame relative to the current frame is obtained.

[0035] As an example, for any frame image k, the sum of the rotation weights of all pixels is used as the denominator. The product of the horizontal displacement component of each pixel and its corresponding rotation weight is summed and used as the numerator. The product of the ratio of the fractions and the preset conversion coefficient is used as the instantaneous rotation angle of the corresponding frame image. .

[0036] The corner weighting is not less than 1, and the denominator is not 0. The average displacement of the whole image is calculated by weighted averaging to obtain the representative displacement of the whole image, and then mapped to the physical angle by a preset conversion coefficient. The unit of the preset conversion coefficient is degrees / pixel, which can be obtained by scanning and calibrating a standard model at a known angle, or preset according to the probe frequency and scanning depth parameters. For example, the preset conversion coefficient = 1 / (image center depth × pixel density), and the empirical value is 0.05~0.1 degrees / pixel.

[0037] Finally, based on the index k of each historical frame image, the instantaneous rotation angles from 0 to k are accumulated, and the result is used as the cumulative deflection angle of each historical frame image.

[0038] It should be noted that the cumulative deflection angle of the current frame does not need to be calculated and is directly set to 0. The dense optical flow algorithm is used to analyze the displacement changes between two frames of images, providing a continuous and dense displacement vector for each pixel in the image, thereby extracting the pixel-level displacement field. This is a well-known motion estimation technique in the field. Before optical flow analysis, noise reduction filtering (such as Gaussian filtering) can be performed on the original image to remove speckle noise, and histogram equalization or edge enhancement can be performed to highlight texture features. The preprocessing methods are all conventional techniques in the field of image enhancement and will not be elaborated further.

[0039] After obtaining the precise angle deviation for each frame, the system performs an inverse spatial transformation to unify all historical data to the current coordinate system.

[0040] Preferably, in one embodiment of the present invention, for each historical frame, based on the cumulative deflection angle corresponding to the historical frame and with a preset rotation axis as a reference, the drag factor of the pixels in the historical frame is subjected to inverse rotation transformation to obtain an aligned drag slice that is aligned with the spatial coordinate system of the current frame; the drag slice formed by the drag factor of the current frame and the aligned drag slices corresponding to all historical frames are constructed into a three-dimensional drag data volume according to the time index order. .

[0041] In this example, the system uses the center of the top of the image (i.e. the center point of the probe surface) as the preset rotation axis, which is the mapping position of the origin of the ultrasonic probe's sound beam emission in the image coordinate system (for example, for a sector scan probe, the center of the top of the image or the virtual vertex above); for areas that exceed the image boundary of the current frame after rotation, a preset maximum constant is filled in.

[0042] After spatiotemporal alignment, the spatial misalignment caused by probe swaying was eliminated, and the resistance data volume was... In the vector along the time axis, the acoustic wave transmission resistance of the tissue at the same physical location changes over a period of time, providing a direct data basis for subsequent dynamic stability analysis.

[0043] Step S2: Based on the distribution of drag factors along the time axis within the drag data body, combined with the spatial distribution of pixels, obtain a corrected drag map; extract the minimum drag trajectory of each deep pixel in the cumulative drag map and the preset image depth from the corrected drag map; project the minimum drag trajectory back into each time slice within the drag data body to obtain the instantaneous total drag sequence of each deep pixel; based on the data fluctuations within the instantaneous total drag sequence, combined with the distribution of the cumulative deflection angle, obtain the drag-angle change gradient.

[0044] Because single-frame images are affected by speckle noise and transient artifacts, directly using them for path planning can easily produce non-physical paths. Therefore, based on the distribution of resistance factors along the time axis within the resistance data body, we analyze along the time axis to suppress random fluctuations. In conjunction with the spatial distribution of pixels, we introduce spatial constraints to obtain a corrected resistance map, providing a basic field for subsequent path search.

[0045] Since the target structures for spinal canal puncture (such as the epidural space and subarachnoid space) are located deep within the spine, and superficial tissues (such as subcutaneous fat and muscle layer) are only areas traversed by the puncture path, rather than the target points, the minimum resistance trajectory of each deep pixel in the cumulative resistance map and the preset image is further extracted from the modified resistance map. This represents the theoretically optimal puncture path from the body surface to each deep candidate point, thus identifying a clear analysis object for subsequent dynamic stability analysis.

[0046] Preferably, in one embodiment of the present invention, the arithmetic mean projection of each slice in the resistance data body along the time axis is first performed to generate an average acoustic resistance distribution map, which can suppress random noise and transient artifacts and highlight the stable low-resistance channels.

[0047] The sound beam emitted by an ultrasonic probe is approximately symmetrical about the probe's center, with the effective acoustic energy mainly concentrated in the central region. When constructing a penetration resistance model, relying solely on the average resistance value ignores the geometric characteristics of sound beam propagation—that is, a path deviating from the central axis is actually equivalent to a longer oblique penetration distance, and its physical transmission cost should be higher.

[0048] Therefore, by superimposing a penalty term proportional to the lateral offset on the average acoustic drag, the actual cumulative drag distribution of sound waves propagating from the probe's center source to deeper layers can be more accurately reflected. This makes the corrected drag map conform to the physical laws of acoustic propagation. Thus, based on the average acoustic drag of each pixel in the average acoustic drag distribution map, combined with the lateral deviation of each pixel from the preset central axis, the corrected drag map is obtained. .

[0049] As an example, the arithmetic mean projection is: the drag factor at the same coordinate point in all time slices of the drag data body is taken as the arithmetic mean along the time axis, and used as the data value (average acoustic drag) at the corresponding position in the average acoustic drag distribution map. The preset central axis is the horizontal coordinate of the sound wave emission source, corresponding to the center column of the image. The absolute value of the difference between the horizontal coordinate of each pixel and the horizontal coordinate of the preset central axis (the center column of the image) is divided by the pixel width of the image (the number of pixels in the horizontal direction) as the horizontal deviation. The product of the horizontal deviation and the preset horizontal deviation penalty coefficient is used as the horizontal deviation penalty term. The sum of the horizontal deviation penalty term of each pixel in the average acoustic drag distribution map and the average acoustic drag is used as the corresponding data value in the corrected drag map.

[0050] The preset lateral deviation penalty coefficient is set to 5~20, with the same dimensions as the drag factor. In this example, it is set to 10, which increases the drag in the edge region by about 5 units. Without obscuring the original acoustic information, it effectively guides the path of least drag to shift towards the center region of the probe, which is consistent with the physical characteristics of the ultrasonic beam energy distribution.

[0051] Furthermore, the modified resistance map is used as the local propagation cost field, and the equations are solved using the Fast Marching Method (FMM). The algorithm simulates the wavefront from the body surface ( During the downward propagation process, the output sound wave reaches the cumulative resistance diagram. The value of each pixel in the image represents the minimum cumulative energy loss required for a sound wave to travel from the body surface to that point along the optimal path.

[0052] The preset image depth refers to the region where the image depth y is greater than the preset depth threshold. Pixels in this region are deep pixels. In this example, the preset depth threshold is 150 pixels, which corresponds to a physical depth of about 3 cm in a clinical ultrasound image (the vertical resolution of the image is 50 pixels / cm), covering the anatomical layers of typical adult spinal canal puncture target structures (such as the ligamentum flavum and epidural space).

[0053] Finally, for each deep pixel p, gradient descent backtracking is performed on the cumulative resistance map of the sound wave arrival to obtain the minimum resistance trajectory.

[0054] Specifically, the system in Starting from p, along - The gradient descent direction is used to backtrack until the starting region on the body surface is reached, thus obtaining a unique optimal path from the body surface to p, denoted as the minimum resistance trajectory. This trajectory corresponds to the geometric channel with the lowest sound wave penetration cost, providing a physical path basis for subsequent dynamic stability analysis.

[0055] It should be noted that methods such as using the fast traversal algorithm to solve the equation to generate the cumulative cost map and using gradient field for path backtracking are all well-known techniques in the field. In other embodiments of the present invention, the implementer can adjust the preset depth threshold according to the vertical resolution of the image, which will not be described in detail here.

[0056] Furthermore, the minimum resistance trajectory is projected back into each time slice within the resistance data body. By recalculating the total resistance of the same path in each historical time slice, the instantaneous total resistance sequence of each deep pixel is obtained, forming a resistance sequence that changes over time, thereby capturing the response characteristics of each channel under changes in probe angle.

[0057] Considering that the real anatomical channel has structural stability during probe scanning, its acoustic wave penetration resistance changes slowly with the angle, while the artifact region experiences drastic fluctuations in resistance value due to specific incident conditions, the resistance-angle change gradient is obtained based on the data fluctuations within the instantaneous total resistance sequence and the distribution of cumulative deflection angle. This gradient quantifies the rate of change of total path resistance caused by a unit angle offset, representing the sensitivity of the channel corresponding to the deep candidate point to probe attitude disturbances, and provides a basis for distinguishing artifacts and generating accurate positioning guidance cues.

[0058] Preferably, in one embodiment of the present invention, the minimum resistance trajectory is projected back into each time slice of the resistance data body. That is, for each depth pixel p, the minimum resistance trajectory of the selected pixel in each time slice is selected, and the sum of the resistance factors of the selected pixels in the time slice is taken as the instantaneous total resistance. The instantaneous total resistance is sorted according to the order of the time slices to form an instantaneous total resistance sequence.

[0059] Considering that if the operator holds the probe still, the angle change will approach zero, and directly calculating the gradient will lead to numerical explosion, the rotation range of the cumulative deflection angle in the image sequence is obtained first. When the rotation range is less than the preset rotation threshold, the resistance-angle change gradient of the previous moment is used. When the turning angle range is greater than or equal to the preset turning angle threshold, in order to measure the sensitivity of resistance to changes in angle, the resistance-angle change gradient is obtained based on the difference between two adjacent data in the instantaneous total resistance sequence of each deep pixel and the difference between two adjacent cumulative deflection angles.

[0060] As an example, the preset turning angle threshold is 1 degree, and the initial default resistance-angle change gradient is 0. For a deep pixel p, the sum of the absolute values ​​of the differences between the instantaneous total resistance of all adjacent frames within the instantaneous total resistance sequence is used as the numerator. The sum of the absolute values ​​of the differences between the cumulative deflection angles corresponding to all adjacent frames, plus the sum after a preset positive parameter divided by zero, is used as the denominator. The ratio of the fractions is used as the resistance-angle change gradient of the deep pixel p.

[0061] In this example, the preset division by zero positive parameter is taken as... The unit is the same as the denominator. The drag-angle change gradient quantifies the amount of drag change caused by a unit angle change. The smaller the value, the less sensitive the corresponding channel is to the incident angle, and it is highly likely to be an isotropic true anatomical gap; the larger the value, the more dependent the corresponding channel is on a specific incident angle, and it is highly likely to be an anisotropic acoustic artifact.

[0062] The data fluctuations within the instantaneous total drag sequence are represented by the sum of the absolute values ​​of the differences between drag values ​​between adjacent frames in the instantaneous total drag sequence, while the distribution of the cumulative deflection angle is represented by the sum of the absolute values ​​of the angle range and the differences between the cumulative deflection angles between adjacent frames.

[0063] It should be noted that the analysis process is the same for each deep pixel. Only one example is described here, and it will not be repeated.

[0064] Step S3: Generate a puncture target probability distribution map based on the cumulative resistance map of sound wave arrival and the resistance-angle change gradient; generate positioning guidance prompt information based on the puncture target probability distribution map and the instantaneous total resistance sequence, combined with the cumulative deflection angle.

[0065] Considering that the cumulative resistance map of acoustic wave arrival reflects the energy loss of acoustic wave penetration into tissue, and the resistance-angle change gradient reflects the physical stability of the target structure as the probe angle changes, a probability distribution map of puncture target points is generated based on the cumulative resistance map of acoustic wave arrival and the resistance-angle change gradient. This quantifies the possibility of each location in the image being a real anatomical gap, effectively eliminating realistic acoustic artifacts while locking in the target area with the best acoustic wave penetration and structural stability, thereby improving the accuracy and robustness of positioning.

[0066] Considering that the instantaneous total resistance sequence records the changes in the accessibility of the target path at historical moments, and the cumulative deflection angle records the historical trajectory of the probe attitude, positioning guidance information is generated based on the probability distribution map of the puncture target point and the instantaneous total resistance sequence, combined with the cumulative deflection angle. This transforms the complex image recognition results into intuitive positioning guidance information, which assists in the accurate positioning of the real intervertebral space under imaging conditions with extremely low signal-to-noise ratio and severe artifact interference.

[0067] Preferably, in one embodiment of the present invention, a single indicator (such as low resistance only or high stability only) is insufficient to reliably identify the puncture target. The actual intervertebral space must simultaneously meet two conditions: first, the sound waves can easily penetrate (high accessibility), and second, it is not sensitive to changes in probe angle (high stability). Since cumulative resistance (cumulative energy loss) and the gradient of change (rate of change) belong to different physical dimensions, and the differences in the physical condition of different patients will lead to fluctuations in the numerical range, the statistical mean values ​​of the cumulative resistance map and the resistance-angle change gradient are obtained separately before statistical analysis. Then, the accessibility probability is obtained based on the cumulative resistance map of sound wave arrival and its statistical mean, and the stability probability is obtained based on the resistance-angle change gradient and its statistical mean. As an example, the statistical mean is used as the outlier, eliminating the interference of extreme values ​​such as strong bone reflection and liquid dark areas, so that the statistical mean can more accurately reflect the baseline resistance level of the background soft tissue; in the cumulative resistance map of sound wave arrival, the average value after removing the highest and lowest 10% of data is used as the statistical mean of the cumulative resistance map of sound wave arrival; in the cumulative resistance map of sound wave arrival, the average value after removing the highest and lowest 10% of data is used as the statistical mean of the resistance-angle change gradient. Accessibility probability The probability of accessibility reflects the magnitude of path resistance; the smaller the resistance (the easier it is to reach), the closer the probability value is to 1.

[0068] Stability probability The stability probability reflects the stability of the channel; the smaller the gradient (the more stable), the closer the probability value is to 1.

[0069] in, The coordinates of the cumulative resistance diagram for sound waves are: The data values ​​of the data points; This represents the statistical mean of the cumulative resistance diagram for sound wave arrival. The coordinates in the resistance data volume are The resistance-angle change gradient of the data points; This represents the statistical mean of the resistance-angle change gradient. This is the first sensitivity adjustment coefficient. This is the second sensitivity adjustment coefficient. This represents the preset positive parameter divided by zero; in this example, its value is [value to be filled in]. .

[0070] It should be noted that the sensitivity adjustment coefficient is used to control the decay rate of probability as the value changes. In this example, the value is always 1.0, indicating that the average statistical level of the background tissue is used as the benchmark unit for probability decay, thereby achieving a normalized evaluation relative to the background level. Since the drag-angle change gradient is not an instantaneous quantity for each frame, but a global, aggregated feature quantity calculated based on the dynamic response of the same coordinate point within the drag data body, it can be used for each spatial coordinate. The corresponding resistance-angle change gradient forms a complete two-dimensional gradient graph. .

[0071] In order to eliminate interference from the superficial fat layer and focus on the deep anatomical structure, a depth mask is introduced to exclude superficial non-target areas. Therefore, the accessibility probability, stability probability and the preset depth mask are fused and calculated to generate a puncture target probability distribution map.

[0072] As an example, a preset depth mask Including: when y is less than a preset depth threshold of 150 pixels, When y is greater than or equal to the preset depth threshold of 150 pixels, ; to use the same spatial coordinates , and The product of these factors, taken as the probability of the puncture target, forms the probability distribution map of the puncture target. ; fusion is achieved through multiplication.

[0073] In the probability distribution map of puncture target points, the regions with values ​​close to 1 are more likely to simultaneously meet the two core conditions of "easy sound wave access (low echo)" and "physical stability (no artifacts)". Therefore, in determining the target guidance point in the probability distribution map of puncture target points, the spatial location with the largest value is selected and locked as the target guidance point. The minimum resistance moment is determined based on the instantaneous total resistance sequence corresponding to the target guide point. Finally, positioning guidance information is generated based on the cumulative deflection angle corresponding to the minimum resistance moment.

[0074] It should be noted that the moment of minimum resistance is the frame slice with the smallest data value in the instantaneous total resistance sequence corresponding to the target guidance point. When there are multiple minimum values, the latest frame slice is taken. The moment of minimum resistance physically corresponds to the historical moment when the soft tissue gap is most open to sound waves and the bone blockage is minimal, i.e., the "optimal acoustic window moment".

[0075] Before locking onto the target guidance point, an effectiveness determination is performed: when the probability value of the largest value is less than the preset effectiveness threshold (e.g., 0.6), it is determined that there is no effective target in the current field of view, no further guidance is performed, and no positioning guidance prompt information is generated; the threshold of 0.6 is set to ensure that guidance is triggered only when the accessibility and stability of the target point are significantly better than the background level, so as to avoid generating misleading prompts in low signal-to-noise ratio areas.

[0076] In other embodiments of the present invention, the implementer may adjust the first and second sensitivity adjustment coefficients and simultaneously adjust the preset effectiveness threshold.

[0077] Please see Figure 2 The flowchart illustrates a method for generating location guidance prompts according to an embodiment of the present invention, specifically including: Step S301: Obtain the correlation coefficient between a frame image at the moment of minimum resistance and the current frame image; when the correlation coefficient is greater than or equal to the preset recommended threshold, calculate the difference between the cumulative deflection angle corresponding to the moment of minimum resistance and the cumulative deflection angle of the current frame, and use it as the reset guide deviation angle.

[0078] Considering that when the operator moves the probe too fast, the current cross-section may have changed significantly in spatial position (translation) from the historical best cross-section. At this time, simply adjusting the angle is meaningless. Therefore, we obtain the correlation coefficient between the image at the moment of minimum resistance and the current frame image, and analyze the correlation between the image at the moment of minimum resistance and the current frame image.

[0079] As an example, the Pearson correlation coefficient between the pixel grayscale values ​​of a frame at the moment of minimum resistance and the current frame within the effective imaging region (y is greater than the preset depth threshold of 150 pixels) is used as the correlation coefficient.

[0080] When the correlation coefficient is less than the preset recommended threshold, it indicates that the cross-section has changed, and the system will not output positioning guidance information, remaining silent or prompting "Please rescan"; in this example, the preset recommended threshold is 0.6; When the correlation coefficient is greater than or equal to the preset recommended threshold, it indicates that the current cut surface and the historical best cut surface maintain a high degree of spatial consistency in terms of anatomical structure, with only angular deviation. In order to transform the identified historical best state into an intuitive positioning guidance prompt, the difference between the cumulative deflection angle corresponding to the moment of minimum resistance and the cumulative deflection angle of the current frame is used as the reset guidance deviation angle, providing a basis for the subsequent generation of positioning guidance information.

[0081] It should be noted that 0.6 is based on statistical calibration of clinical ultrasound scan data: in repeated scans with small angle adjustments on the same intervertebral space, the Pearson correlation coefficient of the effective imaging area is generally higher than 0.7; while when the probe is significantly translated or the sectional plane is switched to an adjacent segment, the correlation coefficient is usually lower than 0.5; therefore, taking 0.6 as an intermediate safety threshold can effectively exclude mismatches due to anatomical sectional plane shift while ensuring high recall.

[0082] Step S302: Generate positioning guidance prompt information based on the distribution of the reset guide deviation angle relative to the preset dead zone range.

[0083] As an example, the location guidance information includes icons and arrows, with a preset dead zone range of [missing information]. When the reset guide deviation corner is within the preset dead zone range, a green "Stay" or "Optimal" icon is displayed; when the reset guide deviation corner exceeds the upper limit of the preset dead zone range (+ When the angle of the reset guide deviation exceeds the lower limit of the preset dead zone range (-), a dynamic arrow is displayed. When ), a dynamic arrow is displayed that tilts in a negative angle direction (such as to the left).

[0084] It should be noted that the preset dead zone range is set based on clinical operating experience and system sensitivity requirements: this range is slightly larger than the natural posture fluctuations when the operator holds the probe, to avoid frequent changes in prompt information due to slight shaking; at the same time, it ensures that angle deviations exceeding this range can be effectively identified and guided for correction. The specific threshold can be determined through offline calibration or user adaptation adjustment, which is not the focus of this invention and will not be elaborated further. In other embodiments of this invention, it can be adjusted according to the device accuracy, user habits, or anatomical location.

[0085] In summary, to address the technical problem of inaccurate localization of the true intervertebral space due to static feature confusion in low signal-to-noise ratio ultrasound images, this invention provides an artificial intelligence-based anesthesia puncture-assisted localization method. This invention first extracts the acoustic wave transmission resistance factor for each pixel; then, based on an optical flow algorithm, it captures the cumulative deflection angle of each historical frame relative to the current frame, and constructs a spatiotemporally aligned resistance data volume; further, it obtains a corrected resistance map, extracts the acoustic wave arrival cumulative resistance map and the minimum resistance trajectory; further, it projects the minimum resistance trajectory back into each time slice within the resistance data volume to obtain the instantaneous total resistance sequence for each deep pixel; further, based on the data fluctuations within the instantaneous total resistance sequence and the distribution of the cumulative deflection angle, it obtains the resistance-angle change gradient; finally, based on the acoustic wave arrival cumulative resistance map, the resistance-angle change gradient, and the instantaneous total resistance sequence, it generates localization guidance information in conjunction with the cumulative deflection angle. This scheme, by constructing a resistance-angle dynamic stability model, distinguishes the true intervertebral space from artifacts in low signal-to-noise ratio ultrasound images, generates reliable localization guidance information, and solves the technical problem of difficult puncture localization due to poor image quality.

[0086] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0087] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. An artificial intelligence-based anesthesia puncture auxiliary positioning method, characterized in that, The method includes: The process involves: acquiring the current image sequence to be analyzed; obtaining the acoustic wave transmission drag factor for each pixel based on its grayscale value; capturing pixel displacement features between adjacent frames using an optical flow algorithm to obtain the cumulative deflection angle of each historical frame relative to the current frame; and constructing a drag data volume based on the cumulative deflection angle and the drag factor, using the current frame as a reference for spatiotemporal alignment. Based on the distribution of the drag factor along the time axis within the drag data body, and combined with the spatial distribution of pixels, a corrected drag map is obtained. The minimum drag trajectory of each deep pixel in the cumulative drag map and the preset image depth is extracted from the corrected drag map. The minimum drag trajectory is projected back into each time slice within the drag data body to obtain the instantaneous total drag sequence for each deep pixel. Based on the data fluctuations within the instantaneous total drag sequence, and combined with the distribution of the cumulative deflection angle, the drag-angle change gradient is obtained. Based on the cumulative resistance map of the sound wave arrival and the resistance-angle change gradient, a puncture target probability distribution map is generated; based on the puncture target probability distribution map and the instantaneous total resistance sequence, and combined with the cumulative deflection angle, positioning guidance prompt information is generated; The method for obtaining the resistance factor includes: when the gray value of a pixel is greater than a preset first gray value threshold, setting the resistance factor to a preset maximum constant; when the gray value of a pixel is less than or equal to the preset first gray value threshold, obtaining the resistance factor by positively correlated mapping of the gray values. The method for obtaining the resistance data volume includes: for each historical frame, based on the cumulative deflection angle corresponding to the historical frame, taking a preset rotation axis as a reference, performing an inverse rotation transformation on the resistance factor of the pixels in the historical frame to obtain an aligned resistance slice aligned with the spatial coordinate system of the current frame; and constructing a three-dimensional resistance data volume by combining the slice formed by the resistance factor of the current frame with the aligned resistance slices corresponding to all historical frames in time index order. The method for obtaining the cumulative deflection angle includes: the sequence number in the image sequence decreases over time; the pixel-level displacement field of each frame image relative to the previous frame image is obtained using a dense optical flow algorithm; the rotation angle weighting weight of each pixel is obtained based on the depth coordinates of each pixel; the instantaneous rotation angle of each frame image is obtained based on the horizontal displacement component of each pixel in each frame image, the rotation angle weighting weight, and a preset conversion coefficient; and the cumulative deflection angle of each historical frame relative to the current frame is obtained based on all the instantaneous rotation angles.

2. The artificial intelligence-based anesthesia puncture auxiliary positioning method according to claim 1, characterized in that, Based on the modified resistance map, a fast travel algorithm is used to obtain the cumulative resistance map of sound wave arrival; for each deep pixel, gradient descent backtracking is performed on the cumulative resistance map of sound wave arrival to obtain the minimum resistance trajectory.

3. The anesthesia puncture-assisted positioning method based on artificial intelligence according to claim 1, characterized in that, The method for obtaining the resistance-angle change gradient includes: Obtain the rotation range of the cumulative deflection angle within the image sequence. When the rotation range is less than a preset rotation threshold, use the resistance-angle change gradient from the previous moment. When the turning angle range is greater than or equal to the preset turning angle threshold, the resistance-angle change gradient is obtained based on the difference between two adjacent data in the instantaneous total resistance sequence of each deep pixel and the difference between two adjacent cumulative deflection angles.

4. The anesthesia puncture-assisted positioning method based on artificial intelligence according to claim 1, characterized in that, The method for obtaining the probability distribution map of the puncture target points includes: The statistical mean values ​​of the cumulative resistance map of the sound wave arrival and the resistance-angle change gradient are obtained separately. The accessibility probability is obtained based on the cumulative resistance map of the sound wave arrival and its statistical mean value, and the stability probability is obtained based on the resistance-angle change gradient and its statistical mean value. The accessibility probability, the stability probability, and the preset depth mask are fused and calculated to generate the puncture target point probability distribution map.

5. The anesthesia puncture-assisted positioning method based on artificial intelligence according to claim 1, characterized in that, The method for generating location guidance prompts includes: A target guidance point is determined in the probability distribution map of the puncture target point. The moment of minimum resistance is determined according to the instantaneous total resistance sequence corresponding to the target guidance point. Positioning guidance prompt information is generated based on the cumulative deflection angle corresponding to the moment of minimum resistance.

6. The artificial intelligence-based anesthesia puncture-assisted positioning method according to claim 5, characterized in that, Obtain the correlation coefficient between a frame image at the moment of minimum resistance and the current frame image; when the correlation coefficient is greater than or equal to a preset recommended threshold, calculate the difference between the cumulative deflection angle corresponding to the moment of minimum resistance and the cumulative deflection angle of the current frame, and use it as the reset guide deviation angle; Based on the distribution of the reset guide deviation angle relative to the preset dead zone range, positioning guide prompt information is generated.

7. The anesthesia puncture-assisted positioning method based on artificial intelligence according to claim 1, characterized in that, The method for obtaining the modified resistance map includes: For each slice in the resistance data body, an arithmetic mean projection is performed along the time axis to generate an average acoustic resistance distribution map; based on the average acoustic resistance of each pixel in the average acoustic resistance distribution map, combined with the lateral deviation of each pixel from the preset central axis, a corrected resistance map is obtained.

Images