Computer-aided diagnosis method and system based on medical images

Abstract

This application belongs to the field of medical image processing technology, and provides a computer-aided diagnostic method and system based on medical images, addressing the shortcomings of computer-aided diagnosis. The method includes: acquiring diffusion tensor images of the brain of the target subject and surface electromyography (EMG) signals of associated muscle groups; converting the images into Riemannian manifold data through tensor decomposition and symmetric positive definite matrix mapping; extracting the Hurst exponent of the EMG signals based on rescaled range analysis to generate motion feature vectors; projecting the manifold data and motion features onto the correlation space using Riemann logarithmic mapping and canonical correlation analysis, extracting the maximum correlation component to generate coupled feature vectors; determining the reconstructed site through Riemann exponent mapping, calculating the geodesic length between it and the baseline state point to obtain the deviation value; and quantitatively determining the degree of neural remodeling and motor function level of stroke patients based on the deviation value. This application achieves accurate quantitative assessment of neural remodeling and motor recovery status in stroke patients.

Description

Technical Field

[0001] This application belongs to the field of medical image processing technology, and in particular relates to a computer-aided diagnostic method and system based on medical images. Background Technology

[0002] Computer-aided diagnostic methods based on medical imaging have broad application prospects in the field of stroke rehabilitation, enabling doctors to objectively assess patients' neural remodeling status and rehabilitation progress. With the development of brain science and data analysis technology, establishing a quantitative correlation between brain structural changes and limb function recovery using multimodal data has become an important research direction for achieving precise stroke rehabilitation assessment.

[0003] Existing auxiliary diagnostic methods typically utilize diffusion tensor imaging to analyze the structural integrity of brain white matter fiber tracts, or analyze the contractile characteristics of limb muscles through surface electromyography (EMG) signals. Some existing techniques attempt to perform simple linear splicing or statistical regression analysis of brain imaging features and EMG features in Euclidean space, trying to comprehensively assess the patient's motor function status.

[0004] However, due to the complex manifold geometry of brain tensor data and the strong nonlinear dynamics of electromyography (EMG) signals, there are significant differences between the two in terms of physical dimensions and spatial distribution. Simple linear combination is insufficient to effectively eliminate the differences between heterogeneous data and capture the deep neuromuscular coupling relationship. This processing method is prone to distortion or loss of feature information, making it impossible to accurately measure the true difference between the patient's current state and the health baseline. Therefore, existing technologies suffer from the technical problem of failing to accurately establish the deep coupling relationship between brain structure and motor function due to improper heterogeneous data fusion processing, thus leading to insufficient computer-aided diagnosis. Summary of the Invention

[0005] The purpose of this application is to provide a computer-aided diagnostic method and system based on medical images, so as to solve the problem of the shortcomings of existing computer-aided diagnostic technology.

[0006] To address the aforementioned technical problems, in a first aspect, this application provides a computer-aided diagnostic method based on medical imaging, applicable to the assessment of neural remodeling in stroke patients, comprising:

[0007] Acquire diffusion tensor images of the target subject's brain and surface electromyographic signals of muscle groups innervated by nerve fiber bundles associated with the diffusion tensor images during rehabilitation movements.

[0008] Tensor solving is performed on the diffusion tensor image to determine the diffusion value corresponding to each voxel. The diffusion parameters are generated by calculating the anisotropy fraction and average diffusion rate based on the diffusion values. The diffusion parameters are then transformed into a point set of the Riemannian manifold through symmetric positive definite matrix mapping to obtain the diffusion manifold data.

[0009] The motion feature vector is obtained by calculating the mean recalculated range of surface electromyography signals at different observation lengths and performing double log-linear fitting to solve the Hurst exponent.

[0010] The diffuse manifold data is mapped to the corresponding tangent space through the Riemann logarithmic mapping to generate tangent vectors. The projection transformation matrix is ​​calculated by performing canonical correlation analysis on the tangent vectors and motion feature vectors. After mapping the tangent vectors and motion feature vectors to the correlation space through the projection transformation matrix, the maximum correlation component is extracted to generate coupled feature vectors.

[0011] The coupled feature vectors are mapped to the Riemann manifold space using the Riemann exponent mapping to determine the target reconstruction site. The geodesic length between the target reconstruction site and the reference state point in the preset reference manifold is calculated based on the Riemann metric tensor to obtain the deviation value.

[0012] The target division interval is determined based on the deviation value and the preset division interval, and the diagnostic result is determined based on the motor function level corresponding to the target division interval.

[0013] Optionally, tensor solving is performed on the diffusion tensor image to determine the diffusion value corresponding to each voxel, and diffusion parameters are generated by calculating the anisotropy fraction and average diffusion rate based on the diffusion values. The diffusion parameters are then transformed into a point set of a Riemannian manifold through a symmetric positive definite matrix mapping to obtain diffusion manifold data, including:

[0014] Tensor diagonalization is performed on the diffusion weighted signals of the diffusion tensor image in multiple diffusion gradient directions to determine the mutually orthogonal diffusion values ​​corresponding to each voxel.

[0015] The diffusion anisotropy deviation and average diffusion intensity of the nerve fiber bundle are calculated using diffusion values. The diffusion anisotropy deviation is determined as the anisotropy fraction, and the average diffusion intensity is determined as the average diffusion rate. The anisotropy fraction and the average diffusion rate are then combined using eigenvectors to generate diffusion parameters.

[0016] The three-order symmetric positive definite matrix corresponding to each voxel is constructed using the diffusion parameter. The three-order symmetric positive definite matrix is ​​used as a mapping operator to project the diffusion parameter onto the Riemann geometric space to generate the corresponding spatial coordinate points, thus obtaining the point set of the Riemann manifold.

[0017] Diffuse manifold data is obtained by aggregating the spatial topology of the point set of the Riemannian manifold.

[0018] Optionally, the motion feature vector is obtained by calculating the mean rescaled range of surface electromyography signals at different observation lengths and performing a double log-linear fitting to solve for the Hurst exponent, including:

[0019] The surface electromyography signal is divided into corresponding signal sub-intervals according to multiple preset observation lengths. The cumulative deviation sequence and standard deviation of the sampling points in each signal sub-interval are calculated. The range value is determined according to the deviation distance between the maximum and minimum values ​​in the cumulative deviation sequence. The recalibrated range component corresponding to each signal sub-interval is determined according to the ratio of the range value to the standard deviation.

[0020] Statistical averaging is performed on multiple rescaled range components belonging to the same observation length to obtain the mean of the rescaled range. A logarithmic mapping relationship between the mean of the rescaled range and the corresponding observation length is constructed. The evolution slope of the regression line in the logarithmic mapping relationship is determined by linear regression fitting, and the evolution slope is determined as the Hearst exponent.

[0021] Multiple Hearst indices collected from different muscle groups are vectorized and arranged in a preset order to generate motion feature vectors.

[0022] Optionally, the diffuse manifold data is mapped to the corresponding tangent space using a Riemannian logarithmic mapping to generate a tangent vector. Canonical correlation analysis is then performed on the tangent vector and the motion feature vector to calculate the projection transformation matrix, including:

[0023] Select the reference center point of the manifold space where the diffuse manifold data is located, and use the Riemann logarithmic transform to map the diffuse manifold data to the tangent plane space corresponding to the reference center point to obtain the tangent vector;

[0024] The first covariance matrix is ​​obtained by performing autocorrelation distribution statistics on the tangent vector using the regularization term, the second covariance matrix is ​​obtained by performing autocorrelation distribution statistics on the motion feature vector, and the cross-covariance matrix is ​​obtained by performing cross-correlation statistical operations on the tangent vector and the motion feature vector.

[0025] The correlation response function is constructed using the first covariance matrix, the second covariance matrix, and the cross-covariance matrix. The correlation response function is then subjected to generalized eigenvalue decomposition to calculate the linear transformation coefficients that maximize the correlation between the mapping components of the tangent vector and the motion feature vector, thereby generating the projection transformation matrix.

[0026] Optionally, after mapping the tangent vector and motion feature vector to the association space using a projection transformation matrix, the maximum correlation component is extracted to generate a coupled feature vector, including:

[0027] By performing linear coordinate transformation on the submatrices corresponding to the tangent vector and the motion feature vector in the projection transformation matrix, the first projection vector group and the second projection vector group located in the associated space are obtained.

[0028] Perform pairwise linear correlation operations on the components of each dimension in the first and second projection vector groups, and extract the pair of components with the highest linear correlation coefficient as the maximum correlation feature pair;

[0029] Weighting coefficients are assigned based on the correlation strength of the most correlated component. The weighting coefficients are then used to perform a linear summation operation on each modal feature component in the most correlated feature pair to generate a coupled feature vector.

[0030] Optionally, the coupled feature vector is mapped to the Riemannian manifold space using the Riemannian exponent mapping to determine the target reconstruction site. The geodesic length between the target reconstruction site and the reference state point in the preset reference manifold is calculated based on the Riemannian metric tensor to obtain the deviation value, including:

[0031] Using the reference center point as the origin of the mapping, an exponential coordinate transformation is performed on the coupled feature vector to map the coupled feature vector from the linear tangent space back to the surface coordinate system where the Riemann manifold space is located, thereby determining the target reconstruction point;

[0032] Based on the coordinate difference between the target reconstruction point and the reference state point in the surface coordinate system, the path is discretized to obtain the first path step sequence connecting the target reconstruction point and the reference state point.

[0033] Using the Riemann metric tensor of the location of each discrete step length in the first path step length sequence as a local geometric scaling operator, the coordinate displacement component corresponding to each discrete step length is subjected to a weighted inner product operation with the Riemann metric tensor to calculate the local geometric length of each discrete step length in the Riemann manifold space, thus obtaining the second path step length sequence composed of each local geometric length.

[0034] The second path step sequence is processed by numerical cumulative integration along the Riemannian manifold space surface to solve the minimum trajectory path length between the target reconstruction point and the reference state point, and the minimum trajectory path length is determined as the deviation value.

[0035] Optionally, the muscle group is located in the innervation area corresponding to the target nerve fiber bundle in the diffusion tensor image, and the acquisition site of the surface electromyography signal and the target nerve fiber bundle are on the same motor nerve conduction path in human anatomy.

[0036] Secondly, this application provides a computer-aided diagnostic system based on medical images, comprising:

[0037] The acquisition module is used to acquire diffusion tensor images of the target subject's brain and surface electromyographic signals of muscle groups innervated by nerve fiber bundles associated with the diffusion tensor images during rehabilitation movements.

[0038] The calculation module is used to perform tensor solving on the diffusion tensor image to determine the diffusion value corresponding to each voxel, and generate diffusion parameters by calculating the anisotropy fraction and average diffusion rate based on the diffusion value. The diffusion parameters are then transformed into a point set of the Riemannian manifold through symmetric positive definite matrix mapping to obtain diffusion manifold data.

[0039] The generation module is used to calculate the mean of the recalculated range of surface electromyography signals at different observation lengths and perform double log-linear fitting to solve the Hurst exponent, thereby obtaining the motion feature vector.

[0040] The generation module is also used to map diffuse manifold data to the corresponding tangent space through Riemann logarithmic mapping to generate tangent vectors, perform canonical correlation analysis on tangent vectors and motion feature vectors to calculate the projection transformation matrix, and after mapping tangent vectors and motion feature vectors to the association space through the projection transformation matrix, extract the maximum correlation component to generate coupled feature vectors.

[0041] The generation module is also used to map the coupled feature vector to the Riemann manifold space using the Riemann exponent mapping to determine the target reconstruction site, and to calculate the geodesic length between the target reconstruction site and the reference state point in the preset reference manifold based on the Riemann metric tensor to obtain the deviation value.

[0042] The determination module is used to determine the target division interval based on the deviation value and the preset division interval, and to determine the diagnostic result based on the motor function level corresponding to the target division interval.

[0043] Thirdly, this application provides an electronic device, comprising:

[0044] Memory, used to store computer programs;

[0045] A processor for executing the computer program to implement the steps of the computer-aided diagnostic method based on medical images as described in the first aspect above.

[0046] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of the computer-aided diagnosis method based on medical images as described in the first aspect above.

[0047] The beneficial effects of this application are:

[0048] The computer-aided diagnostic method based on medical images provided in this application transforms diffusion tensor images into Riemannian manifold data using symmetric positive definite matrix mapping and incorporates the Hurst exponent, which reflects the nonlinear characteristics of electromyography, effectively preserving the spatial geometric properties of brain microstructures and the dynamic characteristics of limb movement. Furthermore, by utilizing Riemannian tangent space co-projection and canonical correlation analysis techniques, the method eliminates the differences in physical dimensions and distribution inconsistencies between heterogeneous data within a unified linear tangent space. This allows for the extraction of coupling feature vectors that maximize the reflection of neuromuscular connections, avoiding feature distortions caused by traditional Euclidean linear combination processing.

[0049] Meanwhile, by using the geodesic distance calculated based on the Riemannian metric tensor as a criterion, the degree to which the patient's current state deviates from the healthy baseline can be accurately measured strictly along the intrinsic geometric path of the manifold surface. Therefore, this application can effectively solve the technical problem that improper heterogeneous data fusion processing methods make it difficult to accurately establish the deep coupling relationship between brain structure and motor function, thus leading to insufficient computer-aided diagnosis, and significantly improve the accuracy and objectivity of neural remodeling assessment and motor function diagnosis in stroke patients.

[0050] Furthermore, this application, through tensor diagonalization and feature parameter extraction of the diffusion-weighted signal, can accurately quantify the anisotropic bias and average intensity of nerve fiber bundles from multiple gradient directions, thereby capturing subtle pathological changes in the microstructure of white matter after stroke. By constructing a third-order symmetric positive definite matrix and mapping these physical parameters to Riemannian geometric space, and utilizing manifold geometric properties instead of the traditional Euclidean distance metric, the inherent spatial topology and nonlinear properties of the tensor data are effectively maintained.

[0051] Finally, diffuse manifold data is generated through aggregation using spatial topology. Therefore, this application can significantly improve the accuracy of brain structural feature representation and solve the technical problem of inaccurate image feature extraction caused by neglecting tensor manifold structure in existing technologies. Attached Figure Description

[0052] To more clearly illustrate the technical solutions of the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 A flowchart illustrating a computer-aided diagnostic method based on medical images, provided for an embodiment of this application;

[0054] Figure 2 A flowchart illustrating a method for generating diffuse manifold data, provided in an embodiment of this application;

[0055] Figure 3 A flowchart illustrating a method for determining diagnostic results provided in an embodiment of this application;

[0056] Figure 4 A schematic diagram of the structure of a computer-aided diagnostic system based on medical images provided in an embodiment of this application;

[0057] Figure 5 This is a schematic diagram of the hardware structure of an electronic device provided in one embodiment of this application. Detailed Implementation

[0058] In the field of stroke rehabilitation assessment, while existing computer-aided diagnostic technologies can acquire imaging information of brain structures and electrophysiological signals of limb muscles separately, they still face significant challenges in data fusion and analysis. Conventional methods often ignore the inherent geometric and dynamic differences in the data, attempting to perform simple linear splicing or statistical regression on diffuse tensor data with complex manifold structures and electromyographic signals with strong nonlinear characteristics in Euclidean space.

[0059] This rigid combination method cannot effectively eliminate the physical dimensional barriers between heterogeneous data, causing the deep neuromuscular coupling relationship to be distorted or lost during feature extraction. As a result, it is difficult to accurately measure the real difference between the patient's current state and the health benchmark, resulting in insufficient objectivity and accuracy of the assessment results.

[0060] To address the challenges of heterogeneous data fusion, this application proposes a computer-aided diagnostic method based on medical imaging. It establishes a unified analytical framework between brain structure and limb function using Riemannian geometry, a high-dimensional mathematical tool. Specifically, the method first transforms the brain diffusion tensor into a point set on a Riemannian manifold and extracts the Hearst exponent, reflecting nonlinear muscle dynamics. Then, using Riemannian tangent space mapping and canonical correlation analysis, it constructs neuromuscular coupling features that maximize correlation within a tangent space that eliminates dimensional differences. Finally, by calculating the geodesic distance of the feature vectors on the manifold surface relative to a healthy baseline, it achieves precise quantification of the degree of deviation.

[0061] This application abandons the traditional linear combination model of Euclidean space and utilizes the intrinsic metric properties of manifold geometry to ensure deep fusion of multimodal data while preserving their respective topological structures. It retains the geometric properties of brain microstructures while also taking into account the dynamic characteristics of motor functions, solving the feature distortion problem caused by improper fusion of heterogeneous data in existing technologies, and significantly improving the accuracy of stroke neural remodeling assessment.

[0062] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0063] To address the problems of the prior art, embodiments of this application provide a computer-aided diagnostic method, apparatus, device, computer storage medium, and computer program product based on medical images. The computer-aided diagnostic method based on medical images provided in this application embodiment will be described first below.

[0064] Figure 1 This illustration shows a flowchart of a computer-aided diagnostic method based on medical images according to an embodiment of this application. Figure 1 As shown, the method includes:

[0065] S101. Obtain diffusion tensor images of the target subject's brain and surface electromyographic signals of muscle groups innervated by nerve fiber bundles associated with the diffusion tensor images during rehabilitation movements.

[0066] Optionally, in step S101, the muscle group is located in the innervation region corresponding to the target nerve fiber bundle in the diffusion tensor image, and the acquisition site of the surface electromyography signal and the target nerve fiber bundle are on the same motor nerve conduction path in the human anatomical structure.

[0067] Diffusion tensor imaging refers to three-dimensional image data obtained by using magnetic resonance imaging technology to detect the microscopic geometric features of water molecule diffusion in living tissue. It can reflect the direction of the brain white matter nerve fiber bundles and the structural integrity, and can include raw signal intensity data in multiple diffusion gradient directions.

[0068] Surface electromyography (EMG) signals are bioelectrical signals recorded by electrode sensors attached to the skin surface during nerve and muscle activity. They reflect the ability of spinal motor neurons to control muscle fibers and the dynamic state of muscle contraction, and can be a series of voltage amplitude sequences that vary over time.

[0069] In the specific implementation process, the head of the target object is first scanned using a magnetic resonance imaging (MRI) scanner. Multiple non-collinear diffusion-sensitive gradient directions are set to obtain a raw dataset of diffusion tensors covering the entire brain or a specific region of interest. .

[0070] Simultaneously, based on the neural innervation relationships in human anatomy, limb muscle groups directly functionally associated with the damaged nerve fiber bundles in diffusion tensor imaging were identified. This ensures a strict correspondence between the structural information on the images and the functional information on electromyography (EMG) along the anatomical pathway. For example, if the focus is on the recovery of the corticospinal tract, the wrist extensor muscles of the affected upper limb were selected as the acquisition targets. Then, during the target subject's performance of specific rehabilitation movements, the surface EMG signals of this muscle group were recorded using a bioelectrical signal acquisition device, yielding a time-series vector. .

[0071] For example, for stroke patient A, assuming that the nerve fiber bundle descending from the damaged left motor cortex is selected as the observation object, and electrodes are simultaneously attached to the surface of the extensor muscles of the right forearm, when patient A attempts to extend the wrist, data including... Diffusion image data in each gradient direction and sampling frequency is Duration is voltage sequence .

[0072] S102. Perform tensor solution on the diffusion tensor image to determine the diffusion value corresponding to each voxel, and generate diffusion parameters by calculating the anisotropy fraction and average diffusion rate based on the diffusion value. Transform the diffusion parameters into a point set of Riemannian manifold through symmetric positive definite matrix mapping to obtain diffusion manifold data.

[0073] Optionally, step S102 involves performing tensor calculations on the diffusion tensor image to determine the diffusion value corresponding to each voxel, generating diffusion parameters by calculating the anisotropy fraction and average diffusion rate based on the diffusion values, and transforming the diffusion parameters into a point set of a Riemannian manifold through a symmetric positive definite matrix mapping to obtain the diffusion manifold data. This process can specifically include:

[0074] Figure 2 A flowchart illustrating a method for generating diffuse manifold data according to an embodiment of this application is shown. Figure 2 As shown, the method includes:

[0075] S1021. Perform tensor diagonalization on the diffusion weighted signal of the diffusion tensor image in multiple diffusion gradient directions to determine the mutually orthogonal diffusion values ​​corresponding to each voxel.

[0076] Diffusion-weighted signal refers to the tissue signal intensity obtained after applying a diffusion-sensitive pulse with a specific gradient during magnetic resonance scanning. Its attenuation is negatively correlated with the diffusion ability of water molecules along the gradient direction, and can be a set of gray values ​​that vary with the gradient direction.

[0077] Tensor diagonalization refers to the process of converting a second-order tensor matrix describing the diffusion characteristics in three-dimensional space into a diagonal matrix through linear algebra operations. The aim is to extract intrinsic physical quantities that do not change with coordinate system rotation. The diffusion value refers to the eigenvalue of the diffusion tensor, obtained after diagonalization, representing the independent diffusion intensity of water molecules along the three mutually orthogonal principal axes.

[0078] In the specific implementation process, the diffusion tensor image data is read first. For each voxel in the image space, a signal intensity attenuation model is established based on the Stejskal-Tanner equation. Next, the least squares method is used to fit the multi-directional diffusion-weighted signal, and the second-order symmetric tensor describing the diffusion characteristics of that voxel is calculated. Finally, the Jacobi iteration method or singular value decomposition algorithm is used to process the tensor. Diagonalization is performed, and the characteristic equation is solved. ,in Given a third-order identity matrix, we obtain three non-negative real eigenvalues, i.e., diffusion values.

[0079] For example, suppose the coordinates in a brain image of stroke patient A are... The voxels were analyzed by processing signals in 32 gradient directions to obtain three mutually orthogonal diffusion values, denoted as follows: ,in Represents the longitudinal diffusion intensity along the main axis of the nerve fiber bundle. and It represents the radial diffusion intensity perpendicular to the fiber bundle direction.

[0080] S1022. Calculate the diffusion anisotropy deviation and average diffusion intensity of the nerve fiber bundle using the diffusion value. Determine the diffusion anisotropy deviation as the anisotropy fraction and the average diffusion intensity as the average diffusion rate. Combine the anisotropy fraction and the average diffusion rate using eigenvectors to generate diffusion parameters.

[0081] Diffusion anisotropy bias refers to the non-uniformity of water molecule diffusion rates in different spatial directions, reflecting the directional arrangement characteristics of tissue microstructure; this value is higher in areas with intact nerve fiber bundles. Diffusion mean intensity refers to the overall average level of water molecule diffusion amplitude in all directions, reflecting changes in the degrees of freedom of water molecules or cell density in the interstitial space. The diffusion parameter is a characteristic vector composed of the anisotropy fraction and the mean diffusion rate, used to highly summarize the integrity state of nerve fiber bundles in a dimensionality-reduced manner.

[0082] In the specific implementation process, firstly, assume that the coordinates in the brain image of stroke patient A are... The voxels yielded diffusion values. The average diffusion rate is obtained by calculating the arithmetic mean of the three values. Secondly, the anisotropy fraction is obtained by quantifying the diffusion bias using the formula for calculating the anisotropy fraction. The calculation formula is shown in formula (1) below:

[0083] (1)

[0084] in, This represents the arithmetic mean of the three diffusion values, i.e. = Finally, these two scalar indices are arranged in an ordered manner to generate the diffusion parameter corresponding to the voxel. .

[0085] S1023. Construct the third-order symmetric positive definite matrix corresponding to each voxel using the diffusion parameter, and use the third-order symmetric positive definite matrix as a mapping operator to project the diffusion parameter onto the Riemannian geometric space to generate the corresponding spatial coordinate points, thus obtaining the point set of the Riemannian manifold.

[0086] A third-order symmetric positive definite matrix is ​​a... A real square matrix whose elements are symmetric about the main diagonal. And for any non-zero vector All have That is, all eigenvalues ​​are positive real numbers. Physically, it corresponds to the ellipsoidal model that describes the probability distribution of water molecule diffusion, and mathematically it constitutes the basic elements of the Riemannian manifold space.

[0087] The point set of a Riemannian manifold refers to the set of geometric coordinates of all voxels corresponding to the third-order symmetric positive definite matrices in the space of a curved SPD (Symmetric Positive Definite) manifold.

[0088] In the specific implementation process, firstly, the average diffusion rate in the dispersion parameter is utilized. and anisotropy fraction Assuming uniform lateral diffusion of nerve fiber bundles, the corresponding longitudinal diffusion characteristic values ​​are derived in reverse. and radial diffusion eigenvalues ,in Specifically, based on the anisotropy score... With average diffusivity The definition, in the case of lateral diffusion isotropic, i.e. Under the assumptions, the following equations can be solved simultaneously: = / 3, and the formula defined by the above formula (1) and , The relationship. Thus, we obtain and The value.

[0089] Next, in conjunction with step S1021 Corresponding feature vector A third-order symmetric positive definite matrix can be directly constructed using the tensor product formula. As shown in the following formula (2):

[0090] (2)

[0091] in, It is a third-order identity matrix, with matrix elements. All are based on and The calculated real value satisfies And the positive definiteness condition, the matrix This is a geometric coordinate point in the Riemannian manifold space.

[0092] S1024. Diffuse manifold data is obtained by aggregating the spatial topology of the point set of the Riemannian manifold.

[0093] Spatial topology aggregation refers to the process of combining scattered independent voxel matrices into an anatomically meaningful overall data structure while preserving the three-dimensional spatial relationships of the original images. Diffuse manifold data refers to a structured dataset of the whole brain or a specific region mapped onto a Riemannian manifold space after aggregation processing.

[0094] In the specific implementation process, according to the spatial resolution and voxel arrangement order of the original diffusion tensor image, a third-order symmetric positive definite matrix is ​​generated for each voxel position. Perform one-to-one backfilling and index construction. For example, suppose that for all regions covering the damaged nerve fiber bundles of patient A... Individual elements, their corresponding A third-order symmetric positive definite matrix According to the original three-dimensional coordinates Structured encapsulation is performed to obtain the final diffuse manifold data. .

[0095] This embodiment preserves the geometric and topological properties of the brain structure while realizing the geometric transformation from Euclidean space to nonlinear manifold space, effectively avoiding the loss of spatial structure information caused by traditional linear methods.

[0096] S103. The motion feature vector is obtained by calculating the mean of the recalculated range of the surface electromyography signal under different observation lengths and performing double log-linear fitting to solve the Hurst exponent.

[0097] Optionally, step S103, which involves calculating the mean rescaled range of surface electromyography signals at different observation lengths and performing a double log-linear fitting to obtain the Hurst exponent, may specifically include:

[0098] S1031. Divide the surface electromyography signal into corresponding signal sub-intervals according to multiple preset observation lengths, calculate the cumulative deviation sequence and standard deviation of the sampling points in each signal sub-interval, determine the range value based on the deviation distance between the maximum and minimum values ​​in the cumulative deviation sequence, and determine the recalibrated range component corresponding to each signal sub-interval based on the ratio of the range value to the standard deviation.

[0099] A signal sub-interval refers to a local data segment obtained by cutting the original electromyography signal into non-overlapping or partially overlapping sections according to a predetermined observation length. It is the basic unit for calculating local statistics. The cumulative deviation sequence refers to the point-by-point summation of the differences between the value of each sampling point within a specific sub-interval and the mean of that interval. The rescaled range component refers to the dimensionless ratio obtained by normalizing the range value of a sub-interval (i.e., the fluctuation range of the cumulative deviation sequence) after normalization to the standard deviation of that interval. It is used to eliminate the influence of the absolute magnitude of the signal amplitude.

[0100] The preset observation length refers to the time window scale artificially selected during rescaled range analysis to divide a long-term series into several short-term segments. Different lengths can reveal the fluctuation characteristics of the signal at different time granularities, and can be a set of integer sequences that grow exponentially or are equally spaced. The preset observation length is set according to different time window types, as shown in Table 1 below:

[0101] Table 1: Preset Observation Length Comparison Table

[0102]

[0103] As shown in Table 1, Table 1 lists the different time window scales selected in this embodiment for capturing the multi-scale dynamic characteristics of electromyographic signals.

[0104] In the specific implementation process, firstly, a set of preset observation lengths is defined, for the total number of data points collected. Surface electromyography signal sequence According to each observation length set in Table 1 ,Will Divided into Two non-overlapping signal sub-intervals. Secondly, for a length of... The Sub-intervals First, calculate the interval. Arithmetic mean of internal data Then calculate the cumulative deviation sequence. ,in .

[0105] Then, find the sequence. The maximum value in and minimum value Calculate the difference between the two to obtain the range. Simultaneously calculate this sub-interval. Standard deviation Finally, the range Divide by standard deviation The sub-interval is obtained. Corresponding rescaled range components .

[0106] For example, for a stroke patient A with a signal of 2000 points in total, assuming a short time window is used... At that time, the signal was divided into 20 sub-intervals, and 20 values ​​were calculated. Value. Select a medium to long time window. At that time, the same signal was re-divided into 4 sub-intervals, and 4 values ​​were calculated. value.

[0107] S1032. Statistically average multiple rescaled range components belonging to the same observation length to obtain the mean of the rescaled range, and construct a logarithmic mapping relationship between the mean of the rescaled range and the corresponding observation length. Determine the evolution slope of the regression line in the logarithmic mapping relationship through linear regression fitting, and determine the evolution slope as the Hearst exponent.

[0108] The rescaled range mean is the arithmetic mean of the rescaled range components corresponding to all signal subintervals under a fixed observation length. The logarithmic mapping relationship refers to the linear dependence between the logarithm of the rescaled range mean and the logarithm of the observation length. This relationship is usually represented by a straight line in a double logarithmic coordinate system, and the slope of the line can quantify the long-range correlation or fractal characteristics of the signal.

[0109] The Hearst exponent is the slope of the regression line mentioned above. Its value is usually between 0 and 1. When the value is greater than 0.5, it indicates that the signal has positive correlation persistence, i.e., memory.

[0110] In the specific implementation process, firstly, for each observation length in Table 1... All of them Rescaled range components of each subinterval Calculate the arithmetic mean to obtain the mean of the double-scaled range for that length. Obtain a set of coordinates Next, the logarithmic mapping relationship is constructed as shown in the following formula (3):

[0111] (3)

[0112] in, For the observation length is The mean of the rescaled range at time , For the current observation length, The intercept constant is... That is, the evolution slope to be solved, i.e., the Hearst exponent. .

[0113] Subsequently, the least squares method was used to analyze this set of data. Perform linear regression fitting and calculate the slope of the best-fit line. As the Hearst index .

[0114] S1033. The multiple Hurst indices collected from different muscle groups are vectorized and arranged in a preset order to generate motion feature vectors.

[0115] Motion feature vectors are numerical vectors composed of Hurst exponents from muscle groups from different anatomical locations or functional groups in a specific order. They provide a multi-dimensional perspective to describe the overall coordinated movement state of the affected limb.

[0116] In the specific implementation process, firstly, assuming that the rehabilitation assessment involves the synergistic effect of multiple muscle groups, such as simultaneously monitoring the forearm extensors, biceps brachii, and triceps brachii, then the calculation process of S1031 and S1032 above is repeated for each muscle group to obtain the Hurst index of the forearm extensors. The biceps brachii is The triceps brachii is Secondly, these scalar indices are arranged according to a preset anatomical order to generate a multidimensional motion feature vector. .

[0117] This embodiment not only effectively eliminates the interference caused by the difference in absolute signal amplitude, but also deeply captures the long-range correlation and nonlinear fractal characteristics that reflect the muscle's ability to coordinate contraction, thereby generating motion feature vectors with clear physical meaning.

[0118] S104. The diffuse manifold data is mapped to the corresponding tangent space through the Riemann logarithmic mapping to generate tangent vectors. The projection transformation matrix is ​​calculated by performing canonical correlation analysis on the tangent vectors and motion feature vectors. After mapping the tangent vectors and motion feature vectors to the correlation space through the projection transformation matrix, the maximum correlation component is extracted to generate coupled feature vectors.

[0119] Optionally, the process in step S104 of mapping the diffuse manifold data to the corresponding tangent space using the Riemann logarithmic mapping to generate tangent vectors, and performing canonical correlation analysis on the tangent vectors and motion feature vectors to calculate the projection transformation matrix can specifically include:

[0120] S1041. Select the reference center point of the manifold space where the diffuse manifold data is located, and use the Riemann logarithmic transform to map the diffuse manifold data to the tangent plane space corresponding to the reference center point to obtain the tangent vector.

[0121] The reference centroid is a special point in the Riemannian manifold space that represents the geometric centroid of the entire dataset, usually defined by the Riemann mean. The Riemann logarithmic transform is a mathematical operation that uses the Riemann logarithmic mapping operator to project points on a curved manifold losslessly onto a local Euclidean tangent plane with the reference centroid as the origin.

[0122] A tangent vector is a data in Euclidean vector form located in the tangent space after logarithmic mapping. It preserves the geodesic distance and direction information of the original manifold points relative to the reference center.

[0123] In the specific implementation process, firstly, based on the results obtained in step S102, including diffuse manifold data of each sample point An iterative algorithm is used to calculate the corresponding Riemann mean, which is then used as the reference center point. Next, with Construct a tangent plane space for the tangent point and utilize the Riemann logarithmic mapping operator. Process them one by one Each data point That is, the third-order symmetric positive definite matrix corresponding to each voxel, which will be used to represent each voxel. matrix It is mapped to a symmetric matrix of the same dimension.

[0124] Finally, extract the independent elements of the upper or lower triangular region of the symmetric matrix and stretch them into column vectors to obtain the corresponding tangent vectors. ,in A 6-dimensional column vector, which corresponds to The six independent elements of a symmetric matrix.

[0125] S1042. The first covariance matrix is ​​obtained by performing autocorrelation distribution statistics on the tangent vector using the regularization term, the second covariance matrix is ​​obtained by performing autocorrelation distribution statistics on the motion feature vector, and the cross-covariance matrix is ​​obtained by performing cross-correlation statistical operations on the tangent vector and the motion feature vector.

[0126] The first covariance matrix is ​​a statistical matrix describing the degree of linear correlation between the dimensions within a tangent vector group, reflecting the autocorrelation distribution of brain microstructural features. The second covariance matrix is ​​a statistical matrix describing the degree of linear correlation between the dimensions within a motion feature vector group, reflecting the autocorrelation distribution of limb muscle dynamics features.

[0127] The cross-covariance matrix is ​​a statistical matrix that describes the degree of cross-correlation between the tangent vector group and the motion feature vector group, and is key to uncovering neuromuscular coupling relationships. The regularization term is a small perturbation parameter introduced to prevent matrix invertibility due to insufficient sample size or data collinearity.

[0128] In the specific implementation process, firstly, the set of tangent vectors is calculated. mean vector Set of motion feature vectors mean vector Secondly, a regularization parameter is introduced. Calculate the autocovariance matrix and crosscovariance matrix of both. First covariance matrix. Second covariance matrix Cross-covariance matrix They are represented as follows:

[0129] ,

[0130] ,

[0131] ,

[0132] in For the identity matrix, elements All are based on sets and The real values ​​obtained by statistically calculating the vector components in the vector array. This is a regularization term.

[0133] S1043. Construct the correlation response function using the first covariance matrix, the second covariance matrix, and the cross-covariance matrix. Perform generalized eigenvalue decomposition on the correlation response function to solve for the linear transformation coefficients that maximize the correlation between the mapping components of the tangent vector and the motion feature vector, and generate the projection transformation matrix.

[0134] The correlation response function is an objective function constructed based on the principle of regularized canonical correlation analysis, which aims to find the maximum correlation coefficient between two sets of variables after projection.

[0135] In the specific implementation process, the first step is to construct the associated response function. As shown in the following formula (4):

[0136] (4)

[0137] in and Let be the projection direction vectors to be solved. Next, perform generalized eigenvalue decomposition on the function to calculate the vectors that make ... The set of basis vectors that is maximized is the linear transformation coefficient. Finally, these coefficients are used to construct the projection transformation matrix. Its form is:

[0138]

[0139] The submatrix formed by the first p rows The submatrix formed by the last q rows of the projected tangent vector. Used to project motion feature vectors.

[0140] Optionally, the process of mapping the tangent vector and the motion feature vector to the association space using the projection transformation matrix and then extracting the maximum correlation component to generate the coupled feature vector in step S104 can specifically include:

[0141] S1044. Perform linear coordinate transformation using the submatrices in the projection transformation matrix that correspond to the tangent vector and the motion feature vector, respectively, to obtain the first projection vector group and the second projection vector group located in the associated space.

[0142] The first projection vector set refers to the new vector set obtained by projecting the original high-dimensional tangent vector data into a specific low-dimensional subspace, namely the associated space, through a linear transformation. It represents the expression of brain microstructural features in this shared space. The second projection vector set refers to the new vector set obtained by projecting the original motion feature vector data into the same associated space through a linear transformation. It represents the expression of limb dynamics features in this shared space.

[0143] A submatrix refers to an independent transformation matrix block that is partitioned from the overall projection transformation matrix according to its feature dimension, and each submatrix acts on different modalities of data. The correlation space refers to the latent feature space constructed through canonical correlation analysis, in which two sets of heterogeneous data exhibit the greatest statistical correlation.

[0144] In the specific implementation process, firstly, the projection transformation matrix generated in step S1043 is... The matrix is ​​split into two sub-matrices, denoted as follows: That is, the corresponding tangent vector dimension and This corresponds to the dimension of motion features. Next, for the set of tangent vectors... Each vector Using the matrix multiplication formula Perform coordinate transformation to generate the first projection vector set. Similarly, for the set of motion feature vectors... Each vector Using the formula Perform the transformation to generate the second set of projection vectors. .

[0145] S1045. Perform pairwise linear correlation operations on the components of each dimension in the first projection vector group and the second projection vector group, and extract the pair of components with the highest linear correlation coefficient as the maximum correlation feature pair.

[0146] Linear metric operations refer to the process of calculating the Pearson correlation coefficient or inner product of two vector sequences within a correlation space, aiming to quantify their statistical similarity. The most correlated feature pair refers to the pair of components with the largest absolute value of the linear correlation coefficient among the multidimensional projective components, i.e., the canonical variable pair, which represents the strongest intrinsic coupling pattern between brain structure and limb function.

[0147] In the specific implementation process, firstly, assume that the projection dimension is... For the first projection vector group Second projection vector group Then, extract their component sequences in each dimension. Next, for the... Each dimension is used to calculate the component sequence. and Pearson correlation coefficient between ,in Finally, compare the correlation coefficients across all dimensions and find the one with the largest value. and its corresponding dimensional index The components in this dimension The pair of features was identified as having the highest correlation.

[0148] For example, suppose that the first projection vector set is obtained after projection. Second projection vector group Calculations revealed that the first dimension had the highest correlation, so it was extracted. As the most relevant feature pair.

[0149] S1046. Assign weight coefficients based on the correlation strength of the most correlated component, and use the weight coefficients to perform a linear summation operation on each modal feature component in the most correlated feature pair to generate a coupled feature vector.

[0150] Correlation strength refers to the absolute value of the linear correlation coefficient between the most correlated feature pairs, which measures the tightness of the neuromuscular coupling relationship. A coupled feature vector is a single numerical value or vector generated by a weighted linear combination of the two components of the most correlated feature pair; it can simultaneously represent the recovery status of brain structure and the state of limb motor function.

[0151] In the specific implementation process, firstly, the maximum correlation coefficient determined in step S1045 is obtained. and the corresponding maximum correlation feature pair Secondly, in order to generate a single coupled feature, based on A weighting strategy is set. Typically, to balance the information from the two modes, equal-weighted addition or signal-to-noise ratio-based weighting can be used. In this embodiment, a simple linear weighted summation formula is used to generate the coupled feature vector. ,Right now ,in This is the weighting coefficient, usually taken as 0.5, or according to... The size is dynamically adjusted.

[0152] This embodiment not only effectively eliminates the difference in physical dimensions between brain imaging features and limb electromyography features, but also maximizes the mining of deep coupling relationships in the neuromuscular system, thereby generating high-confidence coupling feature vectors and solving the technical problem of direct fusion and analysis of heterogeneous multimodal data.

[0153] S105. Using the Riemann index mapping, the coupled feature vector is mapped to the Riemann manifold space to determine the target reconstruction site. Based on the Riemann metric tensor, the geodesic length between the target reconstruction site and the reference state point in the preset reference manifold is calculated to obtain the deviation value.

[0154] Optionally, step S105, which uses the Riemann exponent mapping to map the coupled feature vector to the Riemann manifold space to determine the target reconstruction site and calculates the geodesic length between the target reconstruction site and the reference state point in the preset reference manifold based on the Riemann metric tensor to obtain the deviation value, may specifically include:

[0155] S1051. Using the reference center point as the origin of the mapping, perform exponential coordinate transformation on the coupled feature vector to map the coupled feature vector from the linear tangent space back to the surface coordinate system where the Riemann manifold space is located, and determine the target reconstruction point.

[0156] Riemann exponential mapping is a mathematical operation that uses the Riemann exponential operator to project a linear vector in tangent space back onto a Riemann manifold surface. It recovers the nonlinear geometric structure of the data by wrapping the vector back around the manifold along geodesic directions. The exponential coordinate transformation process is the specific computational procedure for implementing this mapping. The target reconstruction point refers to the specific location of the coupled eigenvector in the Riemann manifold space after the exponential mapping, typically represented by a symmetric positive definite matrix.

[0157] In the specific implementation process, firstly, the reference center point determined in step S1041 is invoked. As the origin of the mapping. Then, because... Originally located in the linear tangent space, it needs to first utilize the projection transformation sub-matrix obtained in step S1043. The generalized inverse, will Restored to 6-dimensional tangent space vector form Finally, regarding Perform Riemann index mapping to obtain the target reconstructed site. The mapping formula is shown in formula (5) below:

[0158] (5)

[0159] in It is a matrix exponential function.

[0160] S1052. Based on the coordinate difference between the target reconstruction point and the reference state point in the surface coordinate system, the path is discretized to obtain the first path step length sequence connecting the target reconstruction point and the reference state point.

[0161] The baseline state point refers to a standard geometric point in a predefined baseline manifold that represents a healthy population or an ideal recovery state. It is typically obtained by statistical averaging of a large number of healthy samples. Path discretization refers to the process of dividing the continuous geometric path connecting the target reconstruction site and the baseline state point into several small straight line segments to facilitate the calculation of the path length using numerical methods. The first path step length sequence is an ordered set of tiny coordinate difference vectors obtained after discretization, describing the linear displacement of each step from the pathological position to the healthy position.

[0162] In the specific implementation process, firstly, a reference state point is set. For example, the Riemann mean of the healthy control group. Secondly, in order to calculate... and The distance between them is introduced into a parameterized path in the surface coordinate system of the Riemannian manifold. ,in Parameter range Divided into equal parts Each segment corresponds to a discrete time point. Subsequently, the coordinate differences between corresponding manifold points at adjacent time points are calculated, yielding a series of small displacement vectors. Finally, these This constitutes the first path step length sequence.

[0163] S1053. Using the Riemann metric tensor of the location of each discrete step length in the first path step length sequence as a local geometric scaling operator, perform a weighted inner product operation on the coordinate displacement component corresponding to each discrete step length and the Riemann metric tensor to calculate the local geometric length of each discrete step length in the Riemann manifold space, and obtain the second path step length sequence composed of each local geometric length.

[0164] The Riemannian metric tensor is an inner product function defined at every point on a manifold. It provides a local distance measurement standard for the manifold space and can dynamically adjust the calculation method of vector length according to the curvature of the manifold. In a symmetric positive definite matrix manifold, a point... Riemannian metric tensor at the location As shown in the following formula (6):

[0165] (6)

[0166] in, Indicates that it is located at point In the path integral process of this scheme, two tangent vectors in the tangent space are usually the same small displacement vector. , The trace operation represents the sum of the elements on the main diagonal of a matrix. Indicates at the base point The inner product symbol at the location.

[0167] The local geometric scaling operator refers to the specific numerical matrix of the aforementioned metric tensor at a particular point, which performs a weighted scaling of the Euclidean displacement. Local geometric length refers to the true length of a small path segment in curved space, obtained after correction using the metric tensor.

[0168] In the specific implementation process, firstly, for each tiny displacement in the first path step size sequence... That is, the manifold points at two adjacent time points and The coordinate difference matrix between the two points determines the starting point of the displacement. Right now Next, using the above formula (6) as the local geometric scaling operator, the small displacement is calculated according to the definition of the affine invariant Riemannian metric on a symmetric positive definite matrix manifold. Corresponding local geometric length The specific calculation formula is shown in formula (7) below:

[0169] (7)

[0170] in It is a base point matrix The inverse matrix serves as the expression for the displacement vector. This serves as a standardization mechanism. Finally, for all steps in the first path length sequence... Repeat the above steps for each discrete segment to obtain the second path step length sequence, which includes all local geometric lengths. .

[0171] S1054. Perform numerical cumulative integration processing along the Riemannian manifold space surface on the second path step length sequence to solve the minimum trajectory path length between the target reconstruction point and the reference state point, and determine the minimum trajectory path length as the deviation value.

[0172] The minimum trajectory path length refers to the shortest path among all possible paths connecting two points, i.e., the length of the geodesic, which is an intrinsic distance metric on a Riemannian manifold. Numerical cumulative integration is a mathematical process of summing discrete local geometric lengths to approximate the total length of the continuous path. The deviation value is the final calculated geodesic length, quantifying the degree of difference between the patient's current state and their ideal health state.

[0173] In the specific implementation process, firstly, the second path step length sequence obtained in step S1053 is processed. The target point is obtained by summing all the elements in the summation. With reference point Total path length Secondly, because when the number of discrete parts When large enough, Converging to the geodesic distance between two points The final calculated This value has been identified as a deviation.

[0174] This embodiment strictly follows the intrinsic geometric structure of the data, and avoids the distortion problem of Euclidean distance in curved space by calculating the minimum trajectory length along the manifold surface to measure the difference.

[0175] S106. Determine the target division interval based on the deviation value and the preset division interval, and determine the diagnostic result based on the motor function level corresponding to the target division interval.

[0176] Figure 3 A flowchart illustrating a method for determining diagnostic results according to an embodiment of this application is shown. Figure 3 As shown, diffusion tensor images of the brain and electromyographic signals of the damaged nerve innervation area were acquired simultaneously. Imaging features representing the integrity of brain nerve fibers and signal features representing muscle control ability were extracted respectively.

[0177] Subsequently, the data from these two different sources were fused for a combined brain-muscle assessment analysis to examine the synergistic recovery relationship between central brain commands and peripheral limb responses. Finally, by calculating the degree to which the patient's current state deviates from a healthy baseline, the effect of neural remodeling was quantified, and the corresponding motor function level diagnostic result was obtained. This method includes:

[0178] The pre-defined intervals refer to a set of numerical ranges that are predetermined based on the statistical distribution characteristics of large-scale clinical sample data and used to discretize continuous geodesic distance deviations into different rehabilitation stages. These intervals usually correspond to the medically recognized classification standards for motor dysfunction.

[0179] Motor function level refers to a qualitative or quantitative label describing the degree of recovery of the patient's limb motor control ability, corresponding to each division interval. Examples include specific grades in the Fugl-Meyer Assessment scale or general clinical descriptions such as severe impairment or moderate impairment. The diagnostic result is the final output report generated based on the matched motor function level, combined with automated analysis, regarding the patient's neural remodeling status and recommended rehabilitation strategies. The motor function level division intervals are shown in Table 2 below.

[0180] Table 2: Comparison Table of Motor Function Level Classification Intervals

[0181]

[0182] As shown in Table 2, Table 2 divides the rehabilitation status into four typical grade intervals based on the distribution pattern of geodesic distances.

[0183] In the specific implementation process, the mapping relationship between deviation values ​​and motor function levels is first established, as shown in Table 2 above. Next, the deviation values ​​calculated in step S105 are obtained. .Will The boundary values ​​of each interval in Table 2 are compared one by one to determine the target interval to which it belongs. Specifically, if Then determine the deviation value. The data falls into the corresponding interval number, and the corresponding motor function level is extracted as a preliminary judgment result. Finally, the level label, the corresponding clinical status description, and the specific deviation value are integrated into the final diagnostic result output.

[0184] For example, for stroke patient A, suppose the deviation value is... Referring to Table 2, we can see that... Located in the interval Within this range, it falls into category III, which is moderate impairment. Therefore, patient A's target classification range is determined to be III, and the confirmed diagnosis is moderate motor dysfunction. Targeted, systematic neurorehabilitation training is recommended.

[0185] This embodiment not only makes the diagnostic results more intuitive and hierarchical, but also establishes a quantitative channel from microscopic data characteristics to macroscopic rehabilitation status. This allows doctors to quickly determine the patient's neural remodeling stage based on objective numerical data, thereby formulating more targeted and scientific personalized rehabilitation treatment plans, significantly improving the clinical practical value of auxiliary diagnosis.

[0186] Figure 4 This application provides a schematic diagram of a specific implementation of a computer-aided diagnostic system based on medical images, referring to... Figure 4 The system may include:

[0187] The acquisition module 410 is used to acquire diffusion tensor images of the target object's brain and surface electromyographic signals of muscle groups that are innervated by nerve fiber bundles associated with the diffusion tensor images under rehabilitation movements.

[0188] The calculation module 420 is used to perform tensor solution on the diffusion tensor image to determine the diffusion value corresponding to each voxel, and generate diffusion parameters by calculating the anisotropy fraction and average diffusion rate based on the diffusion value. The diffusion parameters are then transformed into a point set of the Riemannian manifold through symmetric positive definite matrix mapping to obtain diffusion manifold data.

[0189] The generation module 430 is used to calculate the mean of the recalculated range of the surface electromyography signal at different observation lengths and perform double log-linear fitting to solve the Hurst exponent, thereby obtaining the motion feature vector.

[0190] The generation module 430 is also used to map the diffuse manifold data to the corresponding tangent space through the Riemann logarithmic mapping to generate tangent vectors, perform canonical correlation analysis on the tangent vectors and motion feature vectors to calculate the projection transformation matrix, and after mapping the tangent vectors and motion feature vectors to the association space through the projection transformation matrix, extract the maximum correlation component to generate coupled feature vectors.

[0191] The generation module 430 is also used to map the coupled feature vector to the Riemann manifold space using the Riemann exponent mapping to determine the target reconstruction site, and to calculate the geodesic length between the target reconstruction site and the reference state point in the preset reference manifold based on the Riemann metric tensor to obtain the deviation value.

[0192] The determination module 440 is used to determine the target division interval based on the deviation value and the preset division interval, and to determine the diagnostic result based on the motor function level corresponding to the target division interval.

[0193] The computer-aided diagnostic system based on medical images in this application is used to implement the aforementioned computer-aided diagnostic method based on medical images. Therefore, the specific implementation of the computer-aided diagnostic system based on medical images can be found in the embodiment section of the computer-aided diagnostic method based on medical images above. The specific implementation can be referred to the description of the corresponding embodiments, which will not be repeated here.

[0194] Figure 5 A schematic diagram of the hardware structure of an electronic device provided in one embodiment of this application is shown.

[0195] The electronic device may include a processor 510 and a memory 520 storing computer program instructions.

[0196] Specifically, the processor 510 may include a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0197] Memory 520 may include mass storage for data or instructions. For example, and not limitingly, memory 520 may include a hard disk drive (HDD), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 520 may include removable or non-removable (or fixed) media. Where appropriate, memory 520 may be internal or external to the integrated gateway disaster recovery device. In a particular embodiment, memory 520 is non-volatile solid-state memory.

[0198] Memory may include read-only memory (ROM), random access memory (RAM), disk storage media devices, optical storage media devices, flash memory devices, and electrical, optical, or other physical / tangible memory storage devices. Therefore, typically, memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., memory devices) encoded with software including computer-executable instructions, and when the software is executed (e.g., by one or more processors), it is operable to perform the operations described with reference to the method according to the first aspect of this disclosure.

[0199] The processor 510 reads and executes computer program instructions stored in the memory 520 to implement any of the computer-aided diagnostic methods based on medical images in the above embodiments.

[0200] In one example, the electronic device may also include a communication interface 530 and a bus 540. Wherein, such as Figure 5 As shown, the processor 510, memory 520, and communication interface 530 are connected through bus 540 and complete communication with each other.

[0201] The communication interface 530 is mainly used to realize communication between various modules, devices, units and / or equipment in the embodiments of this application.

[0202] Bus 540 includes hardware, software, or both, that couples components of an online data traffic metering device together. For example, and not limitingly, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Enhanced Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), HyperTransport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a Microchannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or other suitable buses, or combinations of two or more of these. Where appropriate, bus 540 may include one or more buses. Although specific buses are described and illustrated in embodiments of this application, any suitable bus or interconnect is contemplated herein.

[0203] The electronic device can execute the computer-aided diagnosis method based on medical images in the embodiments of this application, thereby realizing the computer-aided diagnosis method based on medical images described in conjunction with the accompanying drawings.

[0204] Furthermore, in conjunction with the computer-aided diagnosis method based on medical images in the above embodiments, this application embodiment can provide a computer-readable storage medium for implementation. This computer-readable storage medium stores computer program instructions; when executed by a processor, these computer program instructions implement any of the computer-aided diagnosis methods based on medical images in the above embodiments.

[0205] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0206] The functional blocks shown in the above-described structural diagram can be implemented as hardware, software, firmware, or a combination thereof. When implemented in hardware, they can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. Programs or code segments can be stored on a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried on a carrier wave. "Machine-readable medium" can include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, fiber optic media, radio frequency (RF) links, etc. Code segments can be downloaded via computer networks such as the Internet, intranets, etc.

[0207] It should also be noted that the exemplary embodiments mentioned in this application describe methods or systems based on a series of steps or apparatus. However, this application is not limited to the order of the above steps; that is, the steps can be performed in the order mentioned in the embodiments, or in a different order, or several steps can be performed simultaneously.

[0208] The aspects of this application have been described above with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It should be understood that each block in the flowchart illustrations and / or block diagrams, and 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, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that these instructions, executable via the processor of the computer or other programmable data processing apparatus, enable the implementation of the functions / actions specified in one or more blocks of the flowchart illustrations and / or block diagrams. Such a processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor, or a field-programmable logic circuit. It is also understood that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can also be implemented by dedicated hardware performing the specified functions or actions, or can be implemented by a combination of dedicated hardware and computer instructions.

[0209] The foregoing has provided a detailed description of a computer-aided diagnostic method and system based on medical images provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A computer-aided diagnostic method based on medical imaging, applicable to the assessment of neural remodeling in stroke patients, characterized in that, include: Acquire diffusion tensor images of the target subject's brain and surface electromyographic signals of muscle groups innervated by nerve fiber bundles associated with the diffusion tensor images during rehabilitation movements. Tensor solving is performed on the diffusion tensor image to determine the diffusion value corresponding to each voxel, and diffusion parameters are generated by calculating the anisotropy fraction and average diffusion rate based on the diffusion values. The diffusion parameters are then transformed into a point set of a Riemannian manifold through symmetric positive definite matrix mapping to obtain diffusion manifold data. The motion feature vector is obtained by calculating the mean recalculated range of the surface electromyography signal at different observation lengths and performing double log-linear fitting to solve the Hurst exponent. The diffuse manifold data is mapped to the corresponding tangent space by the Riemann logarithmic mapping to generate a tangent vector. The projection transformation matrix is ​​calculated by performing canonical correlation analysis on the tangent vector and the motion feature vector. After mapping the tangent vector and the motion feature vector to the association space by the projection transformation matrix, the maximum correlation component is extracted to generate a coupled feature vector. The coupled feature vector is mapped to the Riemann manifold space using the Riemann exponent mapping to determine the target reconstruction site, and the geodesic length between the target reconstruction site and the reference state point in the preset reference manifold is calculated based on the Riemann metric tensor to obtain the deviation value; The target division interval is determined based on the deviation value and the preset division interval, and the diagnostic result is determined based on the motor function level corresponding to the target division interval. The motion feature vector is obtained by calculating the mean rescaled range of the surface electromyography signal at different observation lengths and performing a double log-linear fitting to solve for the Hurst exponent, including: The surface electromyography signal is divided into corresponding signal sub-intervals according to multiple preset observation lengths. The cumulative deviation sequence and standard deviation of the sampling points in each signal sub-interval are calculated. The range value is determined according to the deviation distance between the maximum and minimum values ​​in the cumulative deviation sequence. The recalibrated range component corresponding to each signal sub-interval is determined according to the ratio of the range value to the standard deviation. Statistical averaging is performed on multiple rescaled range components belonging to the same observation length to obtain the mean of the rescaled range, and a logarithmic mapping relationship between the mean of the rescaled range and the corresponding observation length is constructed. The evolution slope of the regression line in the logarithmic mapping relationship is determined by linear regression fitting, and the evolution slope is determined as the Hearst exponent. The multiple Hearst indices collected from different muscle groups are vectorized and arranged in a preset order to generate the motion feature vector.

2. The computer-aided diagnostic method based on medical images according to claim 1, characterized in that, The process involves performing tensor solving on the diffuse tensor image to determine the diffusion value corresponding to each voxel, generating diffusion parameters by calculating the anisotropy fraction and average diffusion rate based on the diffusion values, and transforming the diffusion parameters into a point set of a Riemannian manifold through a symmetric positive definite matrix mapping to obtain diffuse manifold data, including: Tensor diagonalization is performed on the diffusion weighted signal of the diffusion tensor image in multiple diffusion gradient directions to determine the mutually orthogonal diffusion values ​​corresponding to each voxel. The diffusion anisotropy deviation and average diffusion intensity of the nerve fiber bundle are calculated using the diffusion value, the diffusion anisotropy deviation is determined as the anisotropy fraction, the average diffusion intensity is determined as the average diffusion rate, and the anisotropy fraction and the average diffusion rate are merged by feature vector merging to generate the diffusion parameter. Using the diffusion parameters, construct the third-order symmetric positive definite matrix corresponding to each voxel. Use the third-order symmetric positive definite matrix as a mapping operator to project the diffusion parameters onto the Riemannian geometric space to generate the corresponding spatial coordinate points, thereby obtaining the point set of the Riemannian manifold. The diffuse manifold data is obtained by performing spatial topological aggregation on the point set of the Riemannian manifold.

3. The computer-aided diagnostic method based on medical images according to claim 1, characterized in that, The step of mapping the diffuse manifold data to the corresponding tangent space using the Riemann logarithmic mapping to generate a tangent vector, and then performing canonical correlation analysis on the tangent vector and the motion feature vector to calculate the projection transformation matrix, includes: Select the reference center point of the manifold space where the diffuse manifold data is located, and use the Riemann logarithmic transform to map the diffuse manifold data to the tangent plane space corresponding to the reference center point to obtain the tangent vector; The first covariance matrix is ​​obtained by performing autocorrelation distribution statistics on the tangent vector using a regularization term, the second covariance matrix is ​​obtained by performing autocorrelation distribution statistics on the motion feature vector, and the cross-covariance matrix is ​​obtained by performing cross-correlation statistical operations on the tangent vector and the motion feature vector. A correlation response function is constructed using the first covariance matrix, the second covariance matrix, and the cross-covariance matrix. The correlation response function is then subjected to generalized eigenvalue decomposition to calculate the linear transformation coefficients that maximize the correlation between the mapping components of the tangent vector and the motion feature vector, thereby generating the projection transformation matrix.

4. The computer-aided diagnostic method based on medical images according to claim 3, characterized in that, The step of mapping the tangent vector and the motion feature vector to the association space using the projection transformation matrix, and then extracting the maximum correlation component to generate the coupled feature vector includes: By performing linear coordinate transformation processing using the submatrices in the projection transformation matrix that correspond to the tangent vector and the motion feature vector respectively, a first projection vector group and a second projection vector group located in the associated space are obtained. Perform pairwise linear correlation operations on each dimension component in the first projection vector group and the second projection vector group, and extract the component pair with the highest linear correlation coefficient as the maximum correlation feature pair; Weighting coefficients are assigned based on the correlation strength of the maximum correlation component, and the modal feature components in the maximum correlation feature pair are linearly summed using the weighting coefficients to generate the coupled feature vector.

5. The computer-aided diagnostic method based on medical images according to claim 3, characterized in that, The process involves using the Riemann index mapping to map the coupled feature vector to the Riemannian manifold space to determine the target reconstruction site, and calculating the geodesic length between the target reconstruction site and the reference state point in the preset reference manifold based on the Riemann metric tensor to obtain the deviation value, including: Using the reference center point as the mapping origin, perform exponential coordinate transformation on the coupled feature vector to map the coupled feature vector from the linear tangent space back to the surface coordinate system where the Riemannian manifold space is located, and determine the target reconstruction site; Based on the coordinate difference between the target reconstruction point and the reference state point in the surface coordinate system, the path is discretized to obtain the first path step sequence connecting the target reconstruction point and the reference state point; Using the Riemann metric tensor at the location of each discrete step length in the first path step length sequence as a local geometric scaling operator, a weighted inner product operation is performed between the coordinate displacement component corresponding to each discrete step length and the Riemann metric tensor to calculate the local geometric length of each discrete step length in the Riemann manifold space, thereby obtaining a second path step length sequence composed of each local geometric length. The second path step sequence is subjected to numerical cumulative integration along the Riemannian manifold space surface to calculate the minimum trajectory path length connecting the target reconstruction point and the reference state point, and the minimum trajectory path length is determined as the deviation value.

6. The computer-aided diagnostic method based on medical images according to claim 1, characterized in that, The muscle group is located in the innervation region corresponding to the target nerve fiber bundle in the diffusion tensor image, and the acquisition site of the surface electromyography signal and the target nerve fiber bundle are on the same motor nerve conduction path in human anatomical structure.

7. A computer-aided diagnostic system based on medical images, characterized in that, include: The acquisition module is used to acquire diffusion tensor images of the target object's brain and surface electromyographic signals of muscle groups innervated by nerve fiber bundles associated with the diffusion tensor images during rehabilitation movements. The calculation module is used to perform tensor solving on the diffusion tensor image to determine the diffusion value corresponding to each voxel, and generate diffusion parameters by calculating the anisotropy fraction and average diffusion rate based on the diffusion value. The diffusion parameters are then transformed into a point set of a Riemannian manifold through a symmetric positive definite matrix mapping to obtain diffusion manifold data. The generation module is used to calculate the mean of the rescaled range of the surface electromyography signal under different observation lengths and perform double log-linear fitting to calculate the Hurst exponent, thereby obtaining the motion feature vector; the generation module is also used to divide the surface electromyography signal into corresponding signal sub-intervals according to multiple preset observation lengths, calculate the cumulative deviation sequence and standard deviation of the sampling points in each signal sub-interval, determine the range value according to the deviation distance between the maximum and minimum values ​​in the cumulative deviation sequence, and determine the rescaled range component corresponding to each signal sub-interval according to the ratio of the range value to the standard deviation; The mean of the rescaled range is obtained by statistically averaging multiple rescaled range components belonging to the same observation length, and a logarithmic mapping relationship between the mean of the rescaled range and the corresponding observation length is constructed. The evolution slope of the regression line in the logarithmic mapping relationship is determined by linear regression fitting, and the evolution slope is determined as the Hearst exponent. Multiple Hearst exponents collected from different muscle groups are vectorized and arranged in a preset order to generate the motion feature vector. The generation module is also used to map the diffuse manifold data to the corresponding tangent space through the Riemann logarithmic mapping to generate a tangent vector, perform canonical correlation analysis on the tangent vector and the motion feature vector to calculate the projection transformation matrix, and after mapping the tangent vector and the motion feature vector to the association space through the projection transformation matrix, extract the maximum correlation component to generate a coupled feature vector. The generation module is also used to map the coupled feature vector to the Riemann manifold space using the Riemann exponent mapping to determine the target reconstruction site, and to calculate the geodesic length between the target reconstruction site and the reference state point in the preset reference manifold based on the Riemann metric tensor to obtain the deviation value. The determination module is used to determine the target division interval based on the deviation value and the preset division interval, and to determine the diagnostic result based on the motor function level corresponding to the target division interval.

8. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of the computer-aided diagnostic method based on medical images as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the implementation of the computer-aided diagnostic method based on medical images as described in any one of claims 1 to 6.

Images