Pet abdominal ultrasound detection method and system based on image analysis

Abstract

This invention discloses a method and system for detecting abdominal ultrasound in pets based on image analysis, belonging to the field of pet abdominal ultrasound image processing technology. This method maps the ultrasound image to the complex frequency domain using a two-dimensional fast Fourier transform, extracting the spatial phase spectrum. Using the zero-frequency center point as the origin of the emission point, low-frequency components of the solid organ contours are filtered out. Non-axial radial sampling rays are traversed within a preset high-frequency band, extracting the phase value sequence and calculating the autocorrelation coefficient. The feature direction angle corresponding to the maximum value is retrieved. A two-dimensional Gaussian filter mask is constructed based on this, and it is multiplied element-wise with a two-dimensional complex frequency domain matrix to specifically reduce physical grid stripe noise in that direction. Finally, a two-dimensional inverse Fourier transform is used to restore the image to the spatial domain and reconstruct an edge suppression feature map, which is then input into a machine learning model to output an auxiliary probability report of the lesion. This invention specifically filters out structured physical grid noise caused by residual hair roots in pets, significantly improving the accuracy of lesion identification.

Description

Technical Field

[0001] This invention relates to the field of pet abdominal ultrasound image processing technology, and more specifically, this application relates to a pet abdominal ultrasound detection method and system based on image analysis. Background Technology

[0002] In clinical diagnosis and treatment of abdominal organs in dogs and cats, ultrasound imaging has become a core basis for assessing the health status of organs. Physicians usually use subsequent computer vision or artificial intelligence analysis networks to identify lesions and measure morphology in the obtained ultrasound image sequences to help improve diagnostic accuracy.

[0003] However, under routine clinical examinations of the abdominal abdomens of dogs and cats, the above assumptions are highly susceptible to systematic failure at the microscopic physical level. Due to the dense and hydrophobic biophysical properties of the fur on the surface of dogs and cats, short hair roots often remain after routine hair removal. These hair roots, under the pressure of the probe, form a multi-interface composite acoustic discontinuous scatterer at the subcutaneous micro-interface with the ultrasound coupling agent. When the ultrasound beam passes through this area, it is interfered with by the local quasi-periodic arrangement of the hair follicles, resulting in accompanying, geometrically oriented, physically structured stripe noise in the imaging data.

[0004] Existing general-purpose noise removal algorithms are limited by the perceptual blind spot of structured physical noise, adhering to the underlying assumption of isotropic denoising. This leads to an unavoidable logical contradiction: when algorithms attempt to blur and suppress stripe interference through spatial smoothing or general feature convolution, their processing mechanisms cannot distinguish between the edges of such physical noise and the extremely small weak boundaries of lesions in the abdominal organs of dogs and cats within the feature space. This causes them to be incorrectly identified as the same type of target in terms of pixel features and processed accordingly. This directly results in the algorithm inevitably erasing the spatial gradient of tiny lesions that truly represent pathological features while smoothing and removing stripe noise, causing global smoothing distortion of the anatomical edges in the processed image.

[0005] In practical applications, this reduction in edge sharpness and loss of anatomical features leads to a high risk of blurred feature perception and missed detection when subsequent lesion identification models encounter small space-occupying lesions, severely limiting the effective application of digital ultrasound imaging in high-precision auxiliary diagnosis in pet clinical practice.

[0006] Therefore, existing technologies cannot effectively distinguish between structured stripe noise caused by residual hair root scattering and weak edges of small lesions in organs, thus damaging anatomical features while suppressing noise, resulting in a decrease in the accuracy of lesion-assisted identification. Summary of the Invention

[0007] To address the aforementioned technical problems, this paper provides a method and system for detecting abdominal ultrasound in pets based on image analysis. This technical solution resolves the issues raised in the background section.

[0008] In a first aspect, embodiments of this application provide a method for detecting abdominal ultrasound in pets based on image analysis, comprising the following steps: acquiring an abdominal ultrasound image of the target pet and mapping it to a complex frequency domain space through a two-dimensional fast Fourier transform to obtain a two-dimensional complex frequency domain matrix and calculating the phase value of each preset spatial frequency point therein to obtain a spatial phase spectrum; using the zero-frequency center point of the two-dimensional complex frequency domain matrix as the origin of the polar coordinate sampling, filtering out low-frequency components characterizing the overall contour of abdominal organs in the spatial phase spectrum according to a preset lower frequency limit to obtain a high-frequency band, and adjusting the deflection angle at a fixed angle within a preset deflection angle range, traversing each non-axial radial sampling ray starting from the origin of the sampling. The high-frequency band phase value sequence is extracted; the autocorrelation coefficient of the high-frequency band phase value sequence of each non-axial radial sampling ray is calculated, and the feature direction angle corresponding to the maximum value of the autocorrelation coefficient is searched among all deflection angles; a two-dimensional Gaussian filter mask with the straight line passing through the origin of the shooting point corresponding to the feature direction angle as the geometric symmetry axis and having the same dimension as the two-dimensional complex frequency domain matrix is ​​constructed; the two-dimensional complex frequency domain matrix and the two-dimensional Gaussian filter mask are multiplied element-wise, and the complex frequency domain matrix after multiplication is restored by two-dimensional inverse Fourier transform, and the edge suppression feature map is reconstructed by taking the real part; the edge suppression feature map is input into the machine learning model to obtain the abdominal auxiliary report of the target pet.

[0009] Secondly, embodiments of this application provide a pet abdominal ultrasound detection system based on image analysis, including: a phase spectrum acquisition module: used to acquire abdominal ultrasound images of the target pet and map them to a complex frequency domain space through a two-dimensional fast Fourier transform to obtain a two-dimensional complex frequency domain matrix and calculate the phase value of each preset spatial frequency point therein to obtain a spatial phase spectrum; a frequency domain map acquisition module: used to use the zero-frequency center point of the two-dimensional complex frequency domain matrix as the origin of the polar coordinate sampling point, filter out low-frequency components that characterize the overall contour of the abdominal organs in the spatial phase spectrum according to a preset lower frequency limit to obtain a high-frequency band, and within a preset deflection angle range, adjust the deflection angle at a fixed angle, traverse each non-axial radial sampling ray starting from the origin of the sampling point, and extract the high-frequency band phase value sequence; Feature direction angle processing module: used to calculate the autocorrelation coefficient of the high-frequency band phase value sequence of each non-axial radial sampling ray, and to search for the feature direction angle that makes the autocorrelation coefficient reach its maximum value among all deflection angles; Mask processing module: used to construct a two-dimensional Gaussian filter mask with the straight line passing through the origin of the shooting point corresponding to the feature direction angle as the geometric symmetry axis and having the same dimension as the two-dimensional complex frequency domain matrix; Suppression feature map processing module: used to multiply the two-dimensional complex frequency domain matrix and the two-dimensional Gaussian filter mask element by element, restore the multiplied complex frequency domain matrix through two-dimensional inverse Fourier transform, and reconstruct the edge suppression feature map by taking the real part; Report output module: used to input the edge suppression feature map into the machine learning model to obtain an auxiliary report of the target pet's abdomen.

[0010] Thirdly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described image analysis-based method for detecting abdominal ultrasound in pets.

[0011] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0012] 1. This scheme maps the abdominal ultrasound image of the target pet to a complex frequency domain space to obtain a two-dimensional complex frequency domain matrix. It then extracts the high-frequency band phase value sequence by traversing non-axial radial sampling rays within the high-frequency band and calculates its autocorrelation coefficient to retrieve the characteristic orientation angle. This step can accurately and unidirectionally lock the specific tilt angle of the structured physical grid stripe noise generated by multi-interface acoustic scattering from residual hair roots on the surface of dogs and cats from the underlying acoustic physical mechanism. This completely solves the technical problem of feature aliasing caused by the inability of traditional spatial domain denoising algorithms to perceive the specific normal direction of structured physical noise.

[0013] 2. This scheme constructs a two-dimensional Gaussian filter mask with the straight line passing through the origin of the shooting point corresponding to the characteristic direction angle as the geometric axis of symmetry, and multiplies it element-wise with a two-dimensional complex frequency domain matrix. This processing method achieves precise narrowband offsetting and reduction of phase-locked coherent energy of physical grid noise, while performing isotropic lossless passage of high-frequency signals in other directions without hair artifacts. This avoids the drawback of conventional image cleaning algorithms that simultaneously erase the discrete mean of the target anatomical boundary due to "one-size-fits-all" approach.

[0014] 3. This scheme restores the complex frequency domain matrix after multiplication to the spatial domain through a two-dimensional inverse Fourier transform, and then reconstructs the edge suppression feature map by taking the real part. This processing chain ensures that the low-frequency large anatomical contour semantics and the weak edges of isotropic small lesions in the reconstructed image are not physically truncated or damaged. This results in the final auxiliary report of the target pet's abdomen output after inputting into the machine learning model having extremely high geometric topological invariance and classification confidence, and significantly reducing the false alarm rate and false negative rate caused by the engulfment of small lesions due to hair residue. Attached Figure Description

[0015] Figure 1 A schematic diagram illustrating the steps of the image analysis-based pet abdominal ultrasound detection method provided in this application embodiment;

[0016] Figure 2 A schematic diagram of the logic flow of the image analysis-based pet abdominal ultrasound detection method provided in the embodiments of this application;

[0017] Figure 3 This is a schematic diagram of the structure of the image analysis-based pet abdominal ultrasound detection system provided in the embodiments of this application. Detailed Implementation

[0018] This application provides a pet abdominal ultrasound detection method and system based on image analysis, which solves the technical problem in the prior art where, in grayscale image processing scenarios after abdominal ultrasound examination of dogs and cats, existing noise reduction techniques cannot specifically perceive and directionally remove structured grid noise caused by residual hair root scattering, resulting in the simultaneous erasure of weak anatomical edges of small lesions sharing the high-frequency feature space during artifact smoothing, thus causing image distortion.

[0019] This solution addresses the technical challenge of physical grid stripe noise induced by multi-interface scattering from residual hair in pet abdominal ultrasound processing, which erodes and engulfs the weak edges of solid organs. It provides an exploratory data processing approach. Since the original image is a pure real matrix in the spatial domain, noise and the spatial contour information of anatomical structures are highly intertwined and overlapping. To transform the directional stripe artifacts into independent features easily separated quantitatively on the machine, this solution first uses a two-dimensional fast Fourier transform to convert the acquired original image into a complex frequency domain. By performing partial differential transformation on the frequency domain features, a two-dimensional complex frequency domain matrix containing real and imaginary parts is obtained. Furthermore, the argument of each frequency point is extracted to obtain the spatial phase spectrum. This spatial topological mapping breaks away from the traditional algorithm's limitation to smoothing based on statistical similarity in the spatial neighborhood, transforming spatial geometric stripes into a specific resonance distribution on the phase spectrum.

[0020] Considering that the isotropic, highly reflective edges of solid organs typically exhibit large-scale anatomical contours, their acoustic energy naturally concentrates in the low-frequency core region near the origin in the frequency domain, while the dense, residual physical grid of hair roots manifests as high-frequency signal components on the periphery. To directly isolate the strong interference from the large anatomical contour waveform in subsequent calculations, this scheme sets a preset lower frequency limit in the spatial phase spectrum to filter out low-frequency components reflecting the overall contour of solid organs, thereby locking in the high-frequency band containing only scattering noise and the tiny edges of the target object. To establish a unique spatial geometric reference for subsequent X-ray scanning, this scheme explicitly sets the zero-frequency center point of the two-dimensional complex frequency domain matrix after digital matrix rearrangement as the origin of the polar coordinate sampling point. Subsequently, within a preset scanning angle deflection range, the computer system controls the computer system to change the deflection direction with a fixed deflection angle step size, cyclically traversing each non-axial radial sampling ray radiating outward from the origin. Since conventional component damage or display trace artifacts are mostly distributed along the horizontal and vertical axes, the setting of non-axial radial rays can perfectly avoid occasional hardware noise. The system extracts the spatial phase along each non-axial radial sampling ray and extracts the corresponding high-frequency band phase value sequence.

[0021] Because residual hair, acting as a discontinuous scatterer, exhibits a quasi-periodic biological distribution in its local area, this physical reflection characteristic leads to severe spatial phase locking of high-frequency phase values ​​on the corresponding fringe normal slope ray. The discrete distribution of its phase collapses abruptly from a disordered state into a narrow interval. This scheme first unwraps the extracted sequence to eliminate periodic pseudo-jumps caused by arctangent calculation truncation, and then calculates the radial first-order phase difference sequence. By statistically analyzing the probability distribution of this difference sequence falling into each statistical interval and calculating the Shannon information entropy, a quantitative measure characterizing the degree of order of the spatial phase spectrum in each non-axial sampling direction is obtained, called the high-frequency non-axial phase entropy. When noise locks in this direction and phase collapse occurs, the high-frequency non-axial phase entropy will exhibit a local steep minimum valley. This scheme then performs an optimization search across all scanning deflection angles to lock the direction that allows the high-frequency non-axial phase entropy to achieve a global minimum, denoted as the characteristic direction angle. This step completes a technological leap from the original disordered data to high-precision, adaptive, and inversely locking the dominant direction of the physical noise source.

[0022] After accurately identifying the characteristic orientation angle, to achieve specific decoupling without damaging other normal high-frequency effective signals, this scheme dynamically constructs a two-dimensional Gaussian filter mask with the straight line corresponding to the characteristic orientation angle as the geometric symmetry center axis, and the matrix dimension of this mask is strictly kept the same as the original two-dimensional complex frequency domain matrix. The two-dimensional complex frequency domain matrix and the two-dimensional Gaussian filter mask are multiplied element-wise using a Hadamard product floating-point operation at the corresponding row and column coordinate positions, thereby specifically reducing and offsetting the coherent frequency energy in the noise normal direction within the frequency domain. Finally, the updated complex frequency domain matrix after multiplication is mapped back to the spatial geometric space through a two-dimensional inverse Fourier transform, and the imaginary part data is discarded, the real part matrix is ​​extracted, and a feature map with suppressed regular substantial edges is reconstructed. At this point, the weak edge gradients of organs that were originally severely obscured and masked by the physical stripe grid can be decoupled and restored without loss in the spatial feature map. Finally, the high-sharpness feature map is input into a pre-defined machine learning model for deep spatial feature extraction and fully connected layer classification using a multi-layer convolutional neural network. The final output is an auxiliary report indicating whether lesions exist in the target pet's abdominal organs and their corresponding maximum confidence probability. This solution uses the most direct frequency domain matrix algebraic cancellation and spatial domain topology reconstruction to completely overcome the industry's shortcomings of conventional noise reduction methods that erase the anatomical boundaries of minute lesions, achieving a substantial leap in edge sharpness and model generalization performance.

[0023] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0024] like Figure 1The diagram illustrates the steps of an image analysis-based ultrasound detection method for pets, as provided in this embodiment. The image analysis-based ultrasound detection method for pets includes the following steps: This embodiment provides an image analysis-based ultrasound detection method for pets. This method accurately locates and reduces structured physical grid stripe noise caused by residual hair roots on the surface of dogs and cats, preserving the weak anatomical edges of organs without damage. The following section provides a black-box analysis of the data flow steps of this method, combining the underlying data processing logic and computer execution timing.

[0025] The overall processing chain of this image analysis-based pet abdominal ultrasound detection method includes the following steps: acquiring an abdominal ultrasound image of the target pet and mapping it to the complex frequency domain using a two-dimensional fast Fourier transform (FFT) to obtain a two-dimensional complex frequency domain matrix, and calculating the phase value of each preset spatial frequency point to obtain the spatial phase spectrum. In the above steps, the abdominal ultrasound image is a spatial grayscale real matrix acquired by an ultrasound transducer and processed by beamforming. The two-dimensional fast Fourier transform is the underlying signal frequency domain mapping algorithm. The preset spatial frequency points refer to the discrete frequency domain grid coordinate points generated after mapping the spatial matrix to the complex frequency domain, and their total number is completely consistent with the resolution of the original spatial image.

[0026] Using the zero-frequency center point of the two-dimensional complex frequency domain matrix as the origin of the polar coordinate sampling point, low-frequency components representing the overall contour of abdominal organs are filtered out in the spatial phase spectrum according to a preset lower frequency limit, resulting in a high-frequency band. Within a preset deflection angle range, the deflection angle is adjusted at a fixed angle, and each non-axial radial sampling ray starting from the origin of the sampling point is traversed to extract the high-frequency band phase value sequence. In this step, the specific setting rule of the preset lower frequency limit is based on the maximum inter-class variance method to estimate the boundary of the low-frequency energy waveform of large clinical anatomical structures. The typical engineering value is between 5% and 12% of the total cutoff frequency radius outside the frequency domain center. This lower limit serves as a hard cut-off threshold to shield the large contour interference of solid organs. The preset deflection angle range defaults to cover a semi-circular polar coordinate system from 0 degrees to 180 degrees. The fixed angle is the scanning step resolution, and to balance computing power and accuracy, its typical engineering value is calibrated to 0.5 degrees. Skipping the non-axial restrictions of 0° and 90° horizontal and vertical in this step is to directly avoid damage to the ultrasonic probe chip array or the inherent orthogonal hardware dead zone noise of the display grid at the underlying timing level.

[0027] Calculate the autocorrelation coefficient of the high-frequency band phase value sequence of each non-axial radial sampling ray, and search for the characteristic direction angle that makes the autocorrelation coefficient reach its maximum value among all deflection angles;

[0028] Through quantization calculations, the system optimizes in the feature space and identifies the direction of strongest correlation that characterizes the normal projection of the physical grid noise array.

[0029] Construct a two-dimensional Gaussian filter mask with the straight line passing through the origin of the shooting point corresponding to the characteristic direction angle as the geometric symmetry axis and having the same dimension as the two-dimensional complex frequency domain matrix;

[0030] The two-dimensional complex frequency domain matrix is ​​multiplied element-wise with the two-dimensional Gaussian filter mask. The resulting complex frequency domain matrix is ​​then restored using a two-dimensional inverse Fourier transform. The real part is then taken to reconstruct the edge suppression feature map.

[0031] Element-wise multiplication achieves specific high-frequency energy blocking in the frequency domain. The two-dimensional inverse Fourier transform restores the data matrix after reducing coherent energy to the spatial coordinate system.

[0032] By inputting the edge suppression feature map into the machine learning model, an auxiliary report of the target pet's abdomen is obtained.

[0033] The machine learning model used here is a convolutional neural network based on a deep residual topology architecture (such as ResNet-50). The training dataset is derived from a clinically annotated set of pet organ pathology ultrasound images. The loss function is cross-entropy loss, and the network convergence termination condition is set to the loss descent gradient being less than a certain value for 10 consecutive iterations on the validation set. The model's input tensor is a single-channel edge suppression feature map, and the output tensor is a multi-class lesion confidence probability array mapped by a normalized exponential function (e.g., Softmax).

[0034] Figure 2 This is a schematic diagram of the logical flow of the image analysis-based ultrasound detection method for pets provided in this application embodiment. Through the above overall technical solution, this embodiment solves the overall technical problem that traditional spatial smoothing and noise reduction methods cannot distinguish between structured physical noise and weak boundaries of small lesions by constructing a high-frequency ray sampling and correlation analysis mechanism in the complex frequency domain. It achieves the unique technical effect of directional and precise stripping of residual hair root grid stripes and preservation of organ lesion anatomical features with extremely high fidelity.

[0035] Furthermore, for clinical canine and feline patients, especially those with obese bodies and extremely thick and uneven fat layers, ultrasound waves experience non-uniform acoustic attenuation when penetrating superficial tissues. Simply performing the aforementioned frequency domain mapping will cause a severe global zero-frequency phase drift at the mapped frequency domain origin due to spatial energy imbalance. This directly renders the geometric center assumption of "taking the zero-frequency center point as the origin of the radiation point" in the previous scheme invalid, leading to scanning ray deviation and failure to lock onto the noise direction.

[0036] Before acquiring abdominal ultrasound images of the target pet and mapping them to the complex frequency domain using a two-dimensional fast Fourier transform, a global zero-frequency drift origin calibration is also included:

[0037] Obtain the geometric center coordinates of the abdominal ultrasound image of the target pet in a Cartesian coordinate system, and extract the grayscale value of each pixel in the abdominal ultrasound image; traverse all pixels in the abdominal ultrasound image, using the grayscale value as the energy weight, and calculate the acoustic energy centroid coordinates of the abdominal ultrasound image using a coordinate weighted average formula, as follows:

[0038] ,in, The x-coordinate of the acoustic energy centroid coordinates The ordinate of the acoustic energy centroid coordinate system is... The x-coordinate of the pixel is The ordinate of the pixel is The grayscale value of a pixel. To perform a summation operation on all pixels in the abdominal ultrasound image; to obtain a spatial offset vector by subtracting the coordinates of the acoustic energy centroid from the coordinates of the geometric center; to perform reverse translation compensation on the pixel matrix of the abdominal ultrasound image using the spatial offset vector, resulting in a translation-compensated image where the acoustic energy centroid coincides with the geometric center; and to input the translation-compensated image into the two-dimensional fast Fourier transform to map it to the complex frequency domain space step.

[0039] In this reverse translation compensation step, the legal boundary condition for the translation displacement is: the magnitude of the spatial offset vector must not exceed 10% of the overall diagonal pixel length of the image. If this hardware calibration threshold is exceeded, the system triggers an abnormal decision interruption, determining that the probe coupling has completely failed.

[0040] This embodiment perfectly offsets the centroid deviation caused by fat layer attenuation by forcibly aligning the acoustic energy centroid with the geometric center in the spatial domain, achieving the unique effect of accurately calibrating the frequency domain center point and ensuring the absolute stability of the subsequent polar coordinate sampling coordinate system.

[0041] Furthermore, relying solely on the single linear feature direction angle detected in the preceding scheme will result in severe end-tracking off-target errors when targeting the phase trajectory of the high-curvature physiological and anatomical plane such as the edge of the pet's gastrointestinal tract, which undergoes nonlinear bending distortion along the high-curvature surface.

[0042] After obtaining the characteristic direction angle, the process also includes: establishing a polar coordinate system with the zero-frequency center point of the two-dimensional complex frequency domain matrix as the origin of the polar coordinate sampling point, and obtaining the radial coordinates of each preset spatial frequency point relative to the zero-frequency center point; dividing the high-frequency band into multiple concentric ring-shaped radial depth sub-bands along the radial direction where the characteristic direction angle is located; the number of radial depth sub-bands is dynamically allocated according to the Nyquist sampling rate, with a typical engineering value of 8 to 16 consecutive concentric ring regions.

[0043] Within each radial depth sub-band, using the characteristic orientation angle as the initial value, a local traversal is performed within a preset angle window. High-frequency band phase value sequences are extracted along each local sampling direction, and local autocorrelation coefficients are calculated. The optimal orientation angle for the sub-band that maximizes the local autocorrelation coefficient is then locked. Here, the preset angle window is forcibly limited to a central axis centered on the characteristic orientation angle, with positive and negative deflection spans restricted to a specific value. Within a very small spatial region. This fine-tuning window constraint can prevent the algorithm from getting trapped in suboptimal solutions due to irrelevant high-frequency clutter during local iterations.

[0044] Using the radial coordinates as independent variables and the optimal direction angle of each sub-band as the dependent variable, the least squares method is used to perform polynomial curve fitting to generate a dynamic bending deflection trajectory function radiating outward from the zero-frequency center point.

[0045] The order of the polynomial fitting of the least squares method is set as a preset hyperparameter and is strictly limited to no more than 3. This ensures the smoothness of the biological surface fitting while preventing curve calculation oscillations caused by high-order Runge phenomena.

[0046] This embodiment solves the problem that linear models cannot match the scattering deformation of tissue surfaces by using loop optimization and polynomial curvature regression, and achieves a progressive technical effect of dynamically adaptively fitting the noise distribution of physiological curved interfaces.

[0047] Furthermore, since the axis of symmetry has been upgraded to a dynamic bending deflection trajectory function in the previous steps, the traditional Euclidean geometric point-to-line vertical distance calculation model completely fails in the mask weighting scenario, directly causing leakage in the Gaussian blocking band at the bending end.

[0048] The specific process of obtaining the two-dimensional Gaussian filter mask is as follows: traverse each preset spatial frequency point in the two-dimensional complex frequency domain matrix, and calculate the polar coordinates of the preset spatial frequency point based on the zero-frequency center point, extracting the radial and angular coordinates; substitute the radial coordinates of the preset spatial frequency point into the dynamic bending deflection trajectory function to calculate the trajectory reference angle under the corresponding radius; calculate the phase angle deviation of the preset spatial frequency point, and multiply the angular deviation by the radial coordinate to obtain the equivalent orthogonal arc length distance from the preset spatial frequency point to the dynamic bending deflection trajectory function; the mathematical relationship for calculating the equivalent orthogonal arc length distance here is perfectly replaced by the computationally expensive numerical integration iterative algorithm for finding the shortest orthogonal distance through the local polar coordinate infinitesimal arc length approximation.

[0049] Substituting the equivalent orthogonal arc length into the preset one-dimensional Gaussian attenuation model, the mask weight value corresponding to the preset spatial frequency point is calculated; the supplementary formula for the one-dimensional Gaussian attenuation model is: ,in, To calculate the obtained mask weight values, For the equivalent orthogonal arc length distance, The preset mask bandwidth hyperparameter is typically 2% of the radius of the spectral matrix, based on engineering experience.

[0050] According to the row and column coordinates of the preset spatial frequency points in the two-dimensional complex frequency domain matrix, the corresponding mask weight values ​​are assigned to the corresponding coordinate positions of the preset zero matrix container, thereby constructing a two-dimensional Gaussian filter mask with the same dimension as the two-dimensional complex frequency domain matrix and with the straight line passing through the origin of the shooting point corresponding to the characteristic direction angle as the geometric symmetry axis.

[0051] This embodiment solves the problem of accurately characterizing the energy stopband near the curved frequency domain trajectory by using an orthogonal arc length mapping Gaussian weighting mechanism, and achieves the superposition effect of extremely precise nonlinear surface adaptive filtering and strict prevention of the false attenuation of signals from small lesions.

[0052] Furthermore, this relates to the computational scenario of extracting polar coordinate phase from the real and imaginary parts in the frequency domain. Traditional single-valued arctangent functions suffer from their domain being truncated. When crossing the boundaries of the second and third quadrants, it can lead to severe phase sign confusion and periodic pseudo-jumps.

[0053] The specific steps for obtaining the spatial phase spectrum include: extracting the real and imaginary data corresponding to each preset spatial frequency point in the two-dimensional complex frequency domain matrix; for each preset spatial frequency point, calling the four-quadrant arctangent function, using the imaginary and real data as input parameters, to calculate the phase value corresponding to that preset spatial frequency point. The calculation formula for the four-quadrant arctangent function is as follows:

[0054] ,in, The phase value at the preset spatial frequency point, It is the arctangent function in the fourth quadrant. For imaginary part data, Real data;

[0055] The phase values ​​calculated from all preset spatial frequency points are arranged in an array according to their positions in a two-dimensional complex frequency domain matrix to generate a spatial phase spectrum.

[0056] This embodiment solves the problem of quadrant assignment ambiguity for complex plane vectors by hard-coding the hardware-level four-quadrant arctangent function, and realizes phase holographic smooth and continuous analysis of discrete signals in the frequency domain.

[0057] Furthermore, in scenarios involving the calculation of discrete statistical regularities in phase sequences, direct accumulation is highly susceptible to the offset of the overall phase baseline, failing to highlight local capillary grid correlations. Additionally, for spatial sequence span processing, there is a problem of "cold start" at the array ends where no data is available.

[0058] To eliminate baseline interference and remove spatial boundary overrun errors, the specific steps for calculating the autocorrelation coefficient of the high-frequency band phase value sequence for each non-axial radial sampling ray are as follows: For each non-axial radial sampling ray, calculate the sequence mean of its corresponding high-frequency band phase value sequence; based on the sequence mean and a preset delay span, calculate the autocorrelation coefficient using the discrete autocorrelation function formula, which is as follows:

[0059] ,in, Represents the autocorrelation coefficient. This represents the phase value at the current sampling point in the high-frequency band phase value sequence. This refers to the phase value at a sampling point in the high-frequency band phase value sequence after a preset delay. The mean of the sequence. This is the preset delay span.

[0060] The preset delay span here The prior value of the frequency domain period corresponding to the physical spatial distance of the acoustic multi-interface caused by residual hair roots needs to be strictly matched, and is usually calibrated to 5% to 10% of the total sequence length. This addresses the "cold start / out-of-bounds" phenomenon at the ends of the spatial sequence, i.e., when the spatial index... When the current ray exceeds its physical boundary, the system forces a zero-fill truncation mechanism, calculating only the overlapping data that satisfies the valid memory index field.

[0061] This embodiment removes baseline effects and integrates spatial delay definition, solving the problem of defocusing due to interference from strong reflection blind zones, and achieving the technical effect of extremely high noise resistance and strict quantitative correlation of hidden physical periodic fringes.

[0062] Furthermore, in frequency domain suppression scenarios involving high-frequency energy operations, using traditional two-dimensional matrix multiplicative convolution will lead to memory and computing power disasters and unwarranted crosstalk across the entire frequency band.

[0063] To facilitate independent, fixed-point, low-time reduction operations on high-dimensional matrices, the element-wise multiplication of a two-dimensional complex frequency domain matrix with a two-dimensional Gaussian filter mask involves: decomposing the two-dimensional complex frequency domain matrix into an initial real part matrix and an initial imaginary part matrix; performing Hadamard product floating-point multiplication on the mask weight scalar values ​​at each coordinate position in the two-dimensional Gaussian filter mask with the corresponding values ​​at the same coordinate positions of the initial real part matrix and the initial imaginary part matrix, respectively, to obtain updated real part matrices and updated imaginary part matrices; and finally, recombining the updated real part matrix and the updated imaginary part matrix into a multiplied complex frequency domain matrix.

[0064] The Hadamard product is a point-to-point floating-point scaling operation directly supported by the underlying parallel registers, ensuring that the energy of each locked discrete frequency point is only subject to its own weighting factor.

[0065] This embodiment solves the technical problems of high memory consumption and cross-response interference, and achieves high-frequency energy nanosecond-level point-to-point lossless isolation in a specific physical direction.

[0066] Furthermore, in the scenario of reconstructing the semantic spatial output from the inverse transformation of the complex domain, due to the numerical precision truncation limitation of the inverse Fourier transformation calculation, the spatial matrix will inevitably be accompanied by residual small complex waste, namely the imaginary part signal, which seriously interferes with the gradient propagation of subsequent network neurons.

[0067] To eliminate the pollution of the spatial domain signal-to-noise ratio by residual imaginary clutter in the frequency domain, the steps to obtain the edge suppression feature map specifically include: taking the multiplied complex frequency domain matrix as input, calling the two-dimensional inverse fast Fourier transform algorithm to perform a mapping operation from the frequency domain to the spatial domain, obtaining a complex spatial domain image matrix containing real and imaginary matrices; extracting the real matrix from the complex spatial domain image matrix; removing the imaginary matrix from the complex spatial domain image matrix, and using the retained real matrix as the output edge suppression feature map.

[0068] The true grayscale ultrasonic reflection energy exists absolutely and completely in the pure real number domain. All accompanying imaginary matrices are merely mathematical remnants of the algorithm caused by phase suppression asymmetry.

[0069] This embodiment forcibly blocks and discards the spatial imaginary part matrix, solving the problem of high-frequency residual waste in the imaginary domain contaminating the confidence of feature extraction in deep networks, and achieving the technical effect of continuously inputting clutter-free, high-fidelity ultrasonic feature maps into the auxiliary reporting model.

[0070] Figure 3This is a schematic diagram of the pet abdominal ultrasound detection system based on image analysis provided in this application embodiment. The pet abdominal ultrasound detection system based on image analysis includes: a phase spectrum acquisition module: used to acquire abdominal ultrasound images of the target pet and map them to a complex frequency domain space through a two-dimensional fast Fourier transform to obtain a two-dimensional complex frequency domain matrix and calculate the phase value of each preset spatial frequency point to obtain a spatial phase spectrum; a frequency domain map acquisition module: used to filter out low-frequency components that characterize the overall contour of abdominal organs in the spatial phase spectrum according to a preset lower frequency limit, using the zero-frequency center point of the two-dimensional complex frequency domain matrix as the origin of the polar coordinate sampling point, to obtain a high-frequency band, and within a preset deflection angle range, to adjust the deflection angle at a fixed angle, traverse each non-axial radial sampling ray starting from the origin of the sampling point, and extract... The system includes: a high-frequency phase value sequence; a feature orientation angle processing module (used to calculate the autocorrelation coefficient of the high-frequency phase value sequence of each non-axial radial sampling ray and to retrieve the feature orientation angle that maximizes the autocorrelation coefficient among all deflection angles); a mask processing module (used to construct a two-dimensional Gaussian filter mask with the straight line passing through the origin of the shooting point corresponding to the feature orientation angle as the geometric symmetry axis and having the same dimension as the two-dimensional complex frequency domain matrix); a suppression feature map processing module (used to multiply the two-dimensional complex frequency domain matrix and the two-dimensional Gaussian filter mask element-wise, restore the multiplied complex frequency domain matrix through a two-dimensional inverse Fourier transform, and reconstruct the edge suppression feature map by taking the real part); and a report output module (used to input the edge suppression feature map into the machine learning model to obtain an auxiliary report of the target pet's abdomen).

[0071] This application also provides a computer-readable storage medium for storing a computer program, which, when executed by a processor, implements an image analysis-based method for detecting abdominal ultrasound in pets.

[0072] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0073] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0074] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0075] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0076] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0077] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for detecting abdominal ultrasound in pets based on image analysis, characterized in that, Includes the following steps: The abdominal ultrasound image of the target pet is acquired and mapped to the complex frequency domain space through a two-dimensional fast Fourier transform to obtain a two-dimensional complex frequency domain matrix. The phase value of each preset spatial frequency point is calculated to obtain the spatial phase spectrum. Using the zero-frequency center point of the two-dimensional complex frequency domain matrix as the origin of the polar coordinate sampling, the low-frequency components that characterize the overall contour of the abdominal organs are filtered out in the spatial phase spectrum according to the preset lower frequency limit to obtain the high-frequency band. Within the preset deflection angle range, the deflection angle is adjusted at a fixed angle, and each non-axial radial sampling ray starting from the origin of the sampling point is traversed to extract the high-frequency band phase value sequence. Calculate the autocorrelation coefficient of the high-frequency band phase value sequence of each non-axial radial sampling ray, and search for the characteristic direction angle that makes the autocorrelation coefficient reach its maximum value among all deflection angles; Construct a two-dimensional Gaussian filter mask with the straight line passing through the origin of the shooting point corresponding to the characteristic direction angle as the geometric symmetry axis and having the same dimension as the two-dimensional complex frequency domain matrix; The two-dimensional complex frequency domain matrix is ​​multiplied element-wise with the two-dimensional Gaussian filter mask. The resulting complex frequency domain matrix is ​​then restored using a two-dimensional inverse Fourier transform. The real part is then taken to reconstruct the edge suppression feature map. By inputting the edge suppression feature map into the machine learning model, an auxiliary report of the target pet's abdomen is obtained.

2. The image analysis-based ultrasound detection method for pet abdomens according to claim 1, characterized in that, Before acquiring abdominal ultrasound images of the target pet and mapping them to the complex frequency domain using a two-dimensional fast Fourier transform, a global zero-frequency drift origin calibration is also included: Obtain the geometric center coordinates of the abdominal ultrasound image of the target pet in a Cartesian coordinate system, and extract the grayscale value of each pixel in the abdominal ultrasound image; By iterating through all pixels in the abdominal ultrasound image and using grayscale values ​​as energy weights, the acoustic energy centroid coordinates of the abdominal ultrasound image are calculated using a coordinate-weighted average formula, as follows: , ,in, The x-coordinate of the acoustic energy centroid coordinates The ordinate of the acoustic energy centroid coordinate system is... The x-coordinate of the pixel is The ordinate of the pixel is The grayscale value of a pixel. To perform a summation operation on all pixels within an abdominal ultrasound image; The spatial offset vector is obtained by subtracting the coordinates of the acoustic energy centroid from the coordinates of the geometric center. The pixel matrix of the abdominal ultrasound image is reverse-translated using a spatial offset vector to obtain a translation-compensated image in which the acoustic energy centroid coincides with the geometric center. The translation-compensated image is then input into a two-dimensional fast Fourier transform to map it to the complex frequency domain.

3. The image analysis-based ultrasound detection method for pet abdomens according to claim 2, characterized in that, After obtaining the characteristic orientation angle, it also includes: Using the zero-frequency center point of the two-dimensional complex frequency domain matrix as the origin of the polar coordinate sampling point, a polar coordinate system is established to obtain the radial coordinates of each preset spatial frequency point relative to the zero-frequency center point. Along the radial direction where the characteristic direction angle is located, the high-frequency band is divided into multiple radial depth sub-bands in a concentric ring shape; Within each radial depth sub-band, using the characteristic orientation angle as the initial value, local traversal is performed within a preset angle window to extract the high-frequency band phase value sequence in each local sampling direction and calculate the local autocorrelation coefficient. The optimal orientation angle of the sub-band that makes the local autocorrelation coefficient of the current sub-band reach its local maximum value is then locked. Using the radial coordinates as independent variables and the optimal direction angle of each sub-band as the dependent variable, the least squares method is used to perform polynomial curve fitting to generate a dynamic bending deflection trajectory function radiating outward from the zero-frequency center point.

4. The image analysis-based ultrasound detection method for pet abdomens according to claim 3, characterized in that, The specific process for obtaining the two-dimensional Gaussian filter mask is as follows: Traverse each preset spatial frequency point in the two-dimensional complex frequency domain matrix, calculate the polar coordinates of the preset spatial frequency point based on the zero-frequency center point, and extract the radial and angular coordinates; Substitute the radial coordinates of the preset spatial frequency point into the dynamic bending deflection trajectory function to calculate the trajectory reference angle at the corresponding radius. Calculate the phase angle deviation at the preset spatial frequency point, and multiply the phase angle deviation by the radial coordinate to obtain the equivalent orthogonal arc length distance from the preset spatial frequency point to the dynamic bending deflection trajectory function; Substitute the equivalent orthogonal arc length into the preset one-dimensional Gaussian attenuation model to calculate the mask weight value corresponding to the preset spatial frequency point. According to the row and column coordinates of the preset spatial frequency points in the two-dimensional complex frequency domain matrix, the corresponding mask weight values ​​are assigned to the corresponding coordinate positions of the preset zero matrix container, thereby constructing a two-dimensional Gaussian filter mask with the same dimension as the two-dimensional complex frequency domain matrix and with the straight line passing through the origin of the shooting point corresponding to the characteristic direction angle as the geometric symmetry axis.

5. The image analysis-based ultrasound detection method for pet abdomens according to claim 1, characterized in that, The specific steps for obtaining the spatial phase spectrum include: Extract the real and imaginary part data corresponding to each preset spatial frequency point in the two-dimensional complex frequency domain matrix; For each preset spatial frequency point, the four-quadrant arctangent function is invoked, using the imaginary and real parts as input parameters, to calculate the phase value corresponding to that preset spatial frequency point. The calculation formula for the four-quadrant arctangent function is as follows: ,in, The phase value at the preset spatial frequency point, It is the arctangent function in the fourth quadrant. For imaginary part data, Real data; The phase values ​​calculated from all preset spatial frequency points are arranged in an array according to their positions in a two-dimensional complex frequency domain matrix to generate a spatial phase spectrum.

6. The image analysis-based ultrasound detection method for pet abdomens according to claim 1, characterized in that, The specific steps for calculating the autocorrelation coefficient of the high-frequency band phase value sequence of each non-axial radial sampling ray are as follows: For each non-axial radial sampling ray, calculate the sequence mean of its corresponding high-frequency band phase value sequence; Based on the sequence mean and a preset delay span, the autocorrelation coefficient is calculated using the discrete autocorrelation function formula, which is as follows: ,in, Represents the autocorrelation coefficient. This represents the phase value at the current sampling point in the high-frequency band phase value sequence. This refers to the phase value at a sampling point in the high-frequency band phase value sequence after a preset delay. The mean of the sequence. This is the preset delay span.

7. The image analysis-based ultrasound detection method for pet abdomens according to claim 1, characterized in that, The steps for element-wise multiplication of a two-dimensional complex frequency domain matrix with a two-dimensional Gaussian filter mask include: Decompose the two-dimensional complex frequency domain matrix into an initial real part matrix and an initial imaginary part matrix; The mask weight scalar values ​​at each coordinate position in the two-dimensional Gaussian filter mask are subjected to Hadamard product floating-point multiplication with the values ​​at the corresponding coordinate positions of the initial real part matrix and the initial imaginary part matrix, respectively, to obtain the updated real part matrix and the updated imaginary part matrix. The updated real part matrix and the updated imaginary part matrix are recombined into a complex frequency domain matrix after multiplication.

8. The image analysis-based ultrasound detection method for pet abdomens according to claim 1, characterized in that, The specific steps to obtain the edge suppression feature map include: The complex frequency domain matrix after multiplication is used as input, and the two-dimensional inverse fast Fourier transform algorithm is called to perform the mapping operation from the frequency domain to the spatial domain, so as to obtain a complex spatial domain image matrix containing the real part matrix and the imaginary part matrix. Extract the real part of the complex spatial domain image matrix; The imaginary part of the complex spatial domain image matrix is ​​removed, and the remaining real part is used as the output edge suppression feature map.

9. A pet abdominal ultrasound detection system based on image analysis, characterized in that, include: Phase spectrum acquisition module: used to acquire the abdominal ultrasound image of the target pet and map it to the complex frequency domain space through two-dimensional fast Fourier transform to obtain a two-dimensional complex frequency domain matrix and calculate the phase value of each preset spatial frequency point to obtain the spatial phase spectrum; Frequency domain map acquisition module: It is used to filter out low-frequency components that characterize the overall contour of abdominal organs in the spatial phase spectrum according to the preset lower frequency limit, using the zero-frequency center point of the two-dimensional complex frequency domain matrix as the origin of the polar coordinate sampling point, to obtain the high-frequency band. Within the preset deflection angle range, the deflection angle is adjusted at a fixed angle, and each non-axial radial sampling ray starting from the origin of the sampling point is traversed to extract the high-frequency band phase value sequence. Feature orientation angle processing module: used to calculate the autocorrelation coefficient of the high-frequency band phase value sequence of each non-axial radial sampling ray, and to search for the feature orientation angle that makes the autocorrelation coefficient reach its maximum value among all deflection angles; Mask processing module: used to construct a two-dimensional Gaussian filter mask with the straight line passing through the origin of the shooting point corresponding to the feature direction angle as the geometric symmetry axis and having the same dimension as the two-dimensional complex frequency domain matrix; The suppression feature map processing module is used to multiply the two-dimensional complex frequency domain matrix with the two-dimensional Gaussian filter mask element by element, restore the multiplied complex frequency domain matrix through two-dimensional inverse Fourier transform, and reconstruct the edge suppression feature map by taking the real part. Report output module: Used to input the edge suppression feature map into the machine learning model to obtain an auxiliary report of the target pet's abdomen.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-8.

Images