A cardiovascular disease large model cognitive system

Abstract

The present application relates to the technical field of pattern recognition, in particular to a cardiovascular disease large model cognitive system, the system comprises: a parameter initialization module constructs a gray matrix and maps a convolution kernel to a virtual flow velocity component, a divergence constraint module calculates a gradient component to generate a divergence value and a correction coefficient, a feature reasoning module generates a feature atlas using convolution kernel operation and outputs a prediction value, and a weight correction module combines a loss value and the correction coefficient to construct a comprehensive gradient to update a parameter matrix in reverse. In the present application, by introducing a fluid mechanics divergence constraint mechanism, the neural network is forced to follow the fluid incompressibility law when extracting features, guiding the evolution of weight parameters in the direction consistent with the fluid mechanics law, ensuring that the feature atlas reflects the hemodynamic state in the coronary artery, inhibiting non-physical noise interference and enhancing the sensitivity of subtle blood flow anomaly perception, and improving the physical interpretability and recognition accuracy of the model in complex angiography image analysis.

Description

Technical Field

[0001] This invention relates to the field of pattern recognition technology, and in particular to a large-scale cognitive system for cardiovascular diseases. Background Technology

[0002] Pattern recognition technology refers to the field of technology that uses computer systems to automatically process, analyze, and classify data samples through mathematical algorithms. It encompasses the entire process from raw signal acquisition, preprocessing, feature extraction to final category determination. Traditional large-scale cognitive systems for cardiovascular diseases refer to computer server terminals deployed based on a local area network architecture. Their hardware structure includes a central processing unit (CPU), random access memory (RAM), and a large-capacity disk array. During system operation, the system retrieves the patient's electronic medical record text and biochemical indicator data through a standard database interface. The CPU executes convolutional neural network or recurrent neural network program code stored on the hard drive, performing layer-by-layer numerical calculations and feature matching on the input data matrix. Finally, the calculated disease classification label is directly output and displayed on the LCD screen of the doctor's workstation.

[0003] Existing cognitive systems based on general neural networks rely solely on pixel-level numerical calculations for feature fitting when processing coronary artery images, neglecting the blood fluid dynamics physical laws inherent in the angiographic images. This results in convolution kernel parameters only capturing surface texture information during iterative updates, failing to characterize the true physical properties of the intravascular flow field. Consequently, when faced with overlapping blood vessels or uneven contrast agent distribution, the model is prone to misjudging image noise as pathological features. The weight matrix, lacking physical constraints, struggles to maintain the flow conservation property in feature extraction, leading to a lack of interpretability in the reasoning logic when dealing with atypical lesion samples. Consequently, the final classification results fail to accurately reflect subtle abnormal changes in coronary hemodynamics, reducing the auxiliary decision-making value of intelligent diagnosis in complex clinical scenarios. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a large-scale cognitive system for cardiovascular diseases.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a large-scale cognitive system for cardiovascular diseases, comprising: The parameter initialization module acquires a digital image of coronary angiography, extracts the pixel brightness information of the digital image of coronary angiography, constructs a grayscale value matrix, initializes the convolution kernel parameter matrix, and maps the numerical elements in the convolution kernel parameter matrix to virtual velocity vector components. The divergence constraint module calculates the difference between adjacent row values ​​in the convolution kernel parameter matrix to generate a horizontal velocity gradient component, calculates the difference between adjacent column values ​​in the convolution kernel parameter matrix to generate a vertical velocity gradient component, sums the horizontal velocity gradient component and the vertical velocity gradient component to generate a divergence value, and calculates the absolute value of the deviation between the divergence value and the zero target to generate a correction coefficient. The feature inference module calls a convolutional neural network, uses the convolutional kernel parameter matrix to perform weighted summation and sliding window operations on the gray-level value matrix to generate feature map data, and inputs the feature map data to a fully connected classification layer for nonlinear mapping to output predicted values. The weight correction module compares the predicted value with the doctor's labeled diagnosis value to calculate the loss value, and weights and combines the correction coefficient with the loss value to construct a comprehensive gradient value. The comprehensive gradient value is then used to inversely update the convolution kernel parameter matrix.

[0006] As a further aspect of the present invention, the grayscale value matrix is ​​constructed from the pixel brightness information of the coronary angiography digital image; the virtual velocity vector component is a numerical element in the convolution kernel parameter matrix; the horizontal velocity gradient component is the difference between adjacent row values ​​in the convolution kernel parameter matrix; the vertical velocity gradient component is the difference between adjacent column values ​​in the convolution kernel parameter matrix; the divergence value is calculated from the horizontal and vertical velocity gradient components; the correction coefficient is determined by the absolute value of the deviation between the divergence value and the zero-value target; the feature map data is the weighted summation of the convolution kernel parameter matrix and the grayscale value matrix, and the result of sliding window operation; the predicted value is the nonlinear mapping output result of the fully connected classification layer; the loss value is calculated from the predicted value and the doctor's labeled diagnostic value; the comprehensive gradient value is composed of the correction coefficient and the loss value; and the convolution kernel parameter matrix is ​​a set of parameters updated inversely using the comprehensive gradient value.

[0007] As a further aspect of the present invention, the parameter initialization module specifically includes: The image preprocessing submodule acquires the digital image of the coronary angiography, performs noise reduction processing on the image using a Gaussian smoothing filter, extracts the brightness information of the pixels using an adaptive grid discretization algorithm, and normalizes its value to the range of zero to one, thereby constructing the gray value matrix that can preserve the texture features of the microvascular edges. The kernel parameter construction submodule generates initial random weight values ​​that follow a uniform distribution based on the preset number of convolutional neural network layers and convolutional kernel size configuration, and constructs the convolutional kernel parameter matrix used to extract spatial morphological features of the image. The vector mapping submodule establishes a topological mapping relationship between the numerical elements of the convolution kernel parameter matrix and the virtual hydrodynamic velocity field. It defines the numerical elements in the odd-numbered columns of the matrix as the virtual velocity vector components in the horizontal direction and the numerical elements in the even-numbered columns of the matrix as the virtual velocity vector components in the vertical direction, thereby generating an initialization parameter set with the potential for physical field evolution.

[0008] As a further aspect of the present invention, the divergence constraint module specifically includes: The horizontal gradient calculation submodule obtains the virtual velocity vector component, performs discrete difference operation along the row direction of the convolution kernel parameter matrix, calculates the numerical difference between the current grid position element and the adjacent grid element on the right, and generates the horizontal velocity gradient component that characterizes the rate of change of the virtual fluid in the horizontal direction. The vertical gradient calculation submodule performs discrete difference operations along the column direction of the convolution kernel parameter matrix to calculate the numerical difference between the current grid position element and the adjacent grid element below, and generates the vertical velocity gradient component that characterizes the rate of change of the virtual fluid in the vertical direction. The divergence correction submodule performs point-by-point superposition and summation of the horizontal velocity gradient component and the vertical velocity gradient component to generate the divergence value reflecting the local fluid source and sink characteristics. It then calculates the absolute value of the deviation between the divergence value and the zero target under ideal incompressible fluid conditions and generates the correction coefficient used to penalize the distribution of non-physical parameters.

[0009] As a further aspect of the present invention, the feature reasoning module specifically includes: The convolution operation submodule obtains the gray value matrix and the convolution kernel parameter matrix, moves a sliding window on the gray value matrix according to a preset step size, performs a Hadamard product operation on the pixel brightness values ​​in the window and the corresponding convolution kernel weights and sums them to generate the feature map data that extracts the features of local vascular stenosis or occlusion. The classification mapping submodule flattens the feature map data into a one-dimensional feature vector, inputs it into a multi-layer fully connected neural network, maps the high-dimensional features to the disease category space through matrix multiplication and bias addition operations, processes the mapping result using the Softmax activation function, and outputs the predicted value representing the probability distribution of cardiovascular disease.

[0010] As a further aspect of the present invention, the weight correction module specifically includes: The loss evaluation submodule obtains the predicted value output by the model and the doctor-annotated diagnosis value pre-labeled by clinical experts, calculates the probability distribution difference between the two using the cross-entropy loss function, and generates the loss value that can quantify the model's classification prediction error. The gradient fusion submodule introduces a physical constraint regularization weight factor to perform a weighted linear combination of the correction coefficient and the loss value, thereby constructing the comprehensive gradient value that takes into account both the data fitting accuracy target and the fluid physics consistency constraint target. The backpropagation submodule calculates the partial derivative of the comprehensive gradient value with respect to each weight element in the convolution kernel parameter matrix according to the chain rule, and uses the stochastic gradient descent algorithm to adjust the weight values ​​in reverse along the negative gradient direction to generate the convolution kernel parameter matrix after physical correction.

[0011] As a further aspect of the present invention, the generation process of the horizontal velocity gradient component and the vertical velocity gradient component specifically includes: Obtain the coordinates in the convolution kernel parameter matrix. The parameter elements are defined as horizontal velocity components, and the coordinates are obtained. The adjacent parameter elements are used to calculate the difference between them to quantify the horizontal velocity increment, thus obtaining the horizontal velocity gradient component. Obtain the coordinates in the convolution kernel parameter matrix. The parameter elements are defined as vertical velocity components, and the coordinates are obtained. The adjacent parameter elements are used to calculate the difference between them to quantify the velocity increment in the vertical direction, thus obtaining the vertical velocity gradient component. The horizontal velocity gradient component and the vertical velocity gradient component are spatially aligned and combined to construct a gradient tensor that can characterize the flow flux balance state of the virtual flow field within the micro-pixel grid, providing basic vector data for subsequent evaluation of whether the parameter matrix satisfies the fluid divergence-free constraint.

[0012] As a further aspect of the present invention, the process of calculating the correction coefficient by the divergence correction submodule specifically includes: The horizontal velocity gradient component and the vertical velocity gradient component are obtained. Based on the principle of incompressibility in fluid dynamics, the correction coefficient, which characterizes the degree of physical violation of the parameter matrix, is calculated using the following formula: ; in, This represents the correction factor. This represents the size dimension of the convolution kernel parameter matrix. Represents grid points The virtual velocity vector at that location, Represents the divergence operator. Represents the horizontal velocity gradient component. Represents the vertical velocity gradient component, This represents the absolute value operator.

[0013] As a further aspect of the present invention, the process by which the gradient fusion submodule constructs the comprehensive gradient value specifically includes: Obtain the loss value and the correction coefficient, introduce Lagrange multipliers to balance the data-driven objective and the physical constraint objective, construct the overall objective function value to guide parameter updates according to the following formula, and derive the comprehensive gradient value based on this function value: ; in, The overall objective function value is represented by the combined gradient value. Partial derivatives with respect to the kernel parameters, Represents the stated loss value. This represents the diagnostic values ​​indicated by the doctor. Represents the predicted value, Represents the physical constraint weight hyperparameter. This represents the correction coefficient.

[0014] As a further aspect of the present invention, the process by which the backpropagation submodule updates the convolution kernel parameter matrix in reverse using the comprehensive gradient value specifically includes: Obtain the preset learning rate parameters and momentum decay factor, calculate the parameter update step size of the current iteration step based on the comprehensive gradient value, and combine the historical update momentum to smooth the optimized trajectory. The parameter update stride is used to perform a subtraction operation on each weight element in the convolution kernel parameter matrix, and the weight values ​​after physical constraint correction are reassigned to the convolution kernel parameter matrix. The updated parameter matrix is ​​checked to see if it meets the preset convergence condition or reaches the maximum number of iterations. If it does not meet the condition, the updated matrix is ​​used as the initial state for the next iteration. The gradient calculation and parameter update process is repeated until the model performance is stable and the final optimized convolution kernel parameter matrix is ​​generated.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, by mapping the convolution kernel parameter matrix to virtual velocity vector components and introducing a hydrodynamic divergence constraint mechanism, a velocity field divergence model is constructed using the gradient difference between the horizontal and vertical directions to simulate the physical continuity of blood flow. During backpropagation, a correction coefficient is generated by combining the divergence value with the zero-value target deviation and incorporated into the loss function. This forces the neural network to automatically follow the physical law of fluid incompressibility when extracting features, guiding the weight parameters to evolve in a direction that conforms to the laws of hydrodynamics. This ensures that the feature map truly reflects the blood flow dynamics state within the coronary artery, effectively suppresses non-physical image noise interference, enhances the sensitivity to subtle blood flow abnormalities, and improves the physical interpretability and lesion identification accuracy of the model in complex angiography image analysis. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the overall architecture of the cardiovascular disease large-scale cognitive system of this invention. Figure 2 This is a flowchart of the parameter initialization and virtual flow field construction process of the present invention; Figure 3 This is a flowchart of the fluid divergence calculation and physical constraint generation process of this invention; Figure 4 This is a flowchart of the feature extraction and disease probability reasoning process of this invention; Figure 5 This is a flowchart of gradient fusion and parameter inverse correction of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the software-based technical solution is described in detail below with reference to system architecture diagrams and embodiments. It should be understood that the specific embodiments described herein are only for explaining the technical solutions of this invention and do not constitute a limitation on the scope of protection.

[0018] In the description of this invention, the system architecture relationships or data processing flows indicated by terms such as "layer," "module," "interface," "data flow," "client," and "server" are all defined based on the architecture diagram or flowchart corresponding to the embodiments. This way of describing is only used to clearly illustrate the logical relationships between the elements in the technical solution, and not to limit the physical deployment form. The term "multiple" includes two or more technical units, including but not limited to multiple data nodes, processing threads, service instances, or functional components and other scalable elements. The specific number is determined according to the actual business scenario and needs to be specifically specified.

[0019] Please see Figure 1 and Figure 2 This invention provides a technical solution: a large-scale cognitive system for cardiovascular diseases, comprising: The parameter initialization module acquires digital images of coronary angiography, extracts pixel brightness information from the digital images of coronary angiography, constructs a grayscale value matrix, initializes the convolution kernel parameter matrix, and maps the numerical elements in the convolution kernel parameter matrix to virtual velocity vector components. The grayscale value matrix is ​​constructed from the pixel brightness information of the coronary angiography digital image, and the virtual flow velocity vector component is the numerical element in the convolution kernel parameter matrix; The parameter initialization module specifically includes: The image preprocessing submodule acquires digital images of coronary angiography, uses a Gaussian smoothing filter to reduce noise in the images, employs an adaptive grid discretization algorithm to extract pixel brightness information, and normalizes its values ​​to the range of zero to one, constructing a grayscale matrix that can preserve the texture features of microvascular edges. The kernel parameter construction submodule generates initial random weight values ​​that follow a uniform distribution based on the preset number of convolutional neural network layers and convolutional kernel size configuration, and constructs a convolutional kernel parameter matrix for extracting spatial morphological features of the image using the Xavier initialization method. The vector mapping submodule establishes a topological mapping relationship between the numerical elements of the convolution kernel parameter matrix and the virtual hydrodynamic velocity field. It defines the numerical elements in the odd-numbered columns of the matrix as the virtual velocity vector components in the horizontal direction and the numerical elements in the even-numbered columns as the virtual velocity vector components in the vertical direction, thereby generating an initialization parameter set with the potential for physical field evolution.

[0020] The image preprocessing submodule acquires digital images of coronary angiography, reads the raw DICOM file, and parses its metadata header information to determine the image's bit depth and resolution. The system is designed for resolutions of [missing information - likely related to image processing]. A two-dimensional pixel matrix is ​​constructed from the original image data with pixels and a bit depth of 8 bits. The module selected has a size of [size missing]. A Gaussian convolution kernel performs a discrete convolution operation on a matrix. The standard deviation parameter of the Gaussian kernel is... Set to 1.0. During execution, the weighted average of the current pixel and its 24 surrounding pixels is calculated, with the weights determined by a two-dimensional Gaussian function. The calculations are then performed to complete the image smoothing process. Subsequently, the module performs adaptive grid discretization, traversing the denoised pixel matrix and reading the coordinates of each point. Brightness value at the location And perform linear normalization, dividing the brightness value by the maximum bit depth value of 255, that is... This maps the brightness features of all pixels to a floating-point number range of 0 to 1, thus constructing a grayscale matrix.

[0021] The aforementioned Gaussian smoothing filter refers to a linear smoothing filter that calculates weights using a Gaussian function. It is used to blur images to reduce image noise and detail levels. Its core lies in using the probability density function of a normal distribution to generate a convolutional mask.

[0022] Based on the network architecture design, the kernel parameter construction submodule sets the depth of the convolutional layers to 16 layers and the kernel size of the first layer to... Configured as The module calls the Xavier initialization strategy, first calculating the number of input channels for the current convolutional layer. (1 here) and the number of output channels (16 here). Calculate the boundary limit value of the uniform distribution according to the formula. Substituting the above values ​​into the calculation yields... The module is in the interval. The internal function extracts floating-point values ​​using a pseudo-random number generator and fills them to a size of [size missing]. Construct the convolution kernel parameter matrix from the convolution kernel parameter matrix.

[0023] The Xavier initialization strategy mentioned above refers to a weight initialization method designed to maintain the consistency of the variance of the output of each layer of the neural network. It aims to keep the signal during forward propagation and the gradient during backward propagation within a reasonable range, preventing gradient vanishing or exploding.

[0024] The vector mapping submodule iterates through the generated vectors. Convolution kernel parameter matrix. Column indices of the system-defined matrix. (Counting from 1), for the matrix that satisfies (i.e., odd-numbered columns) element Its value is directly assigned to the horizontal velocity component. For satisfying (i.e., even-numbered columns) element Assign its value to the vertical velocity component. For example, for the data in the first row of the matrix , and Mapped to a horizontal velocity component, It is mapped to a vertical velocity component, generating a set of initialization parameters.

[0025] Please see Figure 1 and Figure 3 The divergence constraint module calculates the difference between adjacent row values ​​in the convolution kernel parameter matrix to generate the horizontal velocity gradient component, calculates the difference between adjacent column values ​​in the convolution kernel parameter matrix to generate the vertical velocity gradient component, sums the horizontal velocity gradient component and the vertical velocity gradient component to generate the divergence value, and calculates the absolute value of the deviation between the divergence value and the zero target to generate the correction coefficient. The horizontal velocity gradient component is the difference between adjacent row values ​​in the convolution kernel parameter matrix, the vertical velocity gradient component is the difference between adjacent column values ​​in the convolution kernel parameter matrix, the divergence value is calculated from the horizontal velocity gradient component and the vertical velocity gradient component, and the correction coefficient is determined by the absolute value of the deviation between the divergence value and the zero target. The divergence constraint module specifically includes: The horizontal gradient calculation submodule obtains the virtual velocity vector components, performs discrete difference operations along the row direction of the convolution kernel parameter matrix, calculates the numerical difference between the current grid position element and the adjacent grid element on the right, and generates the horizontal velocity gradient components that characterize the rate of change of the virtual fluid in the horizontal direction. The vertical gradient calculation submodule performs discrete difference operations along the column direction of the convolution kernel parameter matrix, calculates the numerical difference between the element at the current grid position and the adjacent grid element below, and generates the vertical velocity gradient component that characterizes the rate of change of the virtual fluid in the vertical direction. The divergence correction submodule sums the horizontal and vertical velocity gradient components point by point to generate a divergence value that reflects the local fluid source and sink characteristics. It calculates the absolute value of the deviation between the divergence value and the zero target under ideal incompressible fluid conditions and generates a correction coefficient to penalize the distribution of non-physical parameters. The generation process of the horizontal velocity gradient components and the vertical velocity gradient components specifically includes: Obtain the coordinates in the convolution kernel parameter matrix. The parameter elements are defined as horizontal velocity components, and the coordinates are obtained. The adjacent parameter elements are used to calculate the difference between them to quantify the horizontal velocity increment and obtain the horizontal velocity gradient component. Obtain the coordinates in the convolution kernel parameter matrix. The parameter elements are defined as vertical velocity components, and the coordinates are obtained. The adjacent parameter elements are used to calculate the difference between them to quantify the vertical velocity increment and obtain the vertical velocity gradient component. By spatially aligning and combining the horizontal and vertical velocity gradient components, a gradient tensor is constructed that can characterize the flow flux balance state of the virtual flow field within the micro-pixel grid, providing basic vector data for subsequent evaluation of whether the parameter matrix satisfies the fluid divergence-free constraint. The process of calculating the correction coefficients in the divergence correction submodule specifically includes: Obtain the horizontal and vertical velocity gradient components. Based on the incompressibility principle in fluid dynamics, calculate the correction coefficients that characterize the physical violation of the parameter matrix using the following formula: ; in, Represents the correction factor. This represents the size dimension of the convolution kernel parameter matrix. Represents grid points The virtual velocity vector at that location, Represents the divergence operator. Represents the horizontal velocity gradient component. Represents the vertical velocity gradient component. This represents the absolute value operator.

[0026] The horizontal gradient calculation submodule reads the virtual velocity vector components corresponding to the convolution kernel parameter matrix. The system sets the boundary padding strategy to zero padding to handle the difference operations at the matrix edges. This applies to arbitrary coordinate points within the matrix. The module extracts the parameter elements at that location as the current horizontal velocity component. Extract the right adjacent positions in the same row. parameter elements The module performs forward difference operations, and the calculation formula is as follows: This generates horizontal velocity gradient components.

[0027] The vertical gradient calculation submodule synchronously performs the difference operation along the column direction for the same coordinate point. The module extracts the position parameter element as the vertical velocity component. Extract adjacent positions below in the same column. parameter elements The module performs calculations. This generates vertical velocity gradient components.

[0028] The aforementioned forward difference operation refers to a numerical differentiation method in a discrete grid that approximates the derivative by using the difference between the function value of the current node and its next adjacent node. Here, it is used to quantify the spatial rate of change of velocity components in a virtual flow field.

[0029] The divergence correction submodule spatially aligns and algebraically sums the two components mentioned above. For each grid point... Calculate the local divergence value Subsequently, the module calculates the correction coefficients used for the constraint weights based on the principles of incompressible fluid dynamics. The calculation process involves traversing all convolution kernel matrices. For each element, take the absolute value of the local divergence value at each location. And sum all the absolute values ​​and divide by the total number of elements. Based on the principle of incompressibility in fluid dynamics, the correction factor is calculated using the following formula: ; in, Represents the correction coefficient, used to quantify the overall degree to which the parameter matrix violates the physical divergence-free condition; Represents the side length dimension of the convolution kernel parameter matrix; and These represent the row and column indices of the convolution kernel parameter matrix, respectively. Represents grid points The virtual velocity vector at the location; Represents the divergence operator, used to calculate the degree of divergence of a vector field; Represents the horizontal velocity gradient component; Represents the vertical velocity gradient component; Represents the absolute value operator. This represents the summation operator.

[0030] Table 1 lists a Example numerical values ​​of the convolution kernel parameter matrix and the process of calculating its divergence.

[0031] ; As shown in Table 1, the sum of the absolute values ​​of the local divergence of all grid points is calculated as follows: Calculate the correction factor according to the formula: The result indicates that the current convolution kernel parameter matrix has an average deviation of 0.444 units in terms of physical incompressibility, which will be used as a regularization term in subsequent gradient calculations.

[0032] Please see Figure 1 and Figure 4 The feature inference module calls a convolutional neural network and uses the convolutional kernel parameter matrix to perform weighted summation and sliding window operations on the gray-level numerical matrix to generate feature map data. The feature map data is then input to a fully connected classification layer for nonlinear mapping to output predicted values. The feature map data is the result of weighted summation of the convolution kernel parameter matrix and the gray-level numerical matrix, as well as the result of sliding window operation. The predicted values ​​are the nonlinear mapping output of the fully connected classification layer. The feature reasoning module specifically includes: The convolution operation submodule obtains the gray-level value matrix and the convolution kernel parameter matrix, moves the sliding window on the gray-level value matrix according to the preset step size, performs Hadamard product operation on the pixel brightness values ​​in the window and the corresponding convolution kernel weights and sums them to generate feature map data that extracts local vascular stenosis or occlusion features. The classification mapping submodule flattens the feature map data into a one-dimensional feature vector, inputs it into a multi-layer fully connected neural network, maps the high-dimensional features to the disease category space through matrix multiplication and bias addition operations, and processes the mapping results using the Softmax activation function to output a predicted value representing the probability distribution of cardiovascular disease.

[0033] The convolution operation submodule obtains the preprocessed grayscale value matrix and the convolution kernel parameter matrix under physical constraints. It also sets the sliding window stride. No padding mode. The module overlays the convolution kernel onto the top-left corner of the grayscale matrix. to For the region, perform the Hadamard product operation, that is, multiply the corresponding elements: After completing one calculation, the window moves one step to the right until the entire row has been traversed, then it starts scanning a new row. For example, for a local region of the grayscale matrix... (Hereinafter referred to as) (Simplified diagram of the region) and convolution kernels of corresponding sizes The operation yields the following result: This process generates feature map data.

[0034] The Hadamard product operation mentioned above refers to the operation of multiplying corresponding elements of two matrices of the same order to obtain a new matrix. Here, it is used to calculate the weighted correlation between pixel values ​​in a local region of an image and the weights of the convolution kernel.

[0035] The classification mapping submodule receives the feature map data output after multiple convolutional and pooling operations. Assume the final feature map size is... With 64 channels, the module first performs a Flatten operation, flattening it to a length of... One-dimensional feature vector This vector is input to the fully connected layer, along with the weight matrix. (size Perform matrix multiplication and add the bias vector. (size ), to obtain the logistic regression value vector Finally, the Softmax function is used. Convert the output into a probability distribution. For example, if the output vector of a fully connected layer is... Where index 0 represents "illness" and index 1 represents "health". Calculate the index term: , Calculate the sum of the denominators: Calculate the predicted value: probability of disease. Health probability The module ultimately outputs predicted values ​​[0.832, 0.168].

[0036] The Softmax function mentioned above is a normalization exponential function that can compress a K-dimensional vector containing arbitrary real numbers into another K-dimensional vector, such that each element is in the range (0,1) and the sum of all elements is 1. It is often used for probability output in multi-class classification problems.

[0037] Please see Figure 1 and Figure 5 The weight correction module compares the predicted values ​​with the doctor's labeled diagnosis values ​​to calculate the loss value, and weights and combines the correction coefficients and loss values ​​to construct a comprehensive gradient value. The comprehensive gradient value is then used to inversely update the convolution kernel parameter matrix. The loss value is calculated from the predicted value and the doctor's labeled diagnosis value. The comprehensive gradient value is composed of the correction coefficient and the loss value. The convolution kernel parameter matrix is ​​the set of parameters updated in reverse using the comprehensive gradient value. The weight correction module specifically includes: The loss evaluation submodule obtains the predicted values ​​output by the model and the diagnostic values ​​labeled by doctors in advance by clinical experts. It uses the cross-entropy loss function to calculate the difference in probability distribution between the two and generates a loss value that can quantify the model's classification prediction error. The gradient fusion submodule introduces a physical constraint regularization weight factor, which performs a weighted linear combination of the correction coefficient and the loss value to construct a comprehensive gradient value that takes into account both the data fitting accuracy target and the fluid physics consistency constraint target. The backpropagation submodule calculates the partial derivative of the overall gradient value with respect to each weight element in the convolution kernel parameter matrix according to the chain rule, and uses the stochastic gradient descent algorithm to adjust the weight values ​​in reverse along the negative gradient direction to generate the convolution kernel parameter matrix after physical correction. The process of constructing the comprehensive gradient value by the gradient fusion submodule specifically includes: Obtain the loss value and correction coefficient, introduce Lagrange multipliers to balance the data-driven objective and the physical constraint objective, construct the overall objective function value to guide parameter updates according to the following formula, and derive the comprehensive gradient value based on this function value: ; in, The total objective function value is represented by the combined gradient value. Partial derivatives with respect to the kernel parameters, Represents the value of the loss. The doctor marked the diagnostic values. Represents the predicted value. Represents the physical constraint weight hyperparameter. Represents the correction factor; The backpropagation submodule updates the convolution kernel parameter matrix in reverse using the combined gradient values, specifically including: Obtain the preset learning rate parameters and momentum decay factor, calculate the parameter update step size of the current iteration step based on the comprehensive gradient numerical calculation, and combine the historical update momentum to smooth the optimized trajectory. The parameter update stride is used to subtract each weight element in the convolution kernel parameter matrix, and the weight values ​​after physical constraint correction are reassigned to the convolution kernel parameter matrix. The updated parameter matrix is ​​checked to see if it meets the preset convergence condition or reaches the maximum number of iterations. If it does not meet the condition, the updated matrix is ​​used as the initial state for the next iteration. The gradient calculation and parameter update process is repeated until the model performance is stable and the final optimized convolution kernel parameter matrix is ​​generated.

[0038] The loss evaluation submodule obtains the predicted values ​​output by the feature inference module. The corresponding doctor's diagnostic value is displayed. The doctor's value is represented by a one-hot code; if the diagnosis is "illness," then the value is displayed. The module uses the cross-entropy loss function to calculate... The formula is .in, The total number of categories representing cardiovascular disease diagnostic tasks (in this embodiment) (corresponding to the two categories of "illness" and "health"). Represents the category index variable (value range is...) to ), Representing the The actual labeled values ​​corresponding to each category (using One-hot encoding, the actual category value is taken as...) The values ​​for the other categories are: ), The representative model output after the Softmax activation function is the first The predicted probability values ​​for each category. Substitute the above values ​​into the calculation: This value represents the loss.

[0039] The cross-entropy loss function mentioned above is a function used to measure the difference between two probability distributions. In machine learning, it is often used as a loss function to evaluate how close the probability distribution predicted by the model is to the probability distribution of the true label.

[0040] The gradient fusion submodule is based on the formula Construct the overall objective function. In this formula, This refers to the cross-entropy loss value calculated above; These are the correction coefficients calculated by the divergence constraint module; This is a physical constraint weight hyperparameter used to balance data fitting accuracy with physical consistency. The value of is determined based on experimental verification. A grid search is performed within the range of 0.001 to 1.0 to select the value with the highest F1-score on the validation set. This embodiment sets... The aforementioned calculation results and Substitute into the formula: Subsequently, the module performs calculations. Regarding each weight element in the convolution kernel parameter matrix partial derivatives Generate a comprehensive gradient value.

[0041] The backpropagation submodule updates parameters based on the generated combined gradient values. The learning rate parameter is set. Momentum decay factor For weights If the calculated comprehensive gradient value is And the momentum update of the previous moment First, calculate the current momentum: Then perform a weight update: Assuming the original weights The updated weights are The system checks whether the convergence condition is met: it determines whether the decrease in the validation set loss value over 5 consecutive epochs is less than... The system checks whether the number of iterations has reached 5000. If not, the updated parameter matrix is ​​returned to the feature inference module to start the next round of training until the final optimized convolutional kernel parameter matrix is ​​generated.

[0042] The above embodiments illustrate preferred embodiments of the present invention. Any equivalent adjustments to the technical solution based on software engineering methods are within the scope of protection, including but not limited to: implementing algorithm logic using different programming languages, refactoring functional modules into services, adjusting data interaction protocols, and optimizing resource scheduling strategies. Any implementation scheme derived from reasonable modifications to the data processing flow, service call chain, or system architecture layer without departing from the core technology of the present invention should be considered within the scope of protection defined by the claims of the present invention.

Claims

1. A large-scale cognitive system for cardiovascular diseases, characterized in that, The system includes: The parameter initialization module acquires a digital image of coronary angiography, extracts the pixel brightness information of the digital image of coronary angiography, constructs a grayscale value matrix, initializes the convolution kernel parameter matrix, and maps the numerical elements in the convolution kernel parameter matrix to virtual velocity vector components. The divergence constraint module calculates the difference between adjacent row values ​​in the convolution kernel parameter matrix to generate a horizontal velocity gradient component, calculates the difference between adjacent column values ​​in the convolution kernel parameter matrix to generate a vertical velocity gradient component, sums the horizontal velocity gradient component and the vertical velocity gradient component to generate a divergence value, and calculates the absolute value of the deviation between the divergence value and the zero target to generate a correction coefficient. The feature inference module calls a convolutional neural network, uses the convolutional kernel parameter matrix to perform weighted summation and sliding window operations on the gray-level value matrix to generate feature map data, and inputs the feature map data to a fully connected classification layer for nonlinear mapping to output predicted values. The weight correction module compares the predicted value with the doctor's labeled diagnosis value to calculate the loss value, and weights and combines the correction coefficient with the loss value to construct a comprehensive gradient value. The comprehensive gradient value is then used to inversely update the convolution kernel parameter matrix.

2. The cardiovascular disease large-scale model cognitive system according to claim 1, characterized in that, The grayscale value matrix is ​​constructed from the pixel brightness information of the coronary angiography digital image. The virtual velocity vector component is a numerical element in the convolution kernel parameter matrix. The horizontal velocity gradient component is the difference between adjacent row values ​​in the convolution kernel parameter matrix. The vertical velocity gradient component is the difference between adjacent column values ​​in the convolution kernel parameter matrix. The divergence value is calculated from the horizontal and vertical velocity gradient components. The correction coefficient is determined by the absolute value of the deviation between the divergence value and the zero-value target. The feature map data is the weighted sum of the convolution kernel parameter matrix and the grayscale value matrix, as well as the result of sliding window operation. The predicted value is the nonlinear mapping output of the fully connected classification layer. The loss value is calculated from the predicted value and the doctor's labeled diagnosis value. The comprehensive gradient value is a combination of the correction coefficient and the loss value. The convolution kernel parameter matrix is ​​a set of parameters updated inversely using the comprehensive gradient value.

3. The cardiovascular disease large-scale model cognitive system according to claim 2, characterized in that, The parameter initialization module specifically includes: The image preprocessing submodule acquires the digital image of the coronary angiography, performs noise reduction processing on the image using a Gaussian smoothing filter, extracts the brightness information of the pixels using an adaptive grid discretization algorithm, and normalizes its value to the range of zero to one, thereby constructing the gray value matrix that can preserve the texture features of the microvascular edges. The kernel parameter construction submodule generates initial random weight values ​​that follow a uniform distribution based on the preset number of convolutional neural network layers and convolutional kernel size configuration, and constructs the convolutional kernel parameter matrix used to extract spatial morphological features of the image. The vector mapping submodule establishes a topological mapping relationship between the numerical elements of the convolution kernel parameter matrix and the virtual hydrodynamic velocity field. It defines the numerical elements in the odd-numbered columns of the matrix as the virtual velocity vector components in the horizontal direction and the numerical elements in the even-numbered columns of the matrix as the virtual velocity vector components in the vertical direction, thereby generating an initialization parameter set with the potential for physical field evolution.

4. The cardiovascular disease large-scale model cognitive system according to claim 3, characterized in that, The divergence constraint module specifically includes: The horizontal gradient calculation submodule obtains the virtual velocity vector component, performs discrete difference operation along the row direction of the convolution kernel parameter matrix, calculates the numerical difference between the current grid position element and the adjacent grid element on the right, and generates the horizontal velocity gradient component that characterizes the rate of change of the virtual fluid in the horizontal direction. The vertical gradient calculation submodule performs discrete difference operations along the column direction of the convolution kernel parameter matrix to calculate the numerical difference between the current grid position element and the adjacent grid element below, and generates the vertical velocity gradient component that characterizes the rate of change of the virtual fluid in the vertical direction. The divergence correction submodule performs point-by-point superposition and summation of the horizontal velocity gradient component and the vertical velocity gradient component to generate the divergence value reflecting the local fluid source and sink characteristics. It then calculates the absolute value of the deviation between the divergence value and the zero target under ideal incompressible fluid conditions and generates the correction coefficient used to penalize the distribution of non-physical parameters.

5. The cardiovascular disease large-scale model cognitive system according to claim 4, characterized in that, The feature reasoning module specifically includes: The convolution operation submodule obtains the gray value matrix and the convolution kernel parameter matrix, moves a sliding window on the gray value matrix according to a preset step size, performs a Hadamard product operation on the pixel brightness values ​​in the window and the corresponding convolution kernel weights and sums them to generate the feature map data that extracts the features of local vascular stenosis or occlusion. The classification mapping submodule flattens the feature map data into a one-dimensional feature vector, inputs it into a multi-layer fully connected neural network, maps the high-dimensional features to the disease category space through matrix multiplication and bias addition operations, processes the mapping result using the Softmax activation function, and outputs the predicted value representing the probability distribution of cardiovascular disease.

6. The cardiovascular disease large-scale model cognitive system according to claim 5, characterized in that, The weight correction module specifically includes: The loss evaluation submodule obtains the predicted value output by the model and the doctor-annotated diagnosis value pre-labeled by clinical experts, calculates the probability distribution difference between the two using the cross-entropy loss function, and generates the loss value that can quantify the model's classification prediction error. The gradient fusion submodule introduces a physical constraint regularization weight factor to perform a weighted linear combination of the correction coefficient and the loss value, thereby constructing the comprehensive gradient value that takes into account both the data fitting accuracy target and the fluid physics consistency constraint target. The backpropagation submodule calculates the partial derivative of the comprehensive gradient value with respect to each weight element in the convolution kernel parameter matrix according to the chain rule, and uses the stochastic gradient descent algorithm to adjust the weight values ​​in reverse along the negative gradient direction to generate the convolution kernel parameter matrix after physical correction.

7. The cardiovascular disease large-scale model cognitive system according to claim 6, characterized in that, The generation process of the horizontal velocity gradient component and the vertical velocity gradient component specifically includes: Obtain the coordinates in the convolution kernel parameter matrix. The parameter elements are defined as horizontal velocity components, and the coordinates are obtained. The adjacent parameter elements are used to calculate the difference between them to quantify the horizontal velocity increment, thus obtaining the horizontal velocity gradient component. Obtain the coordinates in the convolution kernel parameter matrix. The parameter elements are defined as vertical velocity components, and the coordinates are obtained. The adjacent parameter elements are used to calculate the difference between them to quantify the velocity increment in the vertical direction, thus obtaining the vertical velocity gradient component. The horizontal velocity gradient component and the vertical velocity gradient component are spatially aligned and combined to construct a gradient tensor that can characterize the flow flux balance state of the virtual flow field within the micro-pixel grid, providing basic vector data for subsequent evaluation of whether the parameter matrix satisfies the fluid divergence-free constraint.

8. The cardiovascular disease large-scale model cognitive system according to claim 7, characterized in that, The process by which the divergence correction submodule calculates the correction coefficient specifically includes: The horizontal velocity gradient component and the vertical velocity gradient component are obtained. Based on the principle of incompressibility in fluid dynamics, the correction coefficient, which characterizes the degree of physical violation of the parameter matrix, is calculated using the following formula: ； in, This represents the correction factor. This represents the size dimension of the convolution kernel parameter matrix. Represents grid points The virtual velocity vector at that location, Represents the divergence operator. Represents the horizontal velocity gradient component, Represents the vertical velocity gradient component, This represents the absolute value operator.

9. The cardiovascular disease large-scale model cognitive system according to claim 8, characterized in that, The process by which the gradient fusion submodule constructs the comprehensive gradient value specifically includes: Obtain the loss value and the correction coefficient, introduce Lagrange multipliers to balance the data-driven objective and the physical constraint objective, construct the overall objective function value to guide parameter updates according to the following formula, and derive the comprehensive gradient value based on this function value: ； in, The overall objective function value is represented by the combined gradient value. Partial derivatives with respect to the kernel parameters, Represents the stated loss value. This represents the diagnostic values ​​indicated by the doctor. Represents the predicted value, Represents the physical constraint weight hyperparameter. This represents the correction coefficient.

10. The cardiovascular disease large-scale model cognitive system according to claim 9, characterized in that, The process by which the backpropagation submodule updates the convolution kernel parameter matrix in reverse using the combined gradient values ​​specifically includes: Obtain the preset learning rate parameters and momentum decay factor, calculate the parameter update step size of the current iteration step based on the comprehensive gradient value, and combine the historical update momentum to smooth the optimized trajectory. The parameter update stride is used to perform a subtraction operation on each weight element in the convolution kernel parameter matrix, and the weight values ​​after physical constraint correction are reassigned to the convolution kernel parameter matrix. The updated parameter matrix is ​​checked to see if it meets the preset convergence condition or reaches the maximum number of iterations. If it does not meet the condition, the updated matrix is ​​used as the initial state for the next iteration. The gradient calculation and parameter update process is repeated until the model performance is stable and the final optimized convolution kernel parameter matrix is ​​generated.

Images