An abnormality analysis method and system based on ultrasonic detection images

Abstract

The application discloses an abnormality analysis method and system based on ultrasonic detection images, relates to the related field of ultrasonic image processing technology, and applies a wide-beam composite imaging method to collect ultrasonic detection images of each time point in a fixed time period, and obtains dynamic change data of a detection target; the ultrasonic detection images collected at different time points are preprocessed in view of the reflection characteristics of ultrasonic detection and motion interference in detection, and are further subjected to spatial registration and fusion; the method obtains a pre-trained abnormality analysis model through a reverse distillation mode, inputs the ultrasonic detection fusion images into the pre-trained model to calculate an abnormality score, and outputs an abnormality detection result; and the abnormality detection result output by the abnormality analysis model is used to assist in completing ultrasonic examination of a patient. The method solves the problem that the quality of ultrasonic detection images is poor, leading to a high misdiagnosis rate, reduces the influence of artifacts and noise on the quality of images, and improves the accuracy and efficiency of abnormality analysis of ultrasonic detection images.

Description

Technical Field

[0001] This application relates to the field of ultrasound image processing technology, and in particular to an anomaly analysis method and system based on ultrasound detection images. Background Technology

[0002] Medical ultrasound is a non-invasive medical examination technique that uses the physical properties of ultrasound waves to image the internal structures of the human body. When ultrasound waves propagate through the human body, they undergo reflection, refraction, absorption, and scattering due to differences in the density, elasticity, and other physical properties of different tissues. Ultrasound equipment obtains information about the internal structures of the human body by receiving these reflected ultrasound signals.

[0003] Traditional methods for analyzing abnormalities in medical ultrasound images rely heavily on the physician's experience, expertise, and familiarity with anatomical structures. This approach results in highly subjective outcomes, as the operator's technique, equipment settings, and ultrasound image quality all influence accuracy. Since the quality of ultrasound images significantly impacts examination results, further improvements in image quality are needed. Furthermore, the introduction of visual inspection-based intelligent algorithms can automate anomaly analysis, enhancing the ability to analyze and process dynamic image data. Summary of the Invention

[0004] To address the technical problems of the prior art, this application provides an anomaly analysis method and system based on ultrasound imaging. This method improves the imaging technique of ultrasound imaging, reduces the impact of artifacts on ultrasound imaging, and simultaneously acquires ultrasound images at different time points to observe the dynamic changes of the detected target. In the image preprocessing stage, improved image processing techniques are used to further enhance image quality and ensure the accuracy of ultrasound imaging data. Furthermore, a deep learning-based anomaly analysis model is introduced to perform real-time detection of abnormal ultrasound images, thereby improving the accuracy and efficiency of ultrasound detection and reducing the probability of misdiagnosis.

[0005] This application provides a method for anomaly analysis of ultrasound imaging, including:

[0006] (1) Using the wide-beam composite imaging method, ultrasound images of each time point within a fixed time period are acquired to obtain dynamic change data of the target being detected;

[0007] (2) To address the reflection characteristics of ultrasonic testing and motion interference during testing, ultrasonic testing images acquired at different time points are preprocessed to reduce noise and artifacts.

[0008] (3) Extract feature points from each ultrasound image to achieve spatial registration, and fuse ultrasound images acquired at different time points after registration;

[0009] (4) Obtain the pre-trained anomaly analysis model by reverse distillation, input the ultrasound detection fusion image into the pre-trained model to calculate the anomaly score, and output the anomaly detection result;

[0010] (5) Based on the abnormal detection results output by the abnormality analysis model, assist in completing the patient's ultrasound examination, collect feedback information, and optimize the hyperparameter settings of the abnormality analysis model.

[0011] Furthermore, wide-beam composite imaging transmits a series of divergent acoustic wave sequences with different deflection angles and synthesizes the images generated from each transmission to form a high-quality composite image. This method not only improves spatial resolution and signal-to-noise ratio but also effectively reduces motion artifacts, maintaining high image quality even in deeper tissue layers. The steps for implementing wide-beam composite ultrasound imaging include:

[0012] Select an ultrasonic imaging device with wide beam transmission and multi-beam reception capabilities, select the ultrasonic probe type according to the ultrasonic detection target, and set the probe parameters;

[0013] The ultrasonic probe unit of the ultrasonic imaging device is divided into multiple groups, each group focusing on a different depth position of the target being detected, forming multiple sub-focal points;

[0014] The emission delay of each probe element is calculated based on the position, focal depth, and probe type of the ultrasonic probe element, so that the sound wave is focused at the predetermined focal point;

[0015] According to the calculated transmission delay, an ultrasonic beam is emitted through the ultrasonic probe element, and the waveforms of multiple transmissions are added together to form a wide-beam composite transmission waveform.

[0016] After emitting sound waves, the ultrasonic probe switches to a receiving mode to receive echo signals from the target area. Multiple receiving channels simultaneously receive echoes from multiple scan lines.

[0017] Beamforming is performed on the echo signals from each receiving channel. A time-delay superposition method is used to merge the signals from multiple sub-beams into the final imaging signal. An image reconstruction algorithm is then used to generate an ultrasound detection image.

[0018] Furthermore, in the preprocessing of ultrasound images, directional weight compensation and Bayesian nonlocal means optimization algorithms are used to handle artifacts and noise in the images. Bayesian nonlocal means optimization is a highly efficient denoising algorithm that fully utilizes nonlocal similarity information in the image, thereby preserving image details and texture as much as possible while denoising.

[0019] When ultrasound waves propagate through tissue, they are reflected multiple times back to the probe. These reflected signals overlap with the real signal, creating reverberation artifacts. Secondly, the non-main beam direction of the ultrasound probe also generates echo signals, interfering with the imaging of the main beam and creating sidelobe artifacts. Furthermore, differences in density, sound velocity, and absorption characteristics within human tissues, as well as patient movement during the examination, all contribute to artifacts in the ultrasound images. Noise, on the other hand, arises from the scattering of ultrasound waves by tiny inhomogeneous structures within the tissue, producing chaotic echo signals that affect image quality.

[0020] Furthermore, the detailed steps for reducing ultrasound imaging artifacts using a directional weight compensation method include:

[0021] The ultrasound images were converted to grayscale images, and histogram equalization was introduced to enhance the contrast of the images so as to more clearly identify artifacts and tissue structures.

[0022] By analyzing the intensity distribution and texture features of ultrasound images, regions containing artifacts can be identified: reverberation artifacts appear as periodic strong echo signals and are usually presented as randomly distributed spots.

[0023] Analysis of artifact propagation direction: The direction of reverberation artifact is consistent with the direction of ultrasonic wave propagation, while the direction of side lobe artifact is at a certain angle to the main beam.

[0024] Based on the direction and intensity of artifacts, a weighting function is designed: smaller weights are assigned to the artifact direction to suppress artifact signals; larger weights are assigned to the non-artifact direction to preserve real tissue signals.

[0025] For each direction of the signal in the image, the signal is multiplied by the corresponding weight according to the relationship between its direction and the artifact direction, and the weighted image signal is summed to obtain the final compensated image.

[0026] Furthermore, the detailed steps for reducing ultrasound imaging noise using the Bayesian nonlocal means optimization algorithm include:

[0027] The ultrasound detection compensation image is divided into overlapping image blocks, each containing multiple pixels;

[0028] A search window is defined for each image patch to search for similar image patches in the compensated image. The size of the search window is larger than the size of the image patch.

[0029] For each reference image patch, calculate its similarity with other image patches in the search window, and select image patches with a similarity metric greater than a threshold to form a set of similar image patches;

[0030] Calculate the prior mean and covariance of a set of similar image patches, and use Bayes' theorem, combined with prior knowledge and image patch pixel data, to update the pixel values ​​of the image patches. Then, aggregate the denoising results of each image patch into the entire image.

[0031] Furthermore, based on the ultrasound images acquired at different time points after preprocessing, a spatial registration method based on feature point matching is adopted to reduce the displacement effect caused by body movement during the detection process. The specific process includes:

[0032] The Harris corner detection method is used to extract feature points from ultrasound images. These feature points should have significant geometric or textural features and be stably identifiable in images at different time points.

[0033] Extract a descriptor for each feature point for subsequent feature matching: the descriptor can uniquely represent the local information of the feature point;

[0034] The matching algorithm is used to match feature points in the reference image with feature points in other target images to select reliable matching point pairs.

[0035] Using the selected matching point pairs, the parameters of the rigid transformation model are calculated by the least squares method;

[0036] Based on the calculated transformation parameters, a rigid transformation is performed on the target image to align it with the reference image;

[0037] The transformed image is resampled to eliminate pixel distortion introduced by the transformation.

[0038] Furthermore, the anomaly analysis model obtains more discriminative feature representations by generating and exploring samples that exceed the normal distribution. It also addresses the scale variation problem of anomalies by soft-weighting the contrastive backdistillation loss at different scales. The detailed steps for obtaining the pre-trained anomaly analysis model include:

[0039] Normal ultrasound images are used as normal training samples. The location and magnitude of noise are randomly sampled and scaled and superimposed onto the normal training samples to form the corresponding abnormal training samples.

[0040] The clean teacher model encoder in the anomaly analysis model training framework extracts multi-scale normal features from normal training samples, while the noisy teacher model encoder extracts multi-scale anomalous features from anomalous training samples.

[0041] The fusion module in the anomaly analysis model training framework splices together the multi-scale normal features of the clean teacher model encoder and outputs a joint representation. The scale adaptation mechanism generates scale weights from the joint representation, and the student model decoder reconstructs the multi-scale features.

[0042] Calculate the comparative backdistillation loss to make the reconstructed multi-scale features output by the student model decoder approximate the normal multi-scale features and move away from the abnormal multi-scale features.

[0043] The parameters of the student model, fusion module, and scale adaptation mechanism are updated by backpropagation using the stochastic gradient descent algorithm. When the loss function converges, the model parameters are retained as a pre-trained anomaly analysis model.

[0044] This application also provides an anomaly analysis system for ultrasound imaging, including:

[0045] Ultrasonic Detection Imaging Module: Utilizes a wide-beam composite imaging method to acquire ultrasonic detection images at each time point within a fixed time period, obtaining dynamic change data of the detected target;

[0046] Image data processing module: Used to preprocess ultrasound images acquired at different time points to reduce noise and artifacts, taking into account the reflection characteristics of ultrasound detection and motion interference during detection.

[0047] Image fusion and evaluation module: used to extract feature points from each ultrasound inspection image to achieve spatial registration, and fuse ultrasound inspection images acquired at different time points after registration;

[0048] Image Anomaly Analysis Module: This module is used to obtain a pre-trained anomaly analysis model through back-distillation. It inputs the ultrasound detection fused images into the pre-trained model to calculate the anomaly score and outputs the anomaly detection results.

[0049] Feedback optimization module: Based on the abnormality detection results output by the abnormality analysis model, it assists in completing the patient's ultrasound examination, collects feedback information, and optimizes the hyperparameter settings of the abnormality analysis model.

[0050] The present invention discloses the following technical effects:

[0051] This invention proposes an anomaly analysis method and system based on ultrasound imaging. By improving the imaging method, image preprocessing algorithm, and model training method, it enhances the image quality of ultrasound imaging, thereby improving the accuracy and efficiency of anomaly analysis. This invention employs a wide-beam composite imaging technique, transmitting and receiving echo signals from multiple channels to comprehensively cover the detection target and reduce the impact of motion artifacts on image quality. Furthermore, by acquiring ultrasound imaging data at different time points based on the improved imaging method, the dynamic changes of the detection target can be observed, providing physicians with comprehensive patient examination information. In the preprocessing stage, directional weight compensation and Bayesian nonlocal means optimization algorithms are used to process artifacts and noise in the images, preserving the local texture features of the detection target while minimizing noise and artifacts. This invention introduces a deep learning-based anomaly analysis model, using backdistillation training to obtain a pre-trained model for analyzing real-time acquired ultrasound imaging, improving the automation and intelligence level of ultrasound imaging anomaly analysis and ensuring high precision and efficiency. Attached Figure Description

[0052] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. Flowcharts are used in this application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.

[0053] Figure 1 This is a flowchart illustrating an anomaly analysis method based on ultrasound imaging provided in an embodiment of this application.

[0054] Figure 2 This is a schematic diagram of the training framework for the anomaly analysis model provided in the embodiments of this application.

[0055] Figure 3 A detailed structural diagram of the fusion module and scale adaptation mechanism in the anomaly analysis model provided in the embodiments of this application.

[0056] Figure 4 This is a schematic diagram of an anomaly analysis system based on ultrasound detection images, provided as an embodiment of this application. Detailed Implementation

[0057] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application.

[0058] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description of this application will be provided in conjunction with the accompanying drawings. The described embodiments should not be considered as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0059] In the following description, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those explicitly listed, but may include other steps or modules not explicitly listed or inherent to such processes, methods, products, or devices. 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 application belongs. The terminology used herein is for the purpose of describing embodiments of this application only.

[0060] Example 1: This application provides an anomaly analysis method based on ultrasound imaging, such as... Figure 1 As shown, the method includes:

[0061] Step S10: Apply the wide-beam composite imaging method to acquire ultrasound detection images at each time point within a fixed time period to obtain dynamic change data of the detection target.

[0062] In this embodiment, the steps for implementing wide-beam composite ultrasound imaging include:

[0063] Select an ultrasonic imaging device with wide-beam transmission and multi-beam reception capabilities, choose the ultrasonic probe type according to the ultrasonic detection target, and set the probe parameters:

[0064] The ultrasonic imaging system includes an ultrasonic probe, transmitting and receiving circuits, a signal processing unit, and image reconstruction software; a linear array ultrasonic probe is selected, and the probe aperture size and the number of array elements are set.

[0065] The ultrasonic probe of the ultrasonic imaging device is divided into multiple groups, each group focusing on a different depth position of the target, forming multiple sub-focal points:

[0066] With the position of the transmission line as the midpoint, each pair of symmetrically arranged ultrasonic probe elements on both sides forms a probe element group, and each probe element group focuses on each set focal point in sequence according to the arrangement order.

[0067] At the emission line location, foci are arranged at equal intervals along the depth direction: Assuming the number of ultrasonic probe elements is N, the following settings are made: The floor(·) function represents rounding down to the nearest integer, where the focal points are arranged at equal intervals.

[0068] The emission delay of each probe element is calculated based on its position, focal depth, and probe type, so that the sound waves are focused at a predetermined focal point:

[0069] For a linear array ultrasonic probe, the emission delay T delay The calculation formula is:

[0070]

[0071] Among them, F m For the depth of focus, d x denoted as , where is the lateral distance between the probe element and the transmitting line, and c is the ultrasonic wave propagation speed.

[0072] Based on the calculated transmission delay, an ultrasonic beam is emitted through the ultrasonic probe element. The waveforms of multiple transmissions are added together to form a wide-beam composite transmission waveform:

[0073] After emitting sound waves, the ultrasonic probe switches to a receiving mode to receive echo signals from the target area. Multiple receiving channels simultaneously receive echoes from multiple scan lines.

[0074] Beamforming is performed on the echo signals from each receiving channel. A time-delay superposition method is used to merge the signals from multiple sub-beams into the final imaging signal. An image reconstruction algorithm is then used to generate an ultrasound detection image.

[0075] The specific implementation steps of the delay superposition method include:

[0076] Calculate the delay for each receiving channel. Assume the target point of the ultrasonic detection target is located at depth z and lateral position g, the center of the ultrasonic probe is located at the origin, and the probe element spacing is d. Then, the delay τ for the i-th receiving channel is calculated. i The calculation formula is as follows:

[0077]

[0078] The received signal is digitized, and a delay is achieved through digital signal processing. Let the discrete signal be s. i (n), the delayed signal is obtained in the following way:

[0079]

[0080] in, and They are τ i The floor and floor functions, {τ} i} is τ i The decimal part;

[0081] The final imaging signal S(n) is obtained by superimposing all the delayed signals, and gain compensation is performed on the imaging signal: S comp (n) = S(n)·G(z), where S comp (n) is the image signal after gain compensation, and G(z) is the gain function related to depth z, which is a monotonically increasing function.

[0082] The amplified imaging signal is sampled to obtain the pixel values ​​of the ultrasound inspection image. The sampling interval is determined according to the imaging resolution. The pixel values ​​of all points are combined to form a two-dimensional ultrasound inspection image.

[0083] Step S20: Based on the reflection characteristics of ultrasonic testing and motion interference during testing, preprocess the ultrasonic testing images acquired at different time points to reduce noise and artifacts.

[0084] In this embodiment, during the preprocessing of ultrasound images, directional weight compensation method and Bayesian nonlocal mean optimization algorithm are used to process artifacts and noise in the images.

[0085] The detailed steps for reducing ultrasound imaging artifacts using a directional weight compensation method include:

[0086] The ultrasound images were converted to grayscale images, and histogram equalization was introduced to enhance the contrast of the images so as to more clearly identify artifacts and tissue structures.

[0087] By analyzing the intensity distribution and texture features of ultrasound images, regions containing artifacts can be identified: reverberation artifacts appear as periodic strong echo signals and are usually presented as randomly distributed spots.

[0088] Analysis of artifact propagation direction: The direction of reverberation artifact is consistent with the direction of ultrasonic wave propagation, while the direction of side lobe artifact is at a certain angle to the main beam.

[0089] Based on the direction and intensity of artifacts, a weighting function is designed: smaller weights are assigned to the artifact direction to suppress artifact signals; larger weights are assigned to the non-artifact direction to preserve real tissue signals.

[0090] The weighting function ε(θ) is expressed by the formula:

[0091]

[0092] Where θ is the current direction of the target point in the ultrasound image, θ w It is the artifact direction, δ is the weight decay coefficient, and θ t This refers to the range of weight decay.

[0093] For each direction of the signal in the image, the signal is multiplied by the corresponding weight according to the relationship between its direction and the artifact direction, and the weighted image signal is summed to obtain the final compensated image.

[0094] The detailed steps for reducing noise in ultrasound imaging using the Bayesian nonlocal means optimization algorithm include:

[0095] The ultrasound detection compensation image is divided into overlapping image blocks, each containing multiple pixels;

[0096] A search window is defined for each image patch to search for similar image patches in the compensated image. The size of the search window is larger than the size of the image patch.

[0097] In this embodiment, the image patch size is 3×3, and the search window size is 17×17.

[0098] For each reference image patch, calculate its similarity to other image patches within the search window, and select image patches with a similarity metric greater than a threshold to form a set of similar image patches:

[0099] Using Gaussian-weighted Euclidean distance β uv Measuring similarity:

[0100]

[0101] Among them, P u and P v These represent the target image patch and other image patches within the search window, respectively, in matrix form. h is the standard deviation of the Gaussian kernel, which controls the decay rate of the weights.

[0102] Calculate the prior mean and covariance of the set of similar image patches. Using Bayes' theorem, combined with prior knowledge and image patch pixel data, update the pixel values ​​of the image patches. Then, aggregate the denoising results of each image patch into the entire image.

[0103] For each reference image block P u Calculate the prior mean and covariance matrix of the set of similar image patches:

[0104]

[0105] Where, μ u and Σ u Let S represent the prior mean and covariance matrix, respectively. u Is with P u A set of similar image patches; the process of updating pixel values ​​in an image patch is expressed by the formula:

[0106]

[0107] in, σ represents the denoised target image patch. 2 Time-noise variance, 1 is the identity matrix; the denoising result of each image patch. Aggregate the results into the entire image; if multiple image patches cover the same pixel, these results can be weighted and averaged.

[0108] Step S30: Extract feature points from each ultrasound image to achieve spatial registration, and fuse ultrasound images acquired at different time points after registration.

[0109] In this embodiment, based on ultrasound images acquired at different time points after preprocessing, a spatial registration method based on feature point matching is used to reduce the displacement effect caused by body movement during the detection process. The specific process includes:

[0110] The Harris corner detection method was used to extract feature points from ultrasound images. These feature points should have significant geometric or textural features and be stably identifiable in images at different time points.

[0111] The Sobel operator is used to calculate the gradients of ultrasound images in the horizontal and vertical directions. This operator is expressed as:

[0112]

[0113] Among them, G x and G y Let l represent the horizontal Sobel operator and the vertical Sobel operator, respectively. The horizontal gradient l is obtained by convolving these operators with the ultrasound image. x and vertical gradient I y ;

[0114] Choosing a Gaussian window function to calculate the autocorrelation matrix of the gradient within a local region, the window function Win is represented as:

[0115]

[0116] Where, σ g is the standard deviation of the Gaussian function, where x and y represent the x and y coordinates of the pixel, respectively; the formula for calculating the autocorrelation matrix M is:

[0117]

[0118] Where (l,m) is the pixel position within the window;

[0119] Calculate the corner response function R: R = det(M) - γ(trace(M)) 2, where det(M) and trace(M) represent the determinant and trace of the autocorrelation matrix, respectively, and γ is an empirical constant with a value ranging from 0.04 to 0.06.

[0120] Set a threshold, and select points whose response function is greater than the threshold as candidate corner points. Among the candidate corner points, remove duplicate corner points by non-maximum suppression, and the remaining corner points are used as feature points.

[0121] Descriptors are extracted for each feature point for subsequent feature matching: the descriptor uniquely represents the local information of the feature point.

[0122] For each feature point, the gradient direction in the neighborhood is calculated as the main direction of the feature point; a 31×31 region is sampled with each feature point as the center.

[0123] 256 pairs of pixels are randomly selected within the region. The gray values ​​of each pair of pixels are compared: if the gray value of the first pixel is greater than that of the second pixel, it is recorded as 1 in the binary vector; otherwise, it is recorded as 0. A 256-dimensional binary vector is generated as the BRIEF descriptor for the feature point.

[0124] The BFMatcher matching algorithm is used to match feature points in the reference image with feature points in other target images to select reliable matching point pairs.

[0125] Using the selected matching point pairs, the parameters of the rigid transformation model are calculated by the least squares method. The parameters include the translation vector and the rotation matrix.

[0126] Based on the calculated transformation parameters, a rigid transformation is performed on the target image to align it with the reference image;

[0127] The transformed image is resampled to eliminate pixel distortion introduced by the transformation.

[0128] For multiple registered ultrasound images from different time points, the corresponding pixels of each image are weighted and fused: the key time points for the detection of the target are determined, and the ultrasound images at the key time points are assigned a weight greater than 0.5. Let there be m ultrasound images from different time points O1, O2, ..., O m The corresponding weights are ω1, ω2, ... ω m The value of each pixel in the fused image is calculated using the following formula:

[0129]

[0130] Among them, O j (a,b) is the pixel value at position (a,b) in the j-th image, ω jIt is the weight of the j-th image.

[0131] Step S40: Obtain the pre-trained anomaly analysis model through reverse distillation, input the ultrasound detection fusion image into the pre-trained model to calculate the anomaly score, and output the anomaly detection result.

[0132] In this embodiment, the detailed steps for obtaining the pre-trained anomaly analysis model include:

[0133] Normal ultrasound images are used as normal training samples. The location and magnitude of noise are randomly sampled and scaled and superimposed onto the normal training samples to form the corresponding abnormal training samples.

[0134] In the anomaly analysis model training framework, the clean teacher model encoder extracts multi-scale normal features from normal training samples, while the noisy teacher model encoder extracts multi-scale anomalous features from anomalous training samples.

[0135] The fusion module in the anomaly analysis model training framework fuses the features of the clean teacher model encoder and outputs a joint representation. The scale adaptation mechanism generates scale weights from the joint representation, and the student model decoder reconstructs multi-scale features based on the scale weights.

[0136] Calculate the comparative backdistillation loss to make the reconstructed multi-scale features output by the student model decoder approximate the normal multi-scale features and move away from the abnormal multi-scale features.

[0137] The parameters of the student model, fusion module, and scale adaptation mechanism are updated by backpropagation using the stochastic gradient descent algorithm. When the loss function converges, the model parameters are retained as a pre-trained anomaly analysis model.

[0138] Step S50: Based on the abnormal detection results output by the abnormality analysis model, assist in completing the patient's ultrasound examination, collect feedback information, and optimize the hyperparameter settings of the abnormality analysis model.

[0139] Example 2: This application provides an anomaly analysis method based on ultrasound imaging. The training framework of the anomaly analysis model is as follows: Figure 2 As shown:

[0140] In this embodiment, the training framework comprises two teacher models and a student model, which generate and explore samples that deviate from the normal distribution to obtain more discriminative feature representations. The teacher models are divided into a clean teacher model and a noisy teacher model, whose input features are normal training samples and abnormal training samples of ultrasound detection images, respectively.

[0141] The encoder in the clean teacher model extracts multi-scale normal features from normal training samples. The encoder contains a 3×3 kernel convolutional layer, a batch normalization layer, a ReLU activation function, and a max pooling layer. The pooling layer is used to generate intermediate features of different sizes. The encoder of the noisy teacher model and the decoder of the student model adopt the same structure as the encoder of the clean teacher model. They are used to extract multi-scale abnormal features from abnormal training samples and generate reconstructed multi-scale features, respectively.

[0142] The clean teacher model is followed by a bottleneck-based fusion module, the detailed structure of which is as follows: Figure 3 The upper part shows that the three normal features at different scales output by the clean teacher model encoder are used as input to the module. First, the three features are unified to the same size through convolution operation. Then, the three features are concatenated along the channel dimension. The scale information between the features is fused through 3×3 convolution depth and the channel dimension is reduced. Finally, the fused feature φ is obtained through the residual module.

[0143] To address the scale variation issue in abnormalities within ultrasound imaging, the obtained fused features are fed into a learnable scale adaptation mechanism to generate scale description features α = [α1,...,α]. K ], where K represents the number of layers. The detailed structure of this mechanism is as follows: Figure 3 The lower half is shown below:

[0144] This process employs global max pooling to spatially compress the feature map φ, generating channel-level statistics. Since φ can be considered a joint representation of multi-scale features, its statistics can effectively reflect the cross-scale content information of the image. Subsequently, a linear layer with softmax normalization maps the channel-level statistics to the scale descriptor α. This process is described as follows:

[0145] α = softmax(W·GMP(φ))

[0146] Where softmax represents softmax normalization; GMP represents global max pooling; and W is the weight matrix of the linear layer.

[0147] Finally, a scale-aware contrastive backdistillation loss function is used. Represented as:

[0148]

[0149] Where, α k As a contrastive input-specific scale weight for backdistillation, a dynamic mechanism based on input condition adjustment is introduced, which helps improve the model's discriminative ability; sim(·) represents cosine similarity, u k z k and v kLet represent the output features of the clean teacher model encoder, the noisy teacher model encoder, and the student model decoder, respectively. Converting these features to a one-dimensional representation facilitates the calculation of cosine similarity; ∈ is the minimum value added to the denominator to avoid division by zero errors during calculation. The goal of this loss function is to reduce u... k and v k The distance is reduced, and Z is simultaneously k and v k The distance is pushed further away.

[0150] In actual anomaly analysis, the anomaly analysis model calculates u at each scale k. k and v k The vector cosine similarity between them. The generated similarity map is then used to calculate the corresponding α. k Weighted similarity maps are applied and upsampled to the input image resolution; weighted similarity maps at all scales are aggregated to construct a complete anomaly map; the anomaly score of the input image is obtained by calculating the average value of this aggregated anomaly map.

[0151] Example 3: The anomaly analysis system based on ultrasound imaging provided in this embodiment of the invention can execute the anomaly analysis method based on ultrasound imaging provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method. The detailed structure of the system is as follows: Figure 4 As shown:

[0152] Ultrasonic Detection Imaging Module: Utilizes a wide-beam composite imaging method to acquire ultrasonic detection images at each time point within a fixed time period, obtaining dynamic change data of the detected target;

[0153] Image data processing module: Used to preprocess ultrasound images acquired at different time points to reduce noise and artifacts, taking into account the reflection characteristics of ultrasound detection and motion interference during detection.

[0154] Image fusion and evaluation module: used to extract feature points from each ultrasound inspection image to achieve spatial registration, and fuse ultrasound inspection images acquired at different time points after registration;

[0155] Image Anomaly Analysis Module: This module is used to obtain a pre-trained anomaly analysis model through back-distillation. It inputs the ultrasound detection fused images into the pre-trained model to calculate the anomaly score and outputs the anomaly detection results.

[0156] Feedback optimization module: Based on the abnormality detection results output by the abnormality analysis model, it assists in completing the patient's ultrasound examination, collects feedback information, and optimizes the hyperparameter settings of the abnormality analysis model.

[0157] Although this application makes various references to certain modules in the system according to the embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy distinction between each other and are not used to limit the scope of protection of this invention.

[0158] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application. In some cases, the actions or steps described in this application can be performed in a different order than that shown in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

Claims

1. An anomaly analysis method based on ultrasound imaging, characterized in that, The method includes: (1) Using the wide-beam composite imaging method, ultrasound images of each time point within a fixed time period are acquired to obtain dynamic change data of the target being detected; (2) To address the reflection characteristics of ultrasonic testing and motion interference during testing, ultrasonic testing images acquired at different time points are preprocessed to reduce noise and artifacts. (3) Extract feature points from each ultrasound image to achieve spatial registration, and fuse ultrasound images acquired at different time points after registration; (4) Obtain the pre-trained anomaly analysis model by reverse distillation, input the ultrasound detection fusion image into the pre-trained model to calculate the anomaly score, and output the anomaly detection result; (5) Based on the abnormal detection results output by the abnormality analysis model, assist in completing the patient's ultrasound examination, collect feedback information, and optimize the hyperparameter settings of the abnormality analysis model.

2. The anomaly analysis method based on ultrasound imaging as described in claim 1, characterized in that, In step (1), the wide-beam composite ultrasound imaging step includes: Select an ultrasonic imaging device with wide beam transmission and multi-beam reception capabilities, select the ultrasonic probe type according to the ultrasonic detection target, and set the probe parameters; The ultrasonic probe unit of the ultrasonic imaging device is divided into multiple groups, each group focusing on a different depth position of the target being detected, forming multiple sub-focal points; The emission delay of each probe element is calculated based on the position, focal depth, and probe type of the ultrasonic probe element, so that the sound wave is focused at the predetermined focal point; According to the calculated transmission delay, an ultrasonic beam is emitted through the ultrasonic probe element, and the waveforms of multiple transmissions are added together to form a wide-beam composite transmission waveform. After emitting sound waves, the ultrasonic probe switches to a receiving mode to receive echo signals from the target area. Multiple receiving channels simultaneously receive echoes from multiple scan lines. Beamforming is performed on the echo signals from each receiving channel. A time-delay superposition method is used to merge the signals from multiple sub-beams into the final imaging signal. An image reconstruction algorithm is then used to generate an ultrasound detection image.

3. The anomaly analysis method based on ultrasound imaging as described in claim 2, characterized in that, The specific implementation steps of the delay superposition method include: Calculate the delay for each receiving channel. Assume the target point of the ultrasonic detection target is located at depth z and lateral position g, the center of the ultrasonic probe is located at the origin, and the probe element spacing is d. Then, the delay τ for the i-th receiving channel is calculated. i The calculation formula is as follows: Where c is the speed at which ultrasound propagates in the medium; The received signal is digitized, and a delay is achieved through digital signal processing. Let the discrete signal be s. i (n), the delayed signal is obtained in the following way: in, and They are τ i The floor and floor functions, {τ} i } is τ i The decimal part; The final imaging signal S(n) is obtained by superimposing all the delayed signals, and gain compensation is performed on the imaging signal: S comp (n) = S(n)·G(z), where S comp (n) is the image signal after gain compensation, and G(z) is the gain function related to depth z, which is a monotonically increasing function.

4. The anomaly analysis method based on ultrasound imaging as described in claim 1, characterized in that, In step (2), a directional weight compensation method is used to reduce artifacts in ultrasound images. The detailed steps include: The ultrasound images were converted to grayscale images, and histogram equalization was introduced to enhance the image contrast. By analyzing the intensity distribution and texture features of ultrasound images, regions containing artifacts can be identified: reverberation artifacts appear as periodic strong echo signals and are usually presented as randomly distributed spots. Analysis of artifact propagation direction: The direction of reverberation artifact is consistent with the direction of ultrasonic wave propagation, while the direction of side lobe artifact is at a certain angle to the main beam. Based on the direction and intensity of artifacts, a weighting function is designed to assign different weights in the artifact direction and non-artifact direction, thereby suppressing artifact signals and preserving real tissue signals. For each direction of the signal in the image, the signal is multiplied by the corresponding weight according to the relationship between its direction and the artifact direction, and the weighted image signal is summed to obtain the final compensated image.

5. The anomaly analysis method based on ultrasound imaging as described in claim 1, characterized in that, In step (2), the Bayesian nonlocal mean optimization algorithm is used to reduce noise in the ultrasound imaging. The detailed steps include: The ultrasound detection compensation image is divided into overlapping image blocks, each containing multiple pixels; Define a search window for each image patch, with the search window size being larger than the image patch size; For each reference image patch, calculate its similarity to other image patches within the search window, and select image patches with a similarity metric greater than a threshold to form a set of similar image patches: Using Gaussian-weighted Euclidean distance β uv Measuring similarity: Among them, P u and P v These represent the target image patch and other image patches within the search window, respectively, in matrix form. h is the standard deviation of the Gaussian kernel, which controls the decay rate of the weights. Calculate the prior mean and covariance of the set of similar image patches. Using Bayes' theorem, combined with prior knowledge and image patch pixel data, update the pixel values ​​of the image patches. Then, aggregate the denoising results of each image patch into the entire image. For each reference image block P u Calculate the prior mean and covariance matrix of the set of similar image patches: Where, μ u and ∑ u Let S represent the prior mean and covariance matrix, respectively. u Is with P u A set of similar image patches; the process of updating pixel values ​​in an image patch is expressed by the formula: in, σ represents the denoised target image patch. 2 Time-noise variance, 1 is the identity matrix; the denoising result of each image patch. Aggregate into the entire image.

6. The anomaly analysis method based on ultrasound imaging as described in claim 1, characterized in that, In step (3), based on the ultrasound images acquired at different time points after preprocessing, a spatial registration method based on feature point matching is adopted to reduce the displacement effect caused by body movement during the detection process. The specific process includes: Feature points in ultrasound images were extracted using the Harris corner detection method. Extract descriptors for each feature point for subsequent feature matching; The matching algorithm is used to match feature points in the reference image with feature points in other target images to select reliable matching point pairs. Using the selected matching point pairs, the parameters of the rigid transformation model are calculated by the least squares method; Based on the calculated transformation parameters, a rigid transformation is performed on the target image to align it with the reference image, and the transformed image is then resampled.

7. The anomaly analysis method based on ultrasound imaging as described in claim 6, characterized in that, The process of extracting feature points using the Harris corner detection method includes: The Sobel operator is used to calculate the gradients of ultrasound images in the horizontal and vertical directions. This operator is expressed as: Among them, G x and G y Let l represent the horizontal Sobel operator and the vertical Sobel operator, respectively. The horizontal gradient l is obtained by convolving these operators with the ultrasound image. x and vertical gradient I y ; Choosing the Gaussian window function, the autocorrelation matrix of the gradient is calculated within a local region. The window function Win is represented as: Where, σ g is the standard deviation of the Gaussian function, where x and y represent the x and y coordinates of the pixel, respectively; the formula for calculating the autocorrelation matrix M is: Where (l,m) is the pixel position within the window; Calculate the corner response function R: R = det(M) - γ(trace(M)) 2 Where det(M) and trace(M) represent the determinant and trace of the autocorrelation matrix, respectively, and γ is an empirical constant with a value ranging from 0.04 to 0.06; Set a threshold, and select points whose response function is greater than the threshold as candidate corner points. Among the candidate corner points, remove duplicate corner points by non-maximum suppression, and the remaining corner points are used as feature points.

8. The anomaly analysis method based on ultrasound imaging as described in claim 1, characterized in that, Step (4) involves obtaining the detailed steps for the pre-trained anomaly analysis model, including: Normal ultrasound images are used as normal training samples. The location and magnitude of noise are randomly sampled and scaled and superimposed onto the normal training samples to form the corresponding abnormal training samples. The clean teacher model encoder in the anomaly analysis model training framework extracts multi-scale normal features from normal training samples, while the noisy teacher model encoder extracts multi-scale anomalous features from anomalous training samples. The fusion module in the anomaly analysis model training framework splices together the multi-scale normal features of the clean teacher model encoder and outputs a joint representation. The scale adaptation mechanism generates scale weights from the joint representation, and the student model decoder reconstructs the multi-scale features. Calculate the comparative backdistillation loss to make the reconstructed multi-scale features output by the student model decoder approximate the normal multi-scale features and move away from the abnormal multi-scale features. The parameters of the student model, fusion module, and scale adaptation mechanism are updated by backpropagation using the stochastic gradient descent algorithm. When the loss function converges, the model parameters are retained as a pre-trained anomaly analysis model.

9. The anomaly analysis method based on ultrasound imaging as described in claim 8, characterized in that, The contrasting reverse distillation loss Represented as: Where, α k As a contrastive input-specific scale weight for backdistillation, a dynamic mechanism based on input condition adjustment is introduced; sim(·) represents cosine similarity, u k z k and v k represents the output features of the clean teacher model encoder, the noisy teacher model encoder, and the student model decoder, respectively, and converts them into one-dimensional representations; ∈ is the minimum value added to the denominator to avoid division by zero errors in the calculation.

10. An anomaly analysis system based on ultrasound imaging, characterized in that, The system is used to implement the anomaly analysis method based on ultrasound detection images according to any one of claims 1-9, the system comprising: Ultrasonic Detection Imaging Module: Utilizes a wide-beam composite imaging method to acquire ultrasonic detection images at each time point within a fixed time period, obtaining dynamic change data of the detected target; Image data processing module: Used to preprocess ultrasound images acquired at different time points to reduce noise and artifacts, taking into account the reflection characteristics of ultrasound detection and motion interference during detection. Image fusion and evaluation module: used to extract feature points from each ultrasound inspection image to achieve spatial registration, and fuse ultrasound inspection images acquired at different time points after registration; Image Anomaly Analysis Module: This module is used to obtain a pre-trained anomaly analysis model through back-distillation. It inputs the ultrasound detection fused images into the pre-trained model to calculate the anomaly score and outputs the anomaly detection results. Feedback optimization module: Based on the abnormality detection results output by the abnormality analysis model, it assists in completing the patient's ultrasound examination, collects feedback information, and optimizes the hyperparameter settings of the abnormality analysis model.

Images