An intelligent evaluation method and system for ultrasound operation based on contrast timing table characterization

Abstract

The application relates to the field of medical image processing and computer vision technology, and particularly discloses an ultrasound operation intelligent evaluation method and system based on contrast timing representation, which extracts an anatomic structure contour and generates a three-dimensional space-time evolution track of a target tissue by fusing interframe pixel change features of an ultrasound contrast video stream and probe spatial pose data; the application extracts a perfusion timing representation parameter by deeply analyzing contrast agent perfusion intensity change, and fuses the parameter with an initial contrast space-time feature sequence to form a contrast timing representation vector; the application uses a convolution layer and a recurrent neural network layer to mine deep data patterns of operation behaviors, outputs an ultrasound operation stability score and a section offset correction suggestion, and generates an intelligent evaluation report; the application overcomes the limitations of traditional technologies, realizes objective and quantitative evaluation of ultrasound operation stability, provides strong support for medical image processing and clinical skill training, and improves the precision and practicability of evaluation.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing and computer vision technology, and relates to an intelligent evaluation method and system for ultrasound operation based on contrast imaging time sequence characterization. Background Technology

[0002] Contrast-enhanced ultrasound imaging is a medical image analysis technique that utilizes microbubbles to enhance backscattered signals, providing information on dynamic blood flow and microvascular perfusion within tissues. In recent years, with the rapid popularization of fifth-generation mobile communication technology (5G) and the widespread application of various intelligent medical terminals (such as portable ultrasound terminals and remote cart terminals) in primary healthcare institutions, remote ultrasound diagnosis and large-scale clinical skills training have experienced rapid development.

[0003] However, the quality of ultrasound image acquisition is highly "operator-dependent," heavily reliant on the scanning techniques and probe control capabilities of the on-site operator. In 5G-enabled telemedicine and mobile diagnostic scenarios, the lack of on-site "hands-on" guidance from senior experts presents a significant challenge. Therefore, how to leverage intelligent algorithms to objectively evaluate the ultrasound procedure at the medical terminal and provide real-time corrective guidance has become a critical issue that urgently needs to be addressed in the fields of medical image processing and smart healthcare.

[0004] Existing technologies based on various ultrasound terminals typically rely on extracting static image features from standard sections or using basic motion tracking algorithms to monitor the physical movement trajectory of the probe in isolation. Some systems assess image quality solely by analyzing the grayscale distribution of a single frame or by performing a static similarity comparison between the currently acquired two-dimensional section and a pre-stored standard anatomical template to determine whether the scanned section conforms to specifications.

[0005] The aforementioned existing technical solutions often completely separate image feature analysis from probe spatial pose tracking, failing to construct a unified three-dimensional spatiotemporal representation model. Particularly during ultrasound contrast imaging, these methods neglect the dynamic microcirculation perfusion characteristics of the contrast agent over time, lacking the image analysis capability to deeply integrate tissue morphological deformation with physiological perfusion changes. Due to the failure to fully explore multi-dimensional spatiotemporal evolution patterns, existing terminal-assisted systems struggle to accurately quantify the stability of the scanning process, and are unable to provide operators with precise, spatially guiding suggestions for section offset correction based on complex dynamic tissue changes. Summary of the Invention

[0006] In view of this, in order to solve the problems mentioned in the background art, a method and system for intelligent evaluation of ultrasound operation based on contrast imaging time sequence characterization is proposed.

[0007] The objective of this invention can be achieved through the following technical solution: The first aspect of this invention provides an intelligent evaluation method for ultrasound operation based on contrast imaging temporal characterization, including: S1, acquiring ultrasound contrast imaging video stream and corresponding probe spatial pose data, extracting inter-frame pixel change features of ultrasound contrast imaging video stream, and generating an initial contrast imaging spatiotemporal feature sequence.

[0008] S2. Extract the anatomical structure contour from the initial spatiotemporal feature sequence of the contrast imaging to obtain the dynamic boundary features of the target tissue, and perform spatial mapping in combination with the probe spatial pose data to generate the three-dimensional spatiotemporal evolution trajectory of the target tissue.

[0009] S3. Analyze the changes in contrast agent perfusion intensity in the three-dimensional spatiotemporal evolution trajectory of the target tissue, extract the perfusion time series characterization parameters, and perform feature fusion with the initial contrast spatiotemporal feature sequence to generate the contrast time series characterization vector.

[0010] S4. Input the contrast time sequence representation vector into the operation evaluation model to extract quality features, output the ultrasound operation stability score and section deviation correction suggestions, and generate an intelligent ultrasound operation evaluation report.

[0011] The second aspect of the present invention provides an intelligent evaluation system for ultrasound operation based on contrast imaging temporal characterization, comprising: an initial contrast imaging spatiotemporal feature generation module, which acquires ultrasound contrast imaging video stream and corresponding probe spatial pose data, extracts inter-frame pixel change features of ultrasound contrast imaging video stream, and generates an initial contrast imaging spatiotemporal feature sequence.

[0012] The three-dimensional spatiotemporal evolution trajectory generation module extracts the anatomical structure contour from the initial contrast spatiotemporal feature sequence to obtain the dynamic boundary features of the target tissue, and performs spatial mapping in combination with probe spatial pose data to generate the three-dimensional spatiotemporal evolution trajectory of the target tissue.

[0013] The contrast time sequence characterization vector generation module analyzes the changes in contrast agent perfusion intensity in the three-dimensional spatiotemporal evolution trajectory of the target tissue, extracts perfusion time sequence characterization parameters, and performs feature fusion with the initial contrast spatiotemporal feature sequence to generate a contrast time sequence characterization vector.

[0014] The ultrasound operation assessment report generation module inputs the contrast time sequence characterization vector into the operation assessment model to extract quality features, outputs an ultrasound operation stability score and a section deviation correction suggestion, and generates an intelligent ultrasound operation assessment report.

[0015] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects: (1) This invention extracts the inter-frame pixel change features of the ultrasound contrast video stream and combines them with the probe spatial pose data to perform edge detection and morphological processing on the initial contrast spatiotemporal feature sequence, thereby obtaining the anatomical structure outline; this multi-dimensional image analysis method transforms the dynamic boundary features of the target tissue into the world coordinate system for spatial mapping, generating the three-dimensional spatiotemporal evolution trajectory of the target tissue, overcoming the limitations of static two-dimensional image analysis, and laying a solid data foundation for subsequent operational evaluation; (2) This invention deeply analyzes the changes in contrast agent perfusion intensity in the three-dimensional spatiotemporal evolution trajectory of the target tissue. By calculating the rate of change of voxel gray value with time and performing curve fitting, the peak time and peak intensity are extracted as perfusion time sequence characterization parameters. The perfusion time sequence characterization parameters are combined with the initial contrast spatiotemporal feature sequence through channel splicing and attention weight allocation, realizing the deep fusion of multi-source features and comprehensively capturing the dynamic physiological changes and spatial deformation laws inside the tissue. (3) The present invention inputs the fused contrast time sequence representation vector into the operation evaluation model for quality feature extraction, and uses convolutional layers and recurrent neural network layers to mine deep data patterns that reflect probe holding stability and scanning continuity; after classifying and regressing the operation behavior feature map, the system outputs ultrasound operation stability score and section offset correction suggestions, transforming subjective clinical scanning experience into objective quantitative evaluation indicators to guide the operator to accurately adjust the probe position and posture. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments 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.

[0017] Figure 1 This is a schematic diagram of the method steps of the present invention.

[0018] Figure 2 This is a schematic diagram of the system structure connection of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1The first aspect of the present invention provides an intelligent evaluation method and system for ultrasound operation based on contrast imaging temporal characterization, comprising: S1, acquiring ultrasound contrast imaging video stream and corresponding probe spatial pose data, extracting inter-frame pixel change features of ultrasound contrast imaging video stream, and generating an initial contrast imaging spatiotemporal feature sequence.

[0021] In a specific embodiment of the present invention, the acquisition of ultrasound contrast imaging video stream and corresponding probe spatial pose data, extraction of inter-frame pixel change features of ultrasound contrast imaging video stream, and generation of initial contrast imaging spatiotemporal feature sequence include: acquiring continuous ultrasound image frames output by ultrasound probe and pose coordinates output by spatial positioning sensor, and constructing ultrasound contrast imaging video stream and probe spatial pose data.

[0022] In a specific embodiment of the present invention, the continuous ultrasound image frames output by the ultrasound probe and the pose coordinates output by the spatial positioning sensor are collected to construct an ultrasound contrast imaging video stream and probe spatial pose data, including: receiving the analog signal output by the ultrasound probe in contrast imaging mode and performing analog-to-digital conversion to obtain continuous ultrasound image frames.

[0023] The translation and rotation of the spatial positioning sensor fixed on the ultrasonic probe in three-dimensional space are read to obtain the pose coordinates.

[0024] By aligning the timestamps of continuous ultrasound image frames with their pose coordinates and packaging the data, an ultrasound contrast video stream and probe spatial pose data are constructed.

[0025] Calculate the pixel motion trajectory distribution of adjacent consecutive ultrasound image frames in an ultrasound contrast video stream and extract inter-frame pixel change features.

[0026] The inter-frame pixel change features are serialized and stitched together according to the timestamp order to generate the initial imaging spatiotemporal feature sequence.

[0027] Specifically, the system receives the analog signal output by the ultrasound probe in contrast-enhanced mode and performs analog-to-digital conversion to obtain continuous ultrasound image frames. Here, the analog signal represents the continuous voltage change waveform converted from the reflected echo from human tissue received by the ultrasound probe, while the analog-to-digital conversion is the process of discretizing the continuous voltage change waveform into a digital pixel matrix. The set of two-dimensional pixel matrices arranged chronologically after analog-to-digital conversion constitutes the continuous ultrasound image frames. The conversion relationship between the analog signal and the continuous ultrasound image frames is established using the following formula: ; in the formula Representing time Continuous ultrasound image frames, dimensionless. Representing time The analog signal is measured in millivolts. This represents the analog-to-digital conversion gain coefficient, expressed in millivolts. Represents the bias constant, dimensionless. Based on calibration test data from clinical ultrasound equipment numbered 1 to 500. and The specific values ​​are used to ensure that the grayscale mapping conforms to human visual perception.

[0028] Further details on analog-to-digital conversion gain coefficients With bias constant The acquisition mechanism and specific setting basis. In ultrasound equipment, the amplitude of the analog radio frequency (RF) signal received by the probe usually has a large dynamic range. Direct linear conversion would result in extremely poor image contrast. Clinical calibration data numbered 1 to 500 are derived from RF echo measurement experiments using standard tissue phantoms at different gain depths. By fitting the logarithmic response characteristics of human vision using the Weber-Fechner law, the system uses the least squares method to perform linear approximate regression of the above 500 sets of analog voltage input values ​​with ideal 8-bit digital grayscale values ​​(0-255) to calculate the optimal conversion coefficient. Typical values ​​and basis: When the dynamic range of the input analog signal is truncated by the hardware front end... At that time, in order to map it to the full grayscale space, The typical value is set as follows ;and Used to filter out the noise floor caused by hardware thermal noise, its typical value is set to... (Dimensionless) Radio frequency signals below this level will be forced to zero, thereby ensuring that the background of the generated image is pure black, which conforms to the visual habits of the human eye and improves the boundary contrast of the target tissue.

[0029] After image conversion, the system reads the translation and rotation of the spatial positioning sensor fixed to the ultrasound probe in three-dimensional space to obtain its pose coordinates. The spatial positioning sensor, an electronic component attached to the ultrasound probe housing, is specifically designed to sense changes in spatial position and attitude in real time. Translation represents the linear distance the spatial positioning sensor moves relative to its initial reference point along three orthogonal spatial axes, while rotation represents the angular value of the sensor's rotation around these axes. The six-degree-of-freedom spatial position state, expressed by translation and rotation, is the pose coordinate. The formula for calculating the pose coordinate is: ; in the formula Representing time The pose coordinates are dimensionless. Representing time The translation amount is expressed in millimeters. Representing time The amount of rotation, in radians. This represents the translation normalization coefficient, expressed in millimeters. This represents the rotation normalization coefficient, expressed in radians. The settings are based on calibration experimental data from spatial positioning sensors numbered 1 to 300. and The numerical values ​​are used to eliminate numerical differences caused by different physical dimensions.

[0030] Further details on the translation normalization coefficients With rotation normalization coefficient The physical meaning and value logic of translation. In six-degree-of-freedom spatial pose tracking, translation... The dimension is millimeters (the order of magnitude is usually in millimeters). ), and rotation amount The dimension is radians (the order of magnitude is usually in 1000-1200 radians). If input directly into the neural network without processing, the order-of-magnitude translation will dominate gradient updates, causing the model to completely ignore rotational information. The calibration data, numbered 1 to 300, are derived from statistics on the physical limits of motion during standard scanning actions (such as sector scans and gliding) performed by experienced ultrasound physicians. Based on this data distribution, the system uses a maximum-minimum normalization strategy to set the coefficients. Typical values ​​and basis: According to statistics, the maximum translational distance during a single gliding motion in clinical scanning is approximately... The maximum deflection angle of the probe is approximately (i.e., 90 degrees). To map both uniformly to... dimensionless characteristic space, translation normalization coefficient The typical value is set to Rotational normalization coefficients The typical value is set to This approach mathematically eliminates the dimensional gap and ensures the balance of spatiotemporal feature fusion.

[0031] After acquiring the above data, the system timestamps and packages the continuous ultrasound image frames and pose coordinates to construct an ultrasound contrast-enhanced video stream and probe spatial pose data. Timestamp alignment synchronizes data from different sources along the timeline by matching identical timestamps. Data packaging encapsulates the synchronized multi-source data into a unified data stream according to a specific structure. The dynamic image sequence containing contrast agent microbubble echo information forms the ultrasound contrast-enhanced video stream, and the set of information recording the spatial state of the ultrasound probe during image acquisition forms the probe spatial pose data.

[0032] For the constructed video stream, the system calculates the pixel motion trajectory distribution of adjacent consecutive ultrasound image frames in the ultrasound contrast video stream, thereby extracting inter-frame pixel change features. Adjacent consecutive ultrasound image frames refer to two two-dimensional pixel matrices that are closely connected on the time axis and appear sequentially. The set of vectors representing the spatial displacement of each pixel position in the image over time is defined as the pixel motion trajectory distribution. The quantitative index reflecting the variation of pixel grayscale and position between adjacent images is the inter-frame pixel change feature. The formula for calculating the pixel motion trajectory distribution is: ; in the formula Representing time The distribution of pixel motion trajectories, in millimeters per second. It represents the change in the horizontal axis and is dimensionless. It represents the change in the vertical axis and is dimensionless. Represents the physical size of a pixel, in millimeters. Represents a time interval, in seconds. Based on dynamic evolution observation data from ultrasound contrast imaging, numbered 1 to 400. The value.

[0033] Further details on pixel physical size The acquisition method and setting basis. The spatial resolution mapping between image space and physical space is determined, and it is not a fixed constant, but highly dependent on the transmission frequency of the ultrasound probe and the depth setting parameters of the equipment. The observation data numbered 1 to 400 are derived from observations using precise spacing (e.g., every...). The calibration results of the metal target phantom in ultrasound contrast imaging mode. The system reads the depth parameters from the current ultrasound image header file and calculates the actual physical distance represented by a single pixel by combining the phantom calibration data. Typical values ​​and basis: For example, when clinically diagnosing superficial organs (such as the thyroid gland), the ultrasound detection depth is set to... And the vertical resolution of the generated image is At pixel level, Precisely set as .

[0034] The formula for calculating the inter-frame pixel change features is: ; in the formula Representing time The inter-frame pixel variation characteristics are dimensionless. The representative feature extraction coefficient is expressed in seconds per millimeter. The experimental setup is based on contrast agent microbubble flow tracking data numbered 1 to 200. The value.

[0035] Further details on feature extraction coefficients The physical mechanisms and defined boundaries. Its function is to represent the actual flow rate of microbubbles in the physical world ( Convert ) into feature values ​​suitable for deep learning models to process (usually in Within a certain range, to prevent numerical explosion leading to gradient vanishing. Experimental data numbered 1 to 200 are derived from Doppler and optical flow tracking statistics of actual blood flow velocities of contrast agents (such as sulfur hexafluoride microbubbles) in microvessels, venules, and arteries. Typical values ​​and basis: Because the perfusion velocity of capillary microcirculation angiography in solid tumors is usually within a certain range... In order to ensure that the extracted inter-frame pixel change features have the optimal activation response range in the neural network, a certain range is set. The principle is to maximize the typical flow rate ( Mapped to eigenvalues .therefore, The typical value is set to This parameter, as a feature scaling operator, not only preserves the sensitivity to microbubble velocity fluctuations but also caters to the sensitivity range of the nonlinear activation function in the subsequent operational evaluation model.

[0036] After feature extraction, the system serializes and concatenates the inter-frame pixel change features according to timestamp order, thus generating the initial spatiotemporal feature sequence for contrast imaging. The timestamp order follows the sequential arrangement rule of the time stamps assigned by the system when the data was generated. Serialization and concatenation combine discrete features into a continuous multidimensional array according to the chronological order. The preliminary feature set, which integrates temporal changes and spatial pixel motion, constitutes the initial spatiotemporal feature sequence for contrast imaging. The formula for calculating the initial spatiotemporal feature sequence is: ; in the formula It represents the initial spatiotemporal characteristic sequence of the imaging, and is dimensionless. An index representing the order of timestamps, dimensionless. Represents the total number of frames, dimensionless. Representative index The corresponding inter-frame pixel variation features are dimensionless. Representative index The corresponding mutually perpendicular unit direction vectors are dimensionless. The experimental data is based on the multidimensional decomposition of the feature matrices numbered 1 to 100. The values ​​are set to ensure information independence during the feature splicing process.

[0037] Further details on mutually perpendicular unit direction vectors The mathematical construction of the model and its technical necessity in temporal modeling are discussed. When serializing and concatenating discrete inter-frame variation features into an initial spatiotemporal feature sequence, simply performing scalar addition can lead to aliasing and overwriting of historical information in the temporal dimension, making it impossible for the model to distinguish the order of events. Experimental data numbered 1 to 100 were obtained by using principal component analysis and Gram-Schmidt orthogonalization to reduce the dimensionality and decouple the ultrasound features across multiple consecutive frames. In reality, it is a set of orthogonal basis vectors in a high-dimensional Hilbert space. Typical values ​​and their rationale: To ensure the... The feature information of the frame and the first The feature information of the frames is absolutely orthogonal in mathematical space (i.e., the inner product is orthogonal). ,when When performing short-time analysis (e.g., extracting 3 frames consecutively), One-hot encoded unit vectors are set to be mutually orthogonal, such as , , By introducing this orthogonal matrix vector for product concatenation, linear correlation and information redundancy in the temporal feature concatenation process are fundamentally eliminated, providing a pure and dimension-independent mathematical tensor foundation for the subsequent extraction of accurate three-dimensional spatiotemporal evolution trajectories.

[0038] For example, the system receives the analog signal output by the ultrasound probe in contrast-enhanced mode and performs analog-to-digital conversion to obtain continuous ultrasound image frames. The analog signal is set to 80 millivolts, the analog-to-digital conversion gain coefficient is set to 2.5 per millivolt, and the bias constant is set to 10. Substituting these values ​​into the formula, the grayscale value of the continuous ultrasound image frame is calculated to be 210. After image conversion, the translation and rotation of the spatial positioning sensor fixed on the ultrasound probe in three-dimensional space are read to obtain the pose coordinates. The translation is set to 50 millimeters, the rotation to 0.5 radians, the translation normalization coefficient to 0.02 per millimeter, and the rotation normalization coefficient to 0.5 per radian. Substituting these values ​​into the formula, the pose coordinate value is calculated to be 1.25. After obtaining the above data, the continuous ultrasound image frames and pose coordinates are timestamped and packaged to construct an ultrasound contrast-enhanced video stream and probe spatial pose data. For the constructed video stream, the pixel motion trajectory distribution of adjacent continuous ultrasound image frames in the ultrasound contrast-enhanced video stream is calculated, and inter-frame pixel change features are extracted. The horizontal coordinate change was set to 3, the vertical coordinate change to 4, the pixel physical size to 0.2 mm, and the time interval to 0.1 seconds. Substituting these values ​​into the formula, the pixel motion trajectory distribution was calculated to be 10 mm per second. The feature extraction coefficient was set to 0.5 seconds per millimeter. Substituting this value into the formula, the inter-frame pixel change feature was calculated to be 5. After feature extraction, the inter-frame pixel change features were serialized and concatenated according to timestamp order to generate the initial angiography spatiotemporal feature sequence. The total number of frames was set to 3. The inter-frame pixel change feature of the first frame was set to 5, the inter-frame pixel change feature of the second frame to 6, and the inter-frame pixel change feature of the last frame to 7. The corresponding mutually perpendicular unit direction vectors were set as independent direction vectors in the three-dimensional spatial coordinate system. Substituting these values ​​into the formula, the initial angiography spatiotemporal feature sequence was calculated as a combination of spatial vectors 5, 6, and 7. The above data calculations verified the rationality of the technical feature terminology and the accuracy of the formula derivation.

[0039] S2. Extract the anatomical structure contour from the initial spatiotemporal feature sequence of the contrast imaging to obtain the dynamic boundary features of the target tissue, and perform spatial mapping in combination with the probe spatial pose data to generate the three-dimensional spatiotemporal evolution trajectory of the target tissue.

[0040] In a specific embodiment of the present invention, the anatomical structure contour is extracted from the initial contrast spatiotemporal feature sequence to obtain the dynamic boundary features of the target tissue, and spatial mapping is performed in combination with the probe spatial pose data to generate the three-dimensional spatiotemporal evolution trajectory of the target tissue, including: edge detection and morphological processing of the initial contrast spatiotemporal feature sequence to extract the anatomical structure contour.

[0041] By tracing the deformation process of the anatomical structure outline over time, the dynamic boundary characteristics of the target tissue can be obtained.

[0042] In a specific embodiment of the present invention, the deformation process of the anatomical structure contour in the time dimension is tracked to obtain the dynamic boundary features of the target tissue, including: calculating the displacement vector of each feature point in the anatomical structure contour at adjacent time nodes to obtain the contour motion field.

[0043] The contour motion field is smoothed and outliers are removed to extract the effective deformation components.

[0044] The effective deformation components are integrated along the time axis to obtain the dynamic boundary characteristics of the target tissue.

[0045] By using probe spatial pose data to construct a world coordinate system, the dynamic boundary features of the target tissue are transformed into the world coordinate system for spatial mapping, generating the three-dimensional spatiotemporal evolution trajectory of the target tissue.

[0046] Specifically, the system acquires the initial spatiotemporal feature sequence generated in the preceding steps, performs edge detection and morphological processing on the initial spatiotemporal feature sequence, and extracts the anatomical structure contour. Edge detection and morphological processing refer to the process of finding boundary lines by calculating pixel gray-level gradients and using structuring elements to perform dilation and erosion operations on the boundary lines to eliminate noise and connect broken parts. The closed or semi-closed curve shape reflecting the geometry of internal organs or lesions obtained after the above processing is the anatomical structure contour. The formula for calculating the anatomical structure contour is: ; in the formula Representing time The anatomical outline is dimensionless. Represents the spatiotemporal feature sequence of the initial imaging in time The characteristic matrix of is dimensionless. Represents the horizontal pixel difference, dimensionless. Represents the vertical pixel difference, dimensionless. Represents the edge enhancement coefficient, dimensionless. Based on experimental data of medical image segmentation numbered 1 to 300. The value is used to ensure the integrity of boundary extraction.

[0047] Further details on edge enhancement coefficient The value selection logic and physical meaning. Ultrasound images inherently contain severe speckle noise, and traditional gradient difference algorithms are prone to causing boundary breaks in target tissues (such as tumor capsules). The medical image segmentation experimental data numbered 1 to 300 were obtained by comparing the real contours manually drawn by experienced ultrasound physicians with the contours extracted by the machine. Based on this data, adaptive histogram equalization and gradient compensation techniques were used for calibration. Typical values ​​and their basis: The typical value is set as follows Between. If Because ultrasound signals attenuate in deep tissues, they can cause the loss of subtle deep boundaries; if This will amplify random acoustic noise errors into anatomical boundaries. Set as It can effectively suppress noise while ensuring the closure and integrity of the extracted anatomical structure contour in terms of topology.

[0048] After acquiring the contour, the system tracks the deformation process of the anatomical structure contour over time, thereby obtaining the dynamic boundary features of the target tissue. The deformation process represents the changes in shape and position of the anatomical structure contour over time. A set of quantified data reflecting the motion and geometric changes of the outer edge of a specific observed object over a continuous time period constitutes the dynamic boundary features of the target tissue. To achieve this, the system calculates the displacement vectors of each feature point in the anatomical structure contour at adjacent time points, obtaining the contour motion field. Feature points refer to discrete pixel coordinates with significant geometric features on the anatomical structure contour. Adjacent time points represent closely connected sampling moments on the time axis. The direction and distance of the positional change of feature points between adjacent time points are defined as displacement vectors. The spatial distribution set formed by the displacement vectors of all feature points on the anatomical structure contour is the contour motion field. The formula for calculating the contour motion field is: ; in the formula Represents the outline of the motion field, measured in millimeters per second. Representing time The coordinates of the feature points are dimensionless. Representing time The coordinates of the feature points are dimensionless. Represents the physical size of a pixel, in millimeters. This represents a time interval, measured in seconds.

[0049] Next, the system performs smoothing filtering and outlier removal on the contour motion field to extract the effective deformation components. Smoothing filtering and outlier removal involve weighted averaging of adjacent data to eliminate high-frequency noise and deleting erroneous data that deviates from the normal range based on statistical thresholds. The vector data that truly reflects the contour shape change after noise filtering and erroneous data cleanup is the effective deformation component. The formula for calculating the effective deformation component is: ; in the formula Represents the effective deformation component, measured in millimeters per second. The index representing a neighboring point is dimensionless. Represents the number of neighboring points, dimensionless. Represents the filter weight coefficient, which is dimensionless. The contour motion field within the neighborhood is represented in millimeters per second. This setting is based on motion artifact elimination test data numbered 1 to 200. The value is set to achieve the best noise reduction effect.

[0050] Further details on the filter weight coefficients Distribution patterns and noise reduction mechanisms. During probe scanning, involuntary minor tremors from patient breathing or the doctor's hand can generate high-frequency motion artifacts, resulting in burrs in the extracted motion field. Spectral analysis was performed on test data numbered 1 to 200 to address various physiological artifacts. Typical values ​​and their basis: This typically involves a discrete weight operator that follows a Gaussian distribution (such as a one-dimensional Gaussian kernel). For example, when the number of neighborhood points... (That is, when examining a total of 3 frames) Typical configuration of the combination is This weight combination assigns the highest confidence level (0.50) to the current center frame, while incorporating the trends from previous and subsequent times (0.25 each), thus mathematically acting as a low-pass filter. This effectively filters out millisecond-level non-real probe abrupt displacements and extracts the smooth and effective deformation components that represent the probe's true movement intention.

[0051] Subsequently, the system integrates the effective deformation components along the time axis to obtain the dynamic boundary characteristics of the target tissue. Integration represents the mathematical calculation of summing the effective deformation components over various time periods to obtain the total cumulative change. The formula for calculating the dynamic boundary characteristics of the target tissue is: ; in the formula Represents the dynamic boundary characteristics of the target organization, in millimeters. An index representing a time point, dimensionless. Represents the total number of time points, dimensionless. Representing time The effective deformation component is expressed in millimeters per second. This represents a time interval, measured in seconds.

[0052] Finally, the system constructs a world coordinate system using probe spatial pose data, transforms the dynamic boundary features of the target tissue into the world coordinate system for spatial mapping, and generates the three-dimensional spatiotemporal evolution trajectory of the target tissue. The world coordinate system is an absolute three-dimensional coordinate reference system established with a fixed spatial reference point as its origin. The mathematical operation of transferring data points from the local coordinate system to the absolute three-dimensional coordinate reference system through translation and rotation is called spatial mapping. The three-dimensional spatiotemporal evolution trajectory of the target tissue is constituted by the continuously changing three-dimensional geometric model of the outer edge of a specific observed object recorded in absolute three-dimensional space. The formula for calculating the three-dimensional spatiotemporal evolution trajectory of the target tissue is: ; in the formula The three-dimensional spatiotemporal evolution trajectory of the target organization, in millimeters. Represents a dimensionless rotation matrix extracted from probe spatial pose data. This represents the translation vector extracted from the probe's spatial pose data, in millimeters. The settings are based on calibration experimental data from spatial positioning sensors numbered 1 to 500. and The values ​​are set to ensure the accuracy of the spatial mapping.

[0053] Further details on spatial mapping parameters (Rotation matrix) and The method for obtaining the (translation vector). These two parameters represent the 'hand-eye calibration' extrinsic parameter matrix for transforming the ultrasound image from a two-dimensional local coordinate system to a three-dimensional absolute coordinate system. Calibration experiments numbered 1 to 500 were conducted using a standard N-Wire spatial calibration phantom, combined with a high-precision optical / electromagnetic tracker. Typical characteristics and basis: It is orthogonal rotation matrix, It is The translation vector (in millimeters). During the initialization phase of system deployment, singular value decomposition (SVD) is used to derive the physical bias constant matrix embedded in the probe hardware structure. Their introduction rigorously ensures the high-precision clinical spatial anatomical accuracy of the calculated three-dimensional spatiotemporal evolution trajectory, eliminating systematic errors caused by the probe's own geometry.

[0054] For example, the system performs edge detection and morphological processing on the initial spatiotemporal feature sequence of the imaging to extract the anatomical structure contour. A horizontal pixel difference value of 3, a vertical pixel difference value of 4, and an edge enhancement coefficient of 1.2 are set, and the calculated value of the anatomical structure contour is 6. The system tracks the deformation process of the anatomical structure contour over time to obtain the dynamic boundary features of the target tissue. Specifically, the displacement vectors of each feature point in the anatomical structure contour at adjacent time nodes are calculated to obtain the contour motion field. The time frame is set... The feature point coordinates are 15, and the time is set. The feature point coordinates are set to 10, the pixel physical size is set to 0.2 mm, and the time interval is set to 0.1 seconds. Substituting these values ​​into the formula, the contour motion field is calculated to be 10 mm / s. Smoothing filtering and outlier removal are applied to the contour motion field to extract the effective deformation component. The number of neighborhood points is set to 3, the contour motion field in the first neighborhood is set to 9 mm / s, the next neighborhood to 10 mm / s, and the last neighborhood to 11 mm / s, with corresponding filtering weight coefficients of 0.3, 0.4, and 0.3 respectively. Substituting these values ​​into the formula, the effective deformation component is calculated to be 10 mm / s. The effective deformation component is integrated along the time axis to obtain the dynamic boundary features of the target tissue. The total number of time nodes is set to 5, the effective deformation component at each time node is set to 10 mm / s, and the time interval is set to 0.1 seconds. Substituting these values ​​into the formula, the dynamic boundary features of the target tissue are calculated to be 5 mm. A world coordinate system is constructed using the probe's spatial pose data. The dynamic boundary features of the target tissue are then transformed into the world coordinate system for spatial mapping, generating the three-dimensional spatiotemporal evolution trajectory of the target tissue. The dynamic boundary feature of the target tissue was set to 5 mm. The rotation matrix extracted from the probe's spatial pose data was defined as a matrix with all diagonal elements equal to 1. The translation vector extracted from the probe's spatial pose data was set to 2 mm. Substituting these values ​​into the formula, the calculated three-dimensional spatiotemporal evolution trajectory of the target tissue was found to be 7 mm. The rationality of the technical feature terminology and the accuracy of the formula derivation were verified through the above closed-loop data calculation.

[0055] S3. Analyze the changes in contrast agent perfusion intensity in the three-dimensional spatiotemporal evolution trajectory of the target tissue, extract the perfusion time series characterization parameters, and perform feature fusion with the initial contrast spatiotemporal feature sequence to generate the contrast time series characterization vector.

[0056] In a specific embodiment of the present invention, the change in contrast agent perfusion intensity in the three-dimensional spatiotemporal evolution trajectory of the target tissue is analyzed, perfusion time sequence characterization parameters are extracted, and feature fusion is performed with the initial contrast spatiotemporal feature sequence to generate a contrast time sequence characterization vector, including: calculating the rate of change of gray value of each voxel in the three-dimensional spatiotemporal evolution trajectory of the target tissue over time, and determining the change in contrast agent perfusion intensity.

[0057] Curve fitting was performed on the changes in contrast agent perfusion intensity, and the time to peak and peak intensity were extracted as parameters characterizing the perfusion time series.

[0058] In a specific embodiment of the present invention, curve fitting is performed on the change in contrast agent perfusion intensity, and the peak time and peak intensity are extracted as perfusion time series characterization parameters, including: mapping the change in contrast agent perfusion intensity to a time intensity coordinate system to generate discrete perfusion data points.

[0059] A mathematical model describing the change of contrast agent concentration in tissue over time was used to perform nonlinear least squares fitting on discrete perfusion data points to obtain continuous perfusion curves.

[0060] The highest point of the continuous perfusion curve is located, and the time and amplitude values ​​corresponding to the highest point are extracted as the peak time and peak intensity, which constitute the perfusion time series characterization parameters.

[0061] The perfusion time-series characterization parameters and the initial angiography spatiotemporal feature sequence are concatenated by channels and attention weights are assigned to achieve feature fusion, generating an angiography time-series characterization vector.

[0062] Specifically, the system acquires the three-dimensional spatiotemporal evolution trajectory of the target tissue generated in the preceding steps, calculates the rate of change of grayscale values ​​of each voxel within the three-dimensional spatiotemporal evolution trajectory of the target tissue over time, and thereby determines the change in contrast agent perfusion intensity. A voxel represents the smallest data unit in three-dimensional space. The calculation result of dividing the grayscale difference of voxels at adjacent time points by the time interval is the rate of change of grayscale values ​​over time. The quantitative index reflecting the dynamic increase or decrease of contrast agent concentration in the tissue over time constitutes the change in contrast agent perfusion intensity. The formula for calculating the change in contrast agent perfusion intensity is: ; in the formula This represents the change in contrast agent perfusion intensity, expressed in gray levels per second. Representing time The voxel grayscale value, in units of grayscale levels. Representing time The voxel grayscale value, in units of grayscale levels. This represents a time interval, measured in seconds.

[0063] After acquiring the aforementioned change data, the system maps the contrast agent perfusion intensity changes to a time-intensity coordinate system, thereby generating discrete perfusion data points. The time-intensity coordinate system is a two-dimensional reference system with time as the horizontal axis and intensity as the vertical axis. In the time-intensity coordinate system, isolated coordinate points formed by specific times and corresponding intensity values ​​are the discrete perfusion data points. Subsequently, the system uses a mathematical model describing the change in contrast agent concentration in the tissue over time to perform nonlinear least-squares fitting on the discrete perfusion data points, obtaining a continuous perfusion curve. The mathematical model describing the change in contrast agent concentration in the tissue over time refers to a mathematical function describing the inflow and outflow patterns of microbubbles. Nonlinear least-squares fitting represents a mathematical optimization method that minimizes the sum of squared errors from all data points to the function curve by continuously adjusting the function parameters. The smooth and continuous mathematical curve reflecting the concentration change trend obtained after curve fitting is the continuous perfusion curve. The formula for calculating the continuous perfusion curve is: ; in the formula Represents the continuous perfusion curve over time The amplitude value, in gray levels. This represents a time value, in seconds. This represents the scaling factor, measured in negative grayscale levels per second. Power of 1. It represents the inflow rate index and is dimensionless. This represents the outflow attenuation coefficient, measured in seconds. Represents a natural constant, dimensionless. Based on clinical contrast agent metabolism observation data numbered 1 to 600. , , The initial iteration values ​​are used to ensure that the fit converges.

[0064] Further details on the parameters in the angiography perfusion mathematical model (Magnification factor) (Inflow rate index) Initial iteration setting rules for (outflow attenuation coefficient). Since nonlinear least squares methods are highly dependent on the initial 'seed value,' a poor initial value can cause the algorithm to get trapped in local minima or fail to converge. Clinical imaging observation data numbered 1 to 600 cover the metabolic cycles of typical organs such as the liver and thyroid using sulfur hexafluoride microbubbles. Typical values ​​and basis: Based on the laws of human macrocirculation hemodynamics, the system's preset initial iteration values ​​are usually anchored as follows: , , This initial parameter combination depicts a standard physiological perfusion baseline curve that peaks at approximately 10-15 seconds and then decays exponentially over the following minute. Using this as a baseline for optimization iterations ensures that the model can achieve rapid and stable convergence within 50 iterations when faced with discrete perfusion data points exhibiting any individual variability.

[0065] After curve fitting, the system locates the highest point of the continuous perfusion curve and extracts the corresponding time and amplitude values ​​as the peak time and peak intensity, thus forming the perfusion time-series characterization parameters. The highest point represents the position with the largest amplitude value on the continuous perfusion curve. The horizontal axis value corresponding to the highest point is the time value, and the vertical axis value is the amplitude value. The time required for the contrast agent concentration to reach its maximum is defined as the peak time, and the maximum value of the contrast agent concentration is defined as the peak intensity. The data set describing the perfusion characteristics, composed of the peak time and peak intensity, constitutes the perfusion time-series characterization parameters. The formula for calculating the perfusion time-series characterization parameters is: ; in the formula The normalized perfusion time series characterization parameter is dimensionless. Represents the time to reach the peak, in seconds. This represents the maximum observation time, expressed in seconds. Represents peak intensity, expressed in gray levels. This represents the system's maximum grayscale level, expressed in grayscale levels. It is based on the hardware parameter settings of ultrasound devices numbered 1 to 400. and The numerical values ​​are used to eliminate differences in physical dimensions.

[0066] Further details on normalization constants and The setting basis. Infusion time is usually within... In a matter of seconds, the pixel grayscale is... If the absolute numerical difference between these two physical quantities (time and brightness) is not eliminated, it will cause dimensionality bias in subsequent deep learning models when extracting features. Typical values ​​and basis: The system obtains the longest recording time set for a single ultrasound angiography machine (usually fixed at 1) by reading the header file of the underlying hardware configuration parameters of the ultrasound equipment numbered 1 to 400. (seconds), and the bit depth of the device image export (for a standard 8-bit ultrasound image, For 12-bit Raw data, By dividing by these two maximum physical extrema respectively, the peak time and peak intensity are strictly mapped to... Within the closed interval, the numerical interference caused by physical dimensions on the feature space measurement of the neural network is eliminated from the bottom layer.

[0067] Finally, the system performs channel concatenation and attention weight allocation on the perfusion time-series representation parameters and the initial spatiotemporal feature sequence to achieve feature fusion, ultimately generating the angiography time-series representation vector. The initial spatiotemporal feature sequence is a continuous multidimensional array generated in the preceding steps. Channel concatenation refers to the operation of directly combining feature data from different sources along a specific data dimension. The mathematical process of assigning corresponding product coefficients to different feature components according to their importance is called attention weight allocation. The processing step of integrating multi-source features into a unified expression is called feature fusion. The comprehensive data sequence that integrates spatial deformation and temporal perfusion characteristics constitutes the angiography time-series representation vector. The formula for calculating the angiography time-series representation vector is: ; in the formula The vector representing the imaging time sequence is dimensionless. The attention weight represents the infusion parameter and is dimensionless. The normalized perfusion time series characterization parameter is dimensionless. Attention weights representing spatiotemporal features are dimensionless. It represents the initial spatiotemporal characteristic sequence of the imaging, and is dimensionless. Represents the channel concatenation operator. Experimental data settings based on feature importance assessment from numbers 1 to 500. and The numerical values ​​are used to highlight key diagnostic information.

[0068] Further details on attention weights (Infusion parameter weights) and The allocation logic of (spatiotemporal feature weights). In contrast-enhanced ultrasound assessment, if the probe shifts, the most direct and fatal consequence is that the target lesion moves out of the cut, causing a sudden interruption or distortion of the contrast-enhanced time-intensity curve. The feature importance assessment experiments numbered 1 to 500, based on random forest and ablation experiments, confirmed that perfusion features contribute more to the discrimination of probe stability. Typical values ​​and basis: Therefore, The typical value is set as follows , Set as This asymmetric weighting mechanism, which emphasizes perfusion over displacement, forces the operational evaluation model to prioritize the core blood perfusion information for clinical diagnosis, allowing the system to grant reasonable tolerance to tolerable operations that involve slight probe slippage but do not lose the core perfusion area of ​​the lesion.

[0069] For example, the system calculates the rate of change of grayscale values ​​of each voxel within the three-dimensional spatiotemporal evolution trajectory of the target tissue over time to determine the change in contrast agent perfusion intensity. (Set time) The voxel grayscale value is 120 gray levels, and the time is set. The voxel grayscale value is 100 gray levels, and the time interval is set to 2 seconds. Substituting these values ​​into the formula, the change in contrast agent perfusion intensity is calculated to be 10 gray levels per second. The change in contrast agent perfusion intensity is mapped onto a time-intensity coordinate system to generate discrete perfusion data points. A mathematical model describing the change in contrast agent concentration in tissue over time is used to perform nonlinear least-squares fitting on the discrete perfusion data points to obtain a continuous perfusion curve. The time value is set to 2 seconds, the amplitude scaling factor is set to 50 gray levels to the power of -2 per second, the inflow rate exponent is set to 2, the outflow attenuation factor is set to 0.5 per second, and the natural constant is set to 2.718. Substituting these values ​​into the formula, the amplitude value of the continuous perfusion curve at time 2 seconds is calculated to be 73.58 gray levels. The highest point of the continuous perfusion curve is located, and the time and amplitude values ​​corresponding to the highest point are extracted as the peak time and peak intensity, constituting the perfusion time-series characterization parameters. The peak time was set to 4 seconds, the maximum observation time to 60 seconds, the peak intensity to 150 gray levels, and the maximum gray level of the system to 255 gray levels. Substituting these values ​​into the formula, the normalized perfusion time-series characterization parameters were calculated as dimensionless vectors containing 0.067 and 0.588. The perfusion time-series characterization parameters and the initial spatiotemporal feature sequence of contrast imaging were then combined using channel concatenation and attention weight allocation to achieve feature fusion, generating a contrast imaging time-series characterization vector. The normalized perfusion time-series characterization parameters were set to the aforementioned vector, the initial spatiotemporal feature sequence of contrast imaging was set to a dimensionless vector containing 0.4 and 0.5, the attention weight for the perfusion parameters was set to 0.6, and the attention weight for the spatiotemporal features was set to 0.4. Substituting these values ​​into the formula, the contrast imaging time-series characterization vector was calculated as a combined vector containing 0.0402, 0.3528, 0.16, and 0.2. The above closed-loop data calculations verified the rationality of the technical feature terminology and the accuracy of the formula derivation.

[0070] S4. Input the contrast time sequence representation vector into the operation evaluation model to extract quality features, output the ultrasound operation stability score and section deviation correction suggestions, and generate an intelligent ultrasound operation evaluation report.

[0071] In a specific embodiment of the present invention, the contrast time sequence representation vector is input into the operation evaluation model for quality feature extraction, and the ultrasound operation stability score and section offset correction suggestions are output to generate an ultrasound operation intelligent evaluation report. This includes: inputting the contrast time sequence representation vector into the convolutional layer and recurrent neural network layer of the operation evaluation model for quality feature extraction to obtain an operation behavior feature map.

[0072] The operation behavior feature map is classified and regressed to output an ultrasound operation stability score and a section deviation correction suggestion.

[0073] In a specific embodiment of the present invention, the operation behavior feature map is classified and regressed to output an ultrasound operation stability score and a section offset correction suggestion. This includes: inputting the operation behavior feature map into a fully connected layer for dimensionality reduction to obtain a dimensionality reduction vector of the operation behavior.

[0074] A regressor is used to numerically map the dimension-reduced vector of the operational behavior to calculate the ultrasonic operation stability score.

[0075] The classifier is used to perform pattern matching on the dimensionality-reduced vector of the operation behavior to determine the offset direction and angle, and output the cross-section offset correction suggestion.

[0076] The ultrasound operation stability score and section deviation correction suggestions are summarized, formatted, and generated into an intelligent ultrasound operation assessment report.

[0077] Specifically, the system acquires the contrast-enhanced temporal representation vector generated in the preceding steps and inputs it into the operation evaluation model for quality feature extraction. The operation evaluation model is a mathematical computational framework used to analyze the standardization of the physician's scanning technique. Quality feature extraction refers to the process of mining deep data patterns reflecting probe holding stability and scanning continuity from the input data. To achieve this, the system inputs the contrast-enhanced temporal representation vector into the convolutional and recurrent neural network layers of the operation evaluation model. Here, the convolutional and recurrent neural network layers represent mathematical logic layers that process data through multiple local feature extractors and time-series memory units. The multidimensional data matrix containing spatiotemporal correlation information obtained after the above processing is the operation behavior feature map. The formula for calculating the operation behavior feature map is: ; in the formula The representation of operational behavior characteristics is dimensionless. The vector representing the imaging time sequence is dimensionless. This represents the hidden layer weight matrix, which is dimensionless. This represents the hidden layer bias vector, which is dimensionless. This represents a non-linear activation function, dimensionless. It is based on a clinical ultrasound operation standard dataset numbered 1 to 800. and The numerical value is used to ensure the accuracy of feature extraction.

[0078] Further details on nonlinear activation functions and hidden layer parameters The nature of the project. Typically, the ReLU activation function (corrected linear unit) is used. Its function is to endow the model with nonlinear mapping capabilities, enabling it to solve complex classification and regression problems. (Hidden layer weight matrix) and The hidden layer bias vector is not a manually specified scalar, but a high-dimensional tensor. The clinical ultrasound operation standard dataset, numbered 1 to 800, contains data on 'gold standard' movements demonstrated by multiple senior experts and 'incorrect' movements by novice trainees. Through tens of thousands of backpropagation iterations trained on this data, the model automatically optimized and solidified. and The internal numerical distribution of these features essentially constitutes a set of high-dimensional 'filters' capable of extremely sensitively capturing specific high-frequency noise patterns in the imaging time-series representation vector that characterize 'hand tremor' or 'scanning inconsistency'.

[0079] After acquiring the feature maps, the system performs classification and regression calculations on the operational behavior feature maps, thereby outputting an ultrasound operation stability score and a section offset correction suggestion. Specifically, the system inputs the operational behavior feature maps into a fully connected layer for dimensionality reduction, thus obtaining a dimensionality-reduced operational behavior vector. A fully connected layer is a mathematical processing unit that maps a multidimensional data matrix into a low-dimensional feature representation through a global weight matrix. Dimensionality reduction represents a mathematical operation that reduces data dimensionality while retaining core feature information. The numerical sequence obtained after processing by the global weight matrix constitutes the dimensionality-reduced operational behavior vector. The formula for calculating the dimensionality-reduced operational behavior vector is: ; in the formula The dimensionless vector represents the reduced-dimensionality of the operational behavior. This represents the dimension-reduced weight matrix, which is dimensionless. This represents the dimension reduction bias vector, which is dimensionless. The data is based on the feature compression experimental data numbered 1 to 500. and The value.

[0080] Further explanation of the dimensionality reduction weight matrix in the above dimensionality reduction formula With dimension reduction bias vector The setting mechanism and its technical role in feature extraction. Dimensionality reduction weight matrix. With bias vector The mathematical properties and physical meanings of the output operation behavior feature map are processed by the preceding convolutional layer and recurrent neural network layer. Typically, this involves a dataset with extremely high dimensionality (e.g., containing 256 or 512-dimensional deep features). Directly using such high-dimensional features for subsequent regression scoring and classification matching can easily lead to an explosion of model parameters and cause severe overfitting—meaning the model only remembers specific noise from the training data and loses its ability to generalize to new samples. Therefore, fully connected layers must be introduced as "information funnels" for dimensionality reduction. It is a broad and highly asymmetric projection matrix (e.g., dimension 1). ), It is the corresponding bias (dimension is) In a physical sense, Each row is equivalent to a "principal component projection basis" in the feature space. Its function is to force the projection of the matrix into a specific feature space through matrix multiplication. The 256 complex feature indicators were linearly fused and condensed into 32 key core dimensions. These 32 dimensions constitute the dimensionality reduction vector of the operational behavior. They eliminated redundant background information unrelated to probe stability, retaining the purest "operational action quality fingerprint".

[0081] and The specific setting mechanism and basis of the numerical values ​​are in the formula. and The values ​​are not specified by humans beforehand, but rather dynamically calculated and solidified during the system pre-training phase using an optimization algorithm based on the "feature compression experimental data numbered 1 to 500". Experimental data background: These 500 sets of experimental data contain 500 standard and non-standard movement trajectories executed by the probe under controlled conditions, and record the high-dimensional initial features generated by these trajectories. Numerical computation (setting) process: In the model building phase, the system typically employs an autoencoder architecture or a backpropagation mechanism combined with overall classification loss to perform unsupervised or supervised learning on these 500 sets of data. The goal of the training process is to force the data through this... Even after constructing a low-dimensional "bottleneck," it is still possible to reconstruct the original important information to the maximum extent or maintain classification accuracy. After thousands of gradient descent iterations, when the reconstruction error or classification error caused by compression reaches its global minimum, the values ​​remaining in the matrix network at this point are the final values ​​that are "set" and solidified. and In conclusion, and The values ​​are high-dimensional feature transformation constants derived through optimization driven by a large amount of real physical operation data. Their precise setting not only significantly reduces the computational complexity of the subsequent system and ensures real-time evaluation efficiency, but also perfectly refines the key operational features required for evaluation.

[0082] Next, the system uses a regressor to numerically map the dimensionality-reduced vector of the operational behavior, thereby calculating the ultrasound operational stability score. The regressor represents a mathematical calculator that establishes a mapping relationship between input features and continuous numerical targets. Numerical mapping refers to the process of converting a multidimensional feature vector into continuous numerical values. The numerical index that quantifies the smoothness of holding and moving the ultrasound probe is the ultrasound operational stability score. The formula for calculating the ultrasound operational stability score is: ; in the formula This represents the stability score of ultrasound operation and is dimensionless. This represents the regression mapping coefficient vector, which is dimensionless. This represents the fundamental regression constant and is dimensionless. It is based on expert scoring records numbered 1 to 600. and The value.

[0083] Further details on the regression mapping coefficient vector With regression fundamental constant The computational logic is as follows. The ultrasound operation stability score needs to output a percentage score (0-100) that aligns with human cognitive intuition. Expert scoring records numbered 1 to 600 provide supervisory labels for the regression analysis. Typical values ​​and their basis: This represents the basic passing score for the system; the typical value is usually fixed at 1. or ;and It is a multidimensional penalty / reward coefficient vector learned during training. When the dimensionality reduction vector... When undesirable features characterizing 'severe slippage' are activated, It will output a negative penalty value (e.g., This resulted in the final score dropping to failing. (Points are awarded for good performance); conversely, bonus points are awarded for poor performance, thus perfectly mapping black-box features into interpretable clinical evaluation scores.

[0084] Simultaneously, the system utilizes a classifier to perform pattern matching on the dimensionality-reduced vector of the operational behavior, determining the offset direction and angle, and then outputting cross-sectional offset correction suggestions. A classifier is a mathematical matcher that divides input features into preset discrete categories. Pattern matching represents the process of calculating the similarity between input features and a preset standard template and determining the category to which they belong. The spatial orientation and rotation degrees of the actual ultrasound probe scanning cross-section deviating from the standard anatomical cross-section constitute the offset direction and angle. The textual instructions guiding the operator to adjust the probe position and orientation to restore the standard cross-section are the cross-sectional offset correction suggestions. The formula for calculating the matching probability of a specific offset category is: ; in the formula Represents the matching probability of a specific offset category; it is dimensionless. Representing the The classification weight vector for each category, dimensionless. Represents the total number of preset offset categories; dimensionless. Represents a natural constant, dimensionless. Based on experimental data from probe offset correction experiments numbered 1 to 400. The system selects the text instruction corresponding to the category with the highest matching probability as the final aspect offset correction suggestion.

[0085] Further details on the first Classification weight vectors for each category The invention includes technical functions. It pre-defines several typical probe offset error categories (e.g., excessive probe tilt, leftward slippage, excessive rotation angle, etc.), totaling... Class. The characteristic references for these typical errors were collected in advance for probe offset experiments numbered 1 to 400. Physical meaning and basis: In essence, it represents the first The 'standard feature template vector' for each error type. The inner product operation in the formula is used when calculating the matching probability (Softmax activation process). In essence, it calculates the 'cosine similarity' between the current test subject's operational feature vector and a preset error template. The higher the similarity, the closer the matching probability of that category is to... Based on this, the system accurately pinpoints the operator's specific erroneous action (e.g., determining it as 'the probe tilted to the left, missing the lesion on the right side'), and then matches and outputs highly instructive, definitive corrective instructions.

[0086] Finally, the system summarizes the ultrasound operation stability score and section deviation correction suggestions, and formats them for layout, ultimately generating an intelligent ultrasound operation assessment report. Formatted layout refers to the process of combining and presenting the assessment results according to preset page layout rules. The final output document, including the stability score and specific correction guidance, constitutes the intelligent ultrasound operation assessment report.

[0087] For example, the system inputs the contrast-enhanced temporal representation vector into the convolutional layer and recurrent neural network layer of the operation evaluation model to extract quality features, obtaining an operation behavior feature map. The contrast-enhanced temporal representation vector is set to a dimensionless vector consisting of 0.5 and 0.8, the hidden layer weight matrix is ​​set to a two-dimensional matrix consisting of 0.2, 0.3, 0.4, and 0.5, the hidden layer bias vector is set to a dimensionless vector consisting of 0.1 and 0.2, and the nonlinear activation function is set to a linear rectified function. Substituting these values ​​into the formula, the operation behavior feature map is calculated as a dimensionless vector consisting of 0.44 and 0.8. The operation behavior feature map is then input into a fully connected layer for dimensionality reduction, obtaining a dimensionality-reduced operation behavior vector. The dimensionality-reduced weight matrix is ​​set to a row vector consisting of 0.5 and 0.6, the dimensionality-reduced bias vector is set to 0.1, and the value of the dimensionality-reduced operation behavior vector is calculated as 0.8. A regressor is used to numerically map the dimensionality-reduced operation behavior vector to calculate the ultrasound operation stability score. The regression mapping coefficient vector was set to 50, and the regression baseline constant was set to 40. Substituting these values ​​into the formula, the ultrasound operation stability score was calculated to be 80. A classifier was used to perform pattern matching on the dimensionality-reduced vector of the operation behavior to determine the offset direction and angle, outputting a section offset correction suggestion. The total number of preset offset categories was set to 2, the classification weight vector of the first category was set to 1.2, the classification weight vector of the second category was set to 0.5, and the natural constant was set to 2.718. Substituting these values ​​into the formula, the matching probability of the first category was calculated to be 0.636, and the matching probability of the second category was 0.364. The system selected the text instruction corresponding to the first category with the highest matching probability as the section offset correction suggestion; this text instruction was to shift to the left and rotate clockwise. The ultrasound operation stability score and the section offset correction suggestion were summarized, formatted, and a smart ultrasound operation assessment report was generated. The above closed-loop data calculation verified the rationality of the technical feature terminology and the accuracy of the formula derivation.

[0088] Reference Figure 2 The second aspect of the present invention provides an intelligent evaluation system for ultrasound operation based on contrast imaging temporal representation, comprising: an initial contrast imaging spatiotemporal feature generation module, a three-dimensional spatiotemporal evolution trajectory generation module, a contrast imaging temporal representation vector generation module, and an ultrasound operation evaluation report generation module.

[0089] The initial contrast imaging spatiotemporal feature generation module is connected to the three-dimensional spatiotemporal evolution trajectory generation module. Both the initial contrast imaging spatiotemporal feature generation module and the three-dimensional spatiotemporal evolution trajectory generation module are connected to the contrast imaging time sequence representation vector generation module. The contrast imaging time sequence representation vector generation module is connected to the ultrasound operation evaluation report generation module.

[0090] The initial contrast imaging spatiotemporal feature generation module acquires the ultrasound contrast imaging video stream and the corresponding probe spatial pose data, extracts the inter-frame pixel change features of the ultrasound contrast imaging video stream, and generates the initial contrast imaging spatiotemporal feature sequence.

[0091] The three-dimensional spatiotemporal evolution trajectory generation module extracts the anatomical structure contour from the initial contrast spatiotemporal feature sequence to obtain the dynamic boundary features of the target tissue, and performs spatial mapping in combination with probe spatial pose data to generate the three-dimensional spatiotemporal evolution trajectory of the target tissue.

[0092] The contrast time sequence characterization vector generation module analyzes the changes in contrast agent perfusion intensity in the three-dimensional spatiotemporal evolution trajectory of the target tissue, extracts perfusion time sequence characterization parameters, and performs feature fusion with the initial contrast spatiotemporal feature sequence to generate a contrast time sequence characterization vector.

[0093] The ultrasound operation assessment report generation module inputs the contrast time sequence characterization vector into the operation assessment model to extract quality features, outputs an ultrasound operation stability score and a section deviation correction suggestion, and generates an intelligent ultrasound operation assessment report.

[0094] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, and all such modifications and additions should fall within the protection scope of the present invention.

Claims

1. A method for intelligent evaluation of ultrasound operation based on contrast imaging time sequence characterization, characterized in that, include: S1. Acquire the ultrasound contrast imaging video stream and the corresponding probe spatial pose data, extract the inter-frame pixel change features of the ultrasound contrast imaging video stream, and generate the initial contrast imaging spatiotemporal feature sequence. S2. Extract the anatomical structure contour from the initial contrast spatiotemporal feature sequence to obtain the dynamic boundary features of the target tissue, and combine the probe spatial pose data for spatial mapping to generate the three-dimensional spatiotemporal evolution trajectory of the target tissue. S3. Analyze the changes in contrast agent perfusion intensity in the three-dimensional spatiotemporal evolution trajectory of the target tissue, extract perfusion time series characterization parameters, and perform feature fusion with the initial contrast spatiotemporal feature sequence to generate a contrast time series characterization vector. S4. Input the contrast time sequence characterization vector into the operation evaluation model to extract quality features, output the ultrasound operation stability score and section deviation correction suggestions, and generate an intelligent ultrasound operation evaluation report.

2. The intelligent evaluation method for ultrasound operation based on contrast imaging time sequence characterization according to claim 1, characterized in that, The process of acquiring the ultrasound contrast imaging video stream and the corresponding probe spatial pose data, extracting the inter-frame pixel change features of the ultrasound contrast imaging video stream, and generating an initial spatiotemporal feature sequence for contrast imaging includes: The continuous ultrasound image frames output by the ultrasound probe and the pose coordinates output by the spatial positioning sensor are collected to construct the ultrasound contrast video stream and probe spatial pose data. Calculate the pixel motion trajectory distribution of adjacent consecutive ultrasound image frames in an ultrasound contrast video stream and extract inter-frame pixel change features; The inter-frame pixel change features are serialized and stitched together according to the timestamp order to generate the initial imaging spatiotemporal feature sequence.

3. The intelligent evaluation method for ultrasound operation based on contrast imaging time sequence characterization according to claim 2, characterized in that, The acquisition of continuous ultrasound image frames output by the ultrasound probe and the pose coordinates output by the spatial positioning sensor are used to construct an ultrasound contrast imaging video stream and probe spatial pose data, including: The system receives the analog signal output by the ultrasound probe in contrast mode and performs analog-to-digital conversion to obtain continuous ultrasound image frames. The spatial positioning sensor fixed on the ultrasonic probe is read to measure the translation and rotation in three-dimensional space to obtain the pose coordinates. By aligning the timestamps of continuous ultrasound image frames with their pose coordinates and packaging the data, an ultrasound contrast video stream and probe spatial pose data are constructed.

4. The intelligent evaluation method for ultrasound operation based on contrast imaging time sequence characterization according to claim 1, characterized in that, The step of extracting the anatomical structure contour from the initial spatiotemporal feature sequence of the contrast imaging to obtain the dynamic boundary features of the target tissue, and combining this with the probe spatial pose data for spatial mapping to generate the three-dimensional spatiotemporal evolution trajectory of the target tissue, includes: Edge detection and morphological processing were performed on the initial spatiotemporal feature sequence of the contrast imaging to extract the contours of the anatomical structures. By tracing the deformation process of the anatomical structure outline over time, the dynamic boundary characteristics of the target tissue can be obtained. By using probe spatial pose data to construct a world coordinate system, the dynamic boundary features of the target tissue are transformed into the world coordinate system for spatial mapping, generating the three-dimensional spatiotemporal evolution trajectory of the target tissue.

5. The intelligent evaluation method for ultrasound operation based on contrast imaging time sequence characterization according to claim 4, characterized in that, The process of tracking the deformation of the anatomical structure contour over time yields the dynamic boundary features of the target tissue, including: The displacement vectors of each feature point in the anatomical structure contour at adjacent time nodes are calculated to obtain the contour motion field; Smoothing filtering and outlier removal are performed on the contour motion field to extract the effective deformation components; The effective deformation components are integrated along the time axis to obtain the dynamic boundary characteristics of the target tissue.

6. The intelligent evaluation method for ultrasound operation based on contrast imaging time sequence characterization according to claim 1, characterized in that, The analysis examines the changes in contrast agent perfusion intensity in the three-dimensional spatiotemporal evolution trajectory of the target tissue, extracts perfusion time-series characterization parameters, and fuses these parameters with the initial spatiotemporal feature sequence of contrast agents to generate a contrast time-series characterization vector, including: Calculate the rate of change of gray values ​​of each voxel over time within the three-dimensional spatiotemporal evolution trajectory of the target tissue to determine the change in contrast agent perfusion intensity; Curve fitting was performed on the changes in contrast agent perfusion intensity, and the time to peak and peak intensity were extracted as perfusion time series characterization parameters. The perfusion time-series characterization parameters and the initial angiography spatiotemporal feature sequence are concatenated by channels and attention weights are assigned to achieve feature fusion, generating an angiography time-series characterization vector.

7. The intelligent evaluation method for ultrasound operation based on contrast imaging time sequence characterization according to claim 6, characterized in that, The process of curve fitting the change in contrast agent perfusion intensity and extracting the peak time and peak intensity as perfusion time-series characterization parameters includes: The contrast agent perfusion intensity variation is mapped to the time-intensity coordinate system to generate discrete perfusion data points; A mathematical model describing the change of contrast agent concentration in tissue over time was used to perform nonlinear least squares fitting on discrete perfusion data points to obtain continuous perfusion curves. The highest point of the continuous perfusion curve is located, and the time and amplitude values ​​corresponding to the highest point are extracted as the peak time and peak intensity, which constitute the perfusion time series characterization parameters.

8. The intelligent evaluation method for ultrasound operation based on contrast imaging time sequence characterization according to claim 1, characterized in that, The process involves inputting the contrast-enhanced temporal representation vector into the operational evaluation model for quality feature extraction, outputting an ultrasound operation stability score and sectional deviation correction suggestions, and generating an intelligent ultrasound operation evaluation report, including: The imaging time sequence representation vector is input into the convolutional layer and recurrent neural network layer of the operation evaluation model to extract quality features and obtain the operation behavior feature map; The operation behavior feature map is classified and regressed to output an ultrasound operation stability score and a section deviation correction suggestion. The ultrasound operation stability score and section deviation correction suggestions are summarized, formatted, and generated into an intelligent ultrasound operation assessment report.

9. The intelligent evaluation method for ultrasound operation based on contrast imaging time sequence characterization according to claim 8, characterized in that, The classification and regression calculation of the operational behavior feature map, outputting an ultrasound operation stability score and section deviation correction suggestions, includes: The operation behavior feature map is input into a fully connected layer for dimensionality reduction to obtain the dimensionality reduction vector of the operation behavior. A regressor is used to numerically map the dimension-reduced vector of the operational behavior to calculate the ultrasound operation stability score. The classifier is used to perform pattern matching on the dimensionality-reduced vector of the operation behavior to determine the offset direction and angle, and output the cross-section offset correction suggestion.

10. An intelligent evaluation system for ultrasound operation based on contrast imaging time sequence characterization, characterized in that, include: The initial contrast imaging spatiotemporal feature generation module acquires the ultrasound contrast imaging video stream and the corresponding probe spatial pose data, extracts the inter-frame pixel change features of the ultrasound contrast imaging video stream, and generates the initial contrast imaging spatiotemporal feature sequence. The three-dimensional spatiotemporal evolution trajectory generation module extracts the anatomical structure contour from the initial contrast spatiotemporal feature sequence to obtain the dynamic boundary features of the target tissue, and performs spatial mapping in combination with probe spatial pose data to generate the three-dimensional spatiotemporal evolution trajectory of the target tissue. The contrast time sequence characterization vector generation module analyzes the changes in contrast agent perfusion intensity in the three-dimensional spatiotemporal evolution trajectory of the target tissue, extracts perfusion time sequence characterization parameters, and performs feature fusion with the initial contrast spatiotemporal feature sequence to generate a contrast time sequence characterization vector. The ultrasound operation assessment report generation module inputs the contrast time sequence characterization vector into the operation assessment model to extract quality features, outputs an ultrasound operation stability score and a section deviation correction suggestion, and generates an intelligent ultrasound operation assessment report.

Images