Medical image disease course prediction system based on industrial neural network

Abstract

The invention relates to the technical field of medical image intelligent analysis and artificial intelligence auxiliary diagnosis, in particular to a medical image disease course prediction system based on an industrial neural network, and the system comprises a reference generation module which is used for obtaining static image data; processing the static image data by using a physical perception neural network to generate a pure ideal state reference; a perturbation simulation module; the industrial kinetic parameters are used as perturbation terms to be superposed to a pure ideal state reference, and a theoretical damaged state is generated; the projection verification module is used for acquiring real multi-modal observation data; generating a real residual error; generating a theoretical residual error based on the theoretical damaged state and the pure ideal state reference; calculating a manifold coupling confidence coefficient; the closed-loop correction module is used for performing inversion optimization on the industrial kinetic parameters; outputting a disease course prediction result according to the manifold coupling confidence coefficient; according to the method, the problem that a traditional medical model lacks physical consistency explanation is solved, and the credibility of artificial intelligence auxiliary diagnosis is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical image analysis and artificial intelligence-assisted diagnosis technology, specifically a medical image disease progression prediction system based on industrial neural networks. Background Technology

[0002] In the current field of intelligent medical image analysis, disease prediction and auxiliary diagnosis based on static image data such as CT or MRI mainly rely on deep learning models. These systems are usually trained by mining statistical patterns from massive amounts of historical data. However, existing solutions generally adopt a purely data-driven black-box paradigm, lacking constraint mechanisms based on human physiological mechanisms or physical laws, resulting in a lack of clear interpretability in model output. Due to the absence of an ideal physical field that strictly satisfies physiological conservation laws as a benchmark reference, existing models struggle to effectively decouple real pathological evolution signals from machine imaging artifacts, random measurement noise, or equipment electromagnetic interference during feature extraction. This can easily lead to misjudging non-pathological image fluctuations as lesion development, failing to guarantee the physical consistency of diagnostic logic. Therefore, how to introduce physical mechanism constraints to construct an interpretable deduction architecture to accurately distinguish between non-pathological artifacts and real lesion evolution, thereby improving the physical consistency and accuracy of auxiliary diagnostic results, has become an urgent technical problem to be solved. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a medical image disease progression prediction system based on industrial neural networks. Specifically, the technical solution of this invention includes: The benchmark generation module is used to acquire static image data; it processes the static image data using a physical sensing neural network to generate a pure ideal benchmark. The perturbation simulation module is used to obtain industrial dynamic parameters; the industrial dynamic parameters are superimposed as perturbation terms onto the pure ideal state reference to generate the theoretical damaged state; The projection verification module is used to acquire real-world multimodal observation data; generate real-world residuals based on the real-world multimodal observation data and the pure ideal state benchmark; generate theoretical residuals based on the theoretical damaged state and the pure ideal state benchmark; and calculate manifold coupling confidence based on the real-world residuals and the theoretical residuals. The closed-loop correction module is used to perform inversion optimization of industrial dynamic parameters based on the difference between the actual residual and the theoretical residual in the latent space; and outputs the disease course prediction results based on the manifold coupling confidence.

[0004] Preferably, the benchmark generation module uses a physical sensing neural network to process static image data and generate a pure, ideal benchmark, including: Determine the set of boundary points of the anatomical structure; Determine the set of points for the internal regions of anatomical structures; At the set of boundary points, the boundary conditions output by the physical sensing neural network are forced to match the static image data; At the set of points in the internal region, the output of the physical perception neural network is forced to satisfy the preset physical control equation; Construct a weighted sum of boundary fitting error and physical equation residuals; Minimize the weighted sum until the output of the physical perception neural network converges, generating a pure ideal state benchmark.

[0005] Preferably, the perturbation simulation module superimposes industrial dynamic parameters as perturbation terms onto a pure ideal state reference to generate a theoretically damaged state, including: Extracting the spontaneous diffusion term coefficient from industrial kinetic parameters; Extracting external disturbance coefficients from industrial dynamic parameters; Based on the spontaneous diffusion term coefficient, calculate the diffusion increment of the theoretically damaged state at the current moment; The source generation increment is calculated based on the external disturbance term coefficient and the pure ideal state benchmark. The diffusion increment and the source generation increment are superimposed on the theoretically damaged state at the current moment to generate the theoretically damaged state at the next moment.

[0006] Preferably, the projection verification module generates real residuals based on real multimodal observation data and a pure ideal state benchmark, and generates theoretical residuals based on theoretical damaged states and a pure ideal state benchmark, including: Calculate the difference between real-world multimodal observation data and a pure ideal-state baseline to generate real-world residuals; The difference between the theoretically damaged state and the pure ideal state baseline is calculated to generate the theoretical residual.

[0007] Preferably, the projection verification module calculates the manifold coupling confidence level based on the actual residual and the theoretical residual, including: Invoke the encoder that includes layer normalization operations; The encoder is used to encode the real residuals to generate real latent space feature vectors; The theoretical residual is encoded using an encoder to generate a feature vector in the theoretical latent space; Calculate the inner product of the feature vectors in the real latent space and the feature vectors in the theoretical latent space; Based on the inner product, the magnitude of the feature vectors in the real latent space, the magnitude of the feature vectors in the theoretical latent space, and a preset distance sensitivity coefficient, a manifold coupling confidence score is generated.

[0008] Preferably, the closed-loop correction module outputs disease course prediction results based on the manifold coupling confidence level, including: Obtain the adaptive decision threshold; If the manifold coupling confidence is greater than the adaptive decision threshold, the actual residual is determined to contain true pathological signals, and the true pathological signals are retained as the disease course prediction result. If the manifold coupling confidence level is less than or equal to the adaptive decision threshold, the actual residual is determined to be a non-pathological artifact, and a filtering removal operation is performed.

[0009] Preferably, the closed-loop correction module performs inversion optimization of industrial dynamic parameters based on the difference between the actual residual and the theoretical residual in the latent space, including: Calculate the Euclidean distance error between the feature vectors of the real latent space and the feature vectors of the theoretical latent space; Calculate the gradient of the Euclidean distance error with respect to the industrial dynamic parameters; The industrial dynamics parameters are updated using gradients and a preset learning rate.

[0010] Preferably, the closed-loop correction module is also used for: When the Euclidean distance error converges, output the industrial dynamic parameters; Industrial kinetic parameters are used as quantified, personalized pathological parameters.

[0011] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention solves the problem of lack of physically consistent interpretation in traditional medical models by introducing industrial physical constraints and dynamic simulation mechanisms. Unlike the pure data-driven black box paradigm, this system uses a physical perception neural network to generate a pure ideal state benchmark that strictly follows the physiological conservation law. This design ensures that the disease course prediction logic conforms to physical laws, makes the diagnostic process transparent and traceable, and significantly improves the credibility of artificial intelligence-assisted diagnosis. 2. This invention utilizes a projection verification module to solve the technical challenge of distinguishing between machine artifacts and real lesions. By comparing the residuals of real-world observation data and theoretically derived data, and calculating the manifold coupling confidence level, the system can identify the consistency of signals in manifold geometry. This mechanism acts as a semantic filter, effectively eliminating non-pathological noise generated by respiratory motion, equipment interference, etc., retaining only true pathological signals that conform to physical laws, thus reducing the false alarm rate. 3. This invention transforms medical diagnosis into an industrial parameter inversion problem through a closed-loop correction module. The system can not only predict lesion development but also optimize convergent industrial dynamic parameters through gradient descent. These parameters can serve as quantified personalized pathological indicators, providing doctors with a second dimension of quantitative information in addition to morphological images, and can more sensitively reflect the severity of early pathological changes. 4. This invention employs a zero-sample solution strategy, avoiding reliance on massive amounts of historical statistical data. The system directly solves the steady-state solution of the physical control equation for the current patient's static images, generating a pure ideal-state benchmark that eliminates imaging noise. This benchmark not only serves as a zero-level reference system but also simulates pathological evolution through perturbation simulation, providing a strictly standard answer based on individual anatomical structure for subsequent signal separation, ensuring the individual specificity and accuracy of the analysis results. Attached Figure Description

[0012] The present invention will be further explained below with reference to the accompanying drawings and embodiments: Figure 1 This is a structural diagram of the system of the present invention. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0014] Example 1: Please see Figure 1 A medical image disease progression prediction system based on industrial neural networks includes: The benchmark generation module is used to acquire static image data; it processes the static image data using a physical sensing neural network to generate a pure ideal benchmark. The perturbation simulation module is used to obtain industrial dynamic parameters; the industrial dynamic parameters are superimposed as perturbation terms onto the pure ideal state reference to generate the theoretical damaged state; The projection verification module is used to acquire real-world multimodal observation data; generate real-world residuals based on the real-world multimodal observation data and the pure ideal state benchmark; generate theoretical residuals based on the theoretical damaged state and the pure ideal state benchmark; and calculate manifold coupling confidence based on the real-world residuals and the theoretical residuals. The closed-loop correction module is used to perform inversion optimization of industrial dynamic parameters based on the difference between the actual residual and the theoretical residual in the latent space; and outputs the disease course prediction results based on the manifold coupling confidence.

[0015] This embodiment discloses a medical image disease progression prediction system based on industrial neural networks. This system constructs an interpretable disease progression deduction architecture by introducing industrial physical constraints and dynamic simulation mechanisms, aiming to solve the technical problems of traditional medical artificial intelligence models lacking physically consistent interpretation and being unable to effectively distinguish between machine artifacts and the evolution of real lesions. The system mainly includes four collaborative core processing modules, and the connection relationship and specific functions of each module are as follows. It should be specifically noted that the terms industrial neural networks, industrial dynamic parameters, and industrial physical constraints used in this specification are not intended to limit the application of this invention to the industrial manufacturing field, but rather refer to the fact that this system uses general physical field control equations, such as reaction-diffusion equations and fluid dynamics equations, which are commonly used in industrial simulations, to construct the neural network architecture, emphasizing its physical interpretability and mechanistic deduction characteristics, which distinguish it from purely data-driven models.

[0016] The system is equipped with a benchmark generation module; this module is configured to acquire the patient's static image data, which can be single-frame slices from computed tomography (CT) or magnetic resonance imaging (MRI); the benchmark generation module processes this static image data using a physical perception neural network (PINN), and generates a pure ideal benchmark by solving the steady-state solution of a preset physical control equation. This benchmark It represents the theoretical physical field distribution of the patient's anatomical structure under the condition that it strictly satisfies the physiological conservation law and there is no pathological interference, and serves as a zero-level reference frame for subsequent analysis; The system is equipped with a perturbation simulation module; this module is used to simulate the dynamic interference of pathological mechanisms on an ideal physiological environment; specifically, it is configured to acquire industrial dynamic parameters. The industrial dynamic parameters are then superimposed as perturbation terms onto the aforementioned pure ideal state reference. Above, using discretized dynamic evolution equations, theoretically damaged states that evolve over time are generated. Here, the theoretically damaged state It provides a standard answer based on mechanism deduction; The system is equipped with a projection verification module; this module is designed to identify real disease progression signals through multi-source data comparison; it is configured to acquire real-world multimodal observation data collected by sensors. Based on real-world multimodal observation data Compared with the pure ideal state benchmark Generate real residual Simultaneously based on the theoretical damaged state Compared with the pure ideal state benchmark Generate theoretical residuals This module is based on real-world residuals. Residuals from Theory Calculate the manifold coupling confidence score based on the consistency of the feature space. ; The system is equipped with a closed-loop correction module; this module is configured to use the gradient descent method to correct industrial dynamic parameters based on the difference between the actual and theoretical residuals in the latent space. Perform inversion optimization; and based on manifold coupling confidence. The signal is filtered, and the denoised disease progression prediction result is output. Through the above settings, this embodiment transforms medical diagnosis into an industrial parameter inversion problem, which can not only predict lesion development but also utilize the parameters obtained from the inversion. Quantifying the degree of pathology significantly improves the interpretability of auxiliary diagnoses.

[0017] Example 2: The benchmark generation module uses a physical sensing neural network to process static image data and generate a pure, ideal benchmark, including: Determine the set of boundary points of the anatomical structure; Determine the set of points for the internal regions of anatomical structures; At the set of boundary points, the boundary conditions output by the physical sensing neural network are forced to match the static image data; At the set of points in the internal region, the output of the physical perception neural network is forced to satisfy the preset physical control equation; Construct a weighted sum of boundary fitting error and physical equation residuals; Minimize the weighted sum until the output of the physical perception neural network converges, generating a pure ideal state benchmark.

[0018] In this embodiment, the benchmark generation module uses a physical sensing neural network to generate a pure ideal state benchmark. The process is designed as a zero-sample solution strategy, which does not rely on the statistical regularity of historical big data, but directly solves the physical steady state of the current anatomical structure; Specifically, this module determines the set of boundary points of anatomical structures from the input static image data using an image segmentation algorithm. For example, the walls of blood vessels or the outer surfaces of organs; simultaneously determining the set of collocation points for internal regions of anatomical structures. For example, the fluid region inside a blood vessel or the interstitial region; the boundary point set is used to provide Dirichlet boundary conditions, while the collocation set of the internal region is used to randomly sample within the anatomical structure to test the residuals of the physical control equations; At the set of boundary points, this module forces the boundary conditions output by the physical sensing neural network to match the observations extracted from the static image data. At the set of points in the internal region, the output of the physical sensing neural network is forced to satisfy the preset physical control equations. ;in It is a nonlinear differential operator, specifically a reaction-diffusion equation operator, and its expression is:

[0019] in, This represents the normalized physical field strength of the organization. Where is the diffusion coefficient. It is the reaction rate coefficient; This is a preset physiological constant; this physiological constant It is a standard reference value obtained by matching from a pre-set medical human tissue physical attribute database based on the anatomical location corresponding to the image to be tested, such as liver, lung or brain tissue; To train the network, the module constructs a physical consistency loss function that includes boundary fitting error and physical equation residuals. The calculation formula is as follows:

[0020] in, Represents the position of the neural network The output field at that point, i.e., the ideal reference to be generated; The known values ​​of the static image at the boundary; and These represent the number of samples for boundary points and internal collocation points, respectively. and These are the boundary term weights and the physical residual term weights, respectively. These two weights are not fixed values, but rather derived from adaptive balance coefficients calculated during training using a gradient norm balancing algorithm. Specifically, in each iteration, the boundary loss gradient is calculated. Gradient of physical residual loss The average modulus length is calculated using the following formula:

[0021] in, To prevent smoothing terms with a denominator of zero, the value is taken as... At the same time, let To dynamically balance the contribution weights of the two in backpropagation, so as to ensure the synchronous convergence of boundary constraints and physical constraints during gradient descent; By minimizing this weighted sum Until the output of the physical sensing neural network converges, the generated That is, the pure ideal state benchmark. This process eliminates imaging noise and artifacts in the original image that do not conform to the laws of physics, resulting in a digital twin that strictly adheres to the laws of physiological conservation.

[0022] Example 3: The perturbation simulation module superimposes industrial dynamic parameters as perturbation terms onto a pure ideal state baseline to generate theoretically damaged states, including: Extracting the spontaneous diffusion term coefficient from industrial kinetic parameters; Extracting external disturbance coefficients from industrial dynamic parameters; Based on the spontaneous diffusion term coefficient, calculate the diffusion increment of the theoretically damaged state at the current moment; The source generation increment is calculated based on the external disturbance term coefficient and the pure ideal state benchmark. The diffusion increment and the source generation increment are superimposed on the theoretically damaged state at the current moment to generate the theoretically damaged state at the next moment.

[0023] In this embodiment, the perturbation simulation module uses industrial dynamic parameters as perturbation terms to superimpose onto a pure ideal state reference to generate a theoretically damaged state. This process employs perturbation theory in nonlinear dynamics to parameterize complex pathological mechanisms. The module extracts industrial kinetic parameters from a medical-industrial parameter mapping library. Two types of key coefficients: spontaneous diffusion term coefficient Its physical meaning corresponds to the lesion infiltration capacity or tissue permeability in medicine; and the coefficient of the external interference term. Its physical meaning corresponds to the lesion growth rate or abnormal metabolic rate in medicine; The module uses the discretized pathological evolution state equation to calculate the state at the next time step; it initializes the initial conditions of the evolution equation, letting... Theoretical damaged state at time Equal to the pure ideal state benchmark output by the benchmark generation module This calculation process is based on the current time. State projection for the next moment The state is given by the following formula:

[0024] in, This represents the theoretically damaged state at the current moment; This is the Laplace operator, used to calculate the diffusion increment; For benchmark The reaction function represents the dependence of lesions on the formation of basic anatomical structures; specifically... Defined as a logistic growth function constrained by a baseline tissue density:

[0025] in, As the tissue's maximum carrying capacity constant, this function ensures that the amount of lesion formation is consistent with the baseline anatomy. Its existence is directly proportional to its size and it will not grow indefinitely; for example, tumors depend on blood vessels for blood supply. This is the preset simulation time step; The specific calculation steps include: based on the spontaneous diffusion term coefficient Calculate the diffusion increment of the theoretically damaged state at the current moment. This description refers to the physical spread of the lesion to its surroundings based on its current morphology; it is based on the coefficient of the external interference term. and pure ideal state benchmark Calculate the source to generate increment This describes the rate of lesion formation; the diffusion increment and the source formation increment are superimposed onto the theoretically damaged state at the current moment. This generates the theoretical damaged state for the next time step. This scheme explicitly separates the diffusion and source generation mechanisms, enabling the final prediction results to be a true simulation of the pathological dynamics.

[0026] Example 4: The projection verification module generates realistic residuals based on real-world multimodal observation data and a pure ideal state benchmark, and generates theoretical residuals based on theoretical damaged states and a pure ideal state benchmark, including: Calculate the difference between real-world multimodal observation data and a pure ideal-state baseline to generate real-world residuals; The difference between the theoretically damaged state and the pure ideal state baseline is calculated to generate the theoretical residual.

[0027] In this embodiment, the projection verification module generates real residuals based on real multimodal observation data and pure ideal state benchmarks, and generates theoretical residuals based on theoretical damaged states and pure ideal state benchmarks. The process aims to separate real pathological signals from environmental noise. This process calculates real-world multimodal observation data. Compared with the pure ideal state benchmark The difference between them generates the actual residual. ,Right now The reality residual Physically, it contains a mixture of signals including real pathological changes, machine imaging artifacts, and random measurement errors; Meanwhile, the module calculates the theoretical damaged state Compared with the pure ideal state benchmark The difference between them generates the theoretical residual. ,Right now The theoretical residual It contains only pure pathological features derived from the kinetic equations, without any noise; by introducing As a common reference frame, this embodiment transforms the complex signal separation problem into a comparison problem of two difference vectors.

[0028] Example 5: The projection verification module calculates the manifold coupling confidence score based on the actual residuals and the theoretical residuals, including: Invoke the encoder that includes layer normalization operations; The encoder is used to encode the real residuals to generate real latent space feature vectors; The theoretical residual is encoded using an encoder to generate a feature vector in the theoretical latent space; Calculate the inner product of the feature vectors in the real latent space and the feature vectors in the theoretical latent space; Based on the inner product, the magnitude of the feature vectors in the real latent space, the magnitude of the feature vectors in the theoretical latent space, and a preset distance sensitivity coefficient, a manifold coupling confidence score is generated.

[0029] In this embodiment, the projection verification module calculates the manifold coupling confidence based on the actual residual and the theoretical residual in order to avoid the error caused by direct comparison in the pixel space. The residual is mapped to the semantic latent space for orthogonal projection verification. The module calls an encoder that includes layer normalization operations. The encoder specifically consists of three consecutive convolutional blocks, each containing a... A convolutional layer, a layer normalization layer, and a ReLU activation function; a global average pooling layer is connected after the convolutional block, and the final output dimension is... 3D feature vectors; using this encoder to obtain real-world residuals Encode to generate real-world latent space feature vectors Using the same encoder for theoretical residuals Encode to generate theoretical latent space feature vectors The purpose of layer normalization is to eliminate the influence of the dimensions and amplitude of the input signal, ensuring that the feature vector focuses on the morphological structure of the signal rather than its absolute intensity. Calculate the inner product of the real latent space feature vector and the theoretical latent space feature vector, and then use this inner product, the magnitude of the real latent space feature vector, the magnitude of the theoretical latent space feature vector, and a preset distance sensitivity coefficient as the basis for the calculation. Generate manifold coupling confidence The calculation formula is as follows:

[0030] The first term represents cosine similarity, used to measure directional consistency; the second term represents distance similarity based on radial basis functions, used to measure spatial overlap; where... and The normalized weight coefficients are preferably all 0.5; the first term represents the cosine similarity. Let be the numerical stability constant, and take . Used to calculate the denominator by Operations prevent division-to-zero anomalies; the second exponential term is a Gaussian radial basis function based on Euclidean distance, measuring the distance between the two in the manifold space; among which, the distance sensitivity coefficient... This is a key hyperparameter, and its value is not arbitrarily set. Instead, it is determined during the system calibration phase by plotting receiver operating characteristic (ROCCurve) curves on a historical validation set containing known pathological labels, and selecting the inverse of the variance of the characteristic distribution that maximizes the Youden index. This confidence level... It acts as a semantic filter, only when real-world observations conform to physical laws. During isomorphism, The value will then approach 1.

[0031] Example 6: The closed-loop correction module outputs disease course prediction results based on manifold coupling confidence, including: Obtain the adaptive decision threshold; If the manifold coupling confidence is greater than the adaptive decision threshold, the actual residual is determined to contain true pathological signals, and the true pathological signals are retained as the disease course prediction result. If the manifold coupling confidence level is less than or equal to the adaptive decision threshold, the actual residual is determined to be a non-pathological artifact, and a filtering removal operation is performed.

[0032] In this embodiment, the process by which the closed-loop correction module outputs the disease course prediction result based on the manifold coupling confidence involves adaptive signal decision, which aims to distinguish between the evolution of real lesions and random noise from a statistical perspective. The module obtains the adaptive manifold decision threshold. The threshold It is a dimensionless scalar, and its value is determined based on the statistical distribution analysis during the system calibration phase; specifically, the system calculates the manifold coupling confidence on a validation set that includes pure noise samples and confirmed pathological samples. The distribution of the data was analyzed, and a receiver operating characteristic (ROCCurve) curve was plotted. The confidence level corresponding to the maximization of the Youden index was selected as the baseline. Or, as an alternative implementation method, It can also be set as a pure ideal state benchmark. The 95th percentile of the confidence distribution generated in non-lesion areas is used to ensure that the false positive rate is controlled at a preset level, such as below 5%. The module makes a logical decision: if the currently calculated manifold coupling confidence level... Greater than the adaptive manifold decision threshold This represents the actual residual. The manifold geometry is highly consistent with theoretical pathological deductions, indicating that it contains true pathological signals. The system retains these signals as valid disease progression prediction results. If the manifold coupling confidence level... Less than or equal to the adaptive manifold decision threshold If the actual residual is determined to be mainly composed of non-pathological artifacts such as respiratory motion artifacts, equipment electromagnetic interference, or random measurement noise, the system performs a filtering and elimination operation to set the signal in that region to zero or mark it as a low-confidence region, thereby significantly reducing the false alarm rate of auxiliary diagnosis.

[0033] Example 7: The closed-loop correction module performs inverse optimization of industrial dynamic parameters based on the difference between actual and theoretical residuals in the latent space, including: Calculate the Euclidean distance error between the feature vectors of the real latent space and the feature vectors of the theoretical latent space; Calculate the gradient of the Euclidean distance error with respect to the industrial dynamic parameters; The industrial dynamics parameters are updated using gradients and a preset learning rate.

[0034] In this embodiment, the closed-loop correction module performs inverse optimization of industrial dynamic parameters based on the difference between the actual residual and the theoretical residual in the latent space, which is essentially solving an inverse problem. Due to the theoretical damaged state Based on time step Before calculating the error, the module first performs a time alignment operation on the evolving dynamic sequence: it aligns the theoretically damaged states of the full-time sequence generated by the perturbation simulation module. The sequence is mapped sequentially to a theoretical latent space feature vector sequence; the relationship between each frame in this sequence and the actual latent space feature vector is calculated. The similarity is used to select the time step with the smallest Euclidean distance. The corresponding feature vector is used as the current theoretical latent space feature vector. It should be noted that when constructing the computation graph, the time step is selected. The operation is treated as a non-differential index selection; When calculating subsequent gradients, It is considered a fixed constant for the current iteration step, and its value is only affected by the selected time step. Regarding parameters Calculate the partial derivative, without considering the time index. The gradient is calculated to ensure the continuity of the backpropagation path, thereby achieving optimal temporal matching between dynamic disease progression and static images. Specifically, this module calculates the real-world latent space feature vector. With theoretical latent space eigenvectors The Euclidean distance error between them, i.e. It should be noted that, due to yes The function, and Industrial dynamic parameters again Therefore, this error is also a function of , and thus the error is also The function; The module uses automatic differentiation technology to calculate the Euclidean distance error with respect to industrial dynamic parameters. gradient ; It should be noted that before performing the first round of inversion optimization, the system retrieves the standard statistical mean of the corresponding human tissue type from the medical-industrial parameter mapping library as the industrial kinetic parameter. initial value , or Initialize the data with normally distributed random numbers of non-zero mean to ensure that the perturbation simulation module has non-zero initial evolution values; use this gradient and a preset learning rate... The industrial dynamic parameters are updated using the gradient descent method, and the update formula is as follows: Through multiple iterations, the theoretical deductions were improved. By continuously approaching the pathological aspects of actual observation, we can learn the real physical parameters that govern the progression of the patient's disease.

[0035] Example 8: The closed-loop correction module is also used for: When the Euclidean distance error converges, output the industrial dynamic parameters; Industrial kinetic parameters are used as quantified, personalized pathological parameters.

[0036] In this embodiment, the closed-loop correction module further performs a parameter output function; when the Euclidean distance error converges, that is, when the error change in two consecutive iterations is less than a preset convergence threshold, for example... When the system stops iterating, it outputs the current industrial dynamic parameters. ; At this time, the industrial dynamic parameters It is presented as a quantified, personalized pathological parameter; for example, if from industrial kinetic parameters The spontaneous diffusion term coefficient extracted corresponds to the elastic modulus damping coefficient of liver tissue, and the system directly outputs the current degree of liver cirrhosis as 0.45. This feature enables the present invention to provide a second dimension of diagnostic information in addition to imaging, and can more sensitively reflect early pathological changes, realizing a technological leap from qualitative image reading to quantitative measurement.

[0037] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A medical image disease progression prediction system based on industrial neural networks, characterized in that, include: The baseline generation module is used to acquire still image data; A physical sensing neural network is used to process static image data to generate a pure ideal state benchmark. The perturbation simulation module is used to obtain industrial dynamic parameters; the industrial dynamic parameters are superimposed as perturbation terms onto the pure ideal state reference to generate the theoretical damaged state; The projection verification module is used to acquire real-world multimodal observation data; Based on real-world multimodal observation data and a pure ideal state benchmark, real-world residuals are generated; Based on the theoretical damaged state and the pure ideal state benchmark, theoretical residuals are generated; Based on the actual residuals and theoretical residuals, the confidence level of manifold coupling is calculated. The closed-loop correction module is used to invert and optimize industrial dynamic parameters based on the difference between the actual residual and the theoretical residual in the latent space. Based on the manifold coupling confidence level, the disease course prediction results are output.

2. The medical image disease progression prediction system based on industrial neural networks according to claim 1, characterized in that, The benchmark generation module uses a physical sensing neural network to process static image data and generate a pure, ideal benchmark, including: Determine the set of boundary points of the anatomical structure; Determine the set of points for the internal regions of anatomical structures; At the set of boundary points, the boundary conditions output by the physical sensing neural network are forced to match the static image data; At the set of points in the internal region, the output of the physical perception neural network is forced to satisfy the preset physical control equation; Construct a weighted sum of boundary fitting error and physical equation residuals; Minimize the weighted sum until the output of the physical perception neural network converges, generating a pure ideal state benchmark.

3. The medical image disease progression prediction system based on industrial neural networks according to claim 1, characterized in that, The perturbation simulation module superimposes industrial dynamic parameters as perturbation terms onto a pure ideal state baseline to generate a theoretically damaged state, including: Extracting the spontaneous diffusion term coefficient from industrial kinetic parameters; Extracting external disturbance coefficients from industrial dynamic parameters; Based on the spontaneous diffusion term coefficient, calculate the diffusion increment of the theoretically damaged state at the current moment; The source generation increment is calculated based on the external disturbance term coefficient and the pure ideal state benchmark. The diffusion increment and the source generation increment are superimposed on the theoretically damaged state at the current moment to generate the theoretically damaged state at the next moment.

4. The medical image disease progression prediction system based on industrial neural networks according to claim 1, characterized in that, The projection verification module generates real residuals based on real multimodal observation data and a pure ideal state benchmark, and generates theoretical residuals based on theoretical damaged states and a pure ideal state benchmark, including: Calculate the difference between real-world multimodal observation data and a pure ideal-state baseline to generate real-world residuals; The difference between the theoretically damaged state and the pure ideal state baseline is calculated to generate the theoretical residual.

5. A medical image disease progression prediction system based on industrial neural networks according to claim 1, characterized in that, The projection verification module calculates the manifold coupling confidence level based on the actual residual and the theoretical residual, including: Invoke the encoder that includes layer normalization operations; The encoder is used to encode the real residuals to generate real latent space feature vectors; The theoretical residual is encoded using an encoder to generate a feature vector in the theoretical latent space; Calculate the inner product of the feature vectors in the real latent space and the feature vectors in the theoretical latent space; Based on the inner product, the magnitude of the feature vectors in the real latent space, the magnitude of the feature vectors in the theoretical latent space, and a preset distance sensitivity coefficient, a manifold coupling confidence score is generated.

6. The medical image disease progression prediction system based on industrial neural networks according to claim 1, characterized in that, The closed-loop correction module outputs disease course prediction results based on manifold coupling confidence, including: Obtain the adaptive decision threshold; If the manifold coupling confidence is greater than the adaptive decision threshold, the actual residual is determined to contain true pathological signals, and the true pathological signals are retained as the disease course prediction result. If the manifold coupling confidence level is less than or equal to the adaptive decision threshold, the actual residual is determined to be a non-pathological artifact, and a filtering removal operation is performed.

7. A medical image disease progression prediction system based on industrial neural networks according to claim 5, characterized in that, The closed-loop correction module performs inverse optimization of industrial dynamic parameters based on the difference between the actual residual and the theoretical residual in the latent space, including: Calculate the Euclidean distance error between the feature vectors of the real latent space and the feature vectors of the theoretical latent space; Calculate the gradient of the Euclidean distance error with respect to the industrial dynamic parameters; The industrial dynamics parameters are updated using gradients and a preset learning rate.

8. A medical image disease progression prediction system based on industrial neural networks according to claim 7, characterized in that, The closed-loop correction module is also used for: When the Euclidean distance error converges, output the industrial dynamic parameters; Industrial kinetic parameters are used as quantified, personalized pathological parameters.

Images