Urinary system tumor big data analysis system

Abstract

The invention relates to the technical field of medical data analysis, in particular to a urinary system tumor big data analysis system which comprises a target kernel generation module, a kernel matrix calculation module, a parameter optimization module and a risk division module. According to the method, a target kernel matrix reflecting clinical prognosis differences is constructed and serves as an optimization reference, a radiomics and genomics feature kernel matrix is generated through hardware acceleration parallel computing, kernel function width parameters are dynamically iteratively updated based on an alignment degree numerical value so as to ensure that multi-modal feature distribution is highly matched with a prognosis label, and the accuracy of the multi-modal feature distribution is improved. A multi-dimensional feature space containing rich pathological information is constructed by combining a weighted fusion mechanism after centralization processing, so that a support vector machine is trained to determine a high-robustness decision boundary, and precise division of tumor risk levels is realized while high-dimensional data calculation delay is greatly reduced; and the reliability and timeliness of auxiliary diagnosis and treatment results in a complex pathological environment are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of medical data analysis technology, and in particular to a big data analysis system for urinary tract tumors. Background Technology

[0002] Medical data analytics is an interdisciplinary field that specializes in processing, analyzing, and interpreting healthcare data. Its core components include patient clinical records, imaging data, genomic data, and public health statistics. This field systematically collects, stores, integrates, and processes this heterogeneous data to discover disease patterns, support clinical decision-making, and optimize healthcare service processes. It is a key support for improving diagnostic efficiency and treatment outcomes in modern healthcare systems. Traditional urological tumor big data analysis systems refer to the pathological information, clinical imaging, and treatment follow-up data accumulated for urological cancers such as kidney cancer, bladder cancer, and prostate cancer. These systems typically employ statistical methods or basic data mining algorithms to perform basic descriptive statistics, correlation analysis, and trend prediction on the small amount of structured clinical data collected, in order to conduct preliminary explorations into the pathogenesis and treatment response of urological tumors.

[0003] Current analyses of urinary tract tumors rely on basic statistical methods to process small amounts of structured clinical data, making it difficult to effectively capture the high-dimensional nonlinear correlations within complex medical data. Descriptive statistics based solely on single-dimensional data cannot deeply explore the potential patterns behind unstructured data such as imaging and genomics. This lack of ability to integrate multi-source heterogeneous information makes it difficult to establish comprehensive and accurate disease evolution models. Consequently, when faced with tumor cases with significant individual differences, it is impossible to accurately predict disease progression and treatment response, thus severely limiting the scientific rigor and predictability in the development of clinical treatment plans. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a big data analysis system for urinary tract tumors.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a urinary tract tumor big data analysis system includes: The target kernel generation module compares the clinical prognostic labels with the central processing unit. If the labels are the same, the value is set to 1; otherwise, the value is set to 0. The target kernel matrix is ​​then constructed. The kernel matrix calculation module, wherein the central processing unit is configured with a reconfigurable hardware acceleration card, the reconfigurable hardware acceleration card generates radiomics kernel matrices and genomics kernel matrices; The parameter optimization module sends the radiomics kernel function width parameter and the genomics kernel function width parameter to the reconfigurable hardware acceleration card to calculate the alignment value with the target kernel matrix, and updates the radiomics kernel function width parameter and the genomics kernel function width parameter based on the alignment value. In the risk classification module, if the central processor meets the stopping condition, it generates a centralized radiomics kernel matrix and a centralized genomics kernel matrix, and performs a weighted summation to form a final fusion kernel matrix. Based on the final fusion kernel matrix and the clinical prognostic labels, a support vector machine is trained to construct a decision boundary, and the decision boundary is used to predict the risk level.

[0006] As a further aspect of the present invention, the target kernel matrix is ​​a symmetric square matrix generated based on whether the clinical prognostic labels of any two patient samples are consistent, and its matrix elements consist of the values ​​1 or 0. The radiomics kernel matrix is ​​a similarity matrix calculated based on radiomics feature data and the width parameter of the radiomics kernel function. The genomics kernel matrix is ​​a similarity matrix calculated based on genomics feature data and the width parameter of the genomics kernel function. The final fusion kernel matrix is ​​a weighted combination of the centralized radiomics kernel matrix and the centralized genomics kernel matrix. The decision boundary is a classification hyperplane constructed based on support vector machine parameters and risk level classification thresholds to divide differentiated risk levels.

[0007] As a further aspect of the present invention, the specific function of the target kernel generation module is as follows: The label vector acquisition submodule retrieves clinical prognostic labels corresponding to multiple patient samples in a specified queue from the patient database in batches. The clinical prognostic labels are binary-classified based on whether the patient has recurrence or metastasis within the postoperative target follow-up time window, and the clinical prognostic labels of all patient samples are organized into a one-dimensional prognostic label vector. The matrix construction submodule obtains the prognostic label vector. The central processing unit traverses all patient sample pairs in the prognostic label vector. For any pair of patient samples, it indexes and compares the corresponding clinical prognostic label values ​​in the prognostic label vector. If the values ​​are the same, a value of 1 is assigned to the corresponding row and column position of the target kernel matrix. If the values ​​are different, a value of 0 is assigned. By performing the comparison operation on all sample pairs until the entire matrix is ​​filled, a complete symmetric square matrix representing the prognostic consistency between samples is generated as the target kernel matrix.

[0008] As a further aspect of the present invention, the specific function of the kernel matrix calculation module is as follows: The image kernel computation submodule, where the central processing unit sends the imageomics feature data extracted from the database and the current imageomics kernel function width parameter to the reconfigurable hardware accelerator card, is based on a Gaussian kernel function, according to the formula: ; The radiomics feature data vectors of any two patient samples are paired to calculate the radiomics similarity between the samples, and a radiomics kernel matrix is ​​generated. in, Representative patient sample With patient samples The radiomics similarity kernel function values ​​between them and Representing patient samples respectively and patient samples Multidimensional image omics feature data vectors This represents the square of the Euclidean distance between two feature data vectors. The width parameter of the radiomics kernel function is used to control the range of the radial action of the Gaussian function; The gene kernel calculation submodule, wherein the central processing unit sends the genomic feature data extracted from the database and the current genomic kernel function width parameter to the reconfigurable hardware acceleration card, wherein the reconfigurable hardware acceleration card performs pairwise calculations on the genomic feature data vectors of any two patient samples based on the Gaussian kernel function to obtain the genomic similarity between the samples and generate a genomic kernel matrix; The calculation of the genomic kernel matrix uses the same formula structure, but the inputs are the genomic feature data and the genomic kernel function width parameter.

[0009] As a further aspect of the present invention, the specific function of the parameter optimization module is as follows: In the alignment calculation submodule, the central processing unit sends the radiomics kernel function width parameter, the genomics kernel function width parameter, and a preset fusion weight coefficient for the current iteration cycle to the reconfigurable hardware accelerator card. The reconfigurable hardware accelerator card performs a weighted summation of the radiomics kernel matrix and the genomics kernel matrix to generate a temporary fusion kernel matrix, and then applies the formula: ; Calculate the similarity between the temporary fusion kernel matrix and the target kernel matrix to generate an alignment score. in, This represents the alignment value. This represents the temporary fusion kernel matrix after central processing. Represents the target kernel matrix, The Frobenius inner product is represented by summing the products of corresponding elements of the two matrices. This represents the Frobenius norm, which is calculated by taking the square root of the sum of the squares of all elements in the matrix. In the width parameter update submodule, the central processing unit receives the alignment value, calculates the partial derivative of the alignment value with respect to the width parameters of the radiomics kernel function and the genomics kernel function using the gradient ascent method, obtains the gradient vector, and iteratively updates the two width parameters with a preset learning rate to generate a set of optimized width parameters for the next round of calculation.

[0010] As a further aspect of the present invention, the specific function of the risk allocation module is as follows: In the matrix centering submodule, after the stopping condition is met, the central processing unit obtains the radiomics kernel matrix and the genomics kernel matrix generated using the final optimized width parameter, and centers them through a predefined centering matrix with the same dimensions as the kernel matrices. Perform the following on the radiomics kernel matrix and the genomics kernel matrix: The transformation operations generate a centralized radiomics kernel matrix and a centralized genomics kernel matrix; The fusion kernel generation submodule, wherein the central processing unit performs a weighted summation of the centralized radiomics kernel matrix and the centralized genomics kernel matrix based on the fusion weight coefficient corresponding to maximizing the alignment value during the parameter optimization process, to obtain the final fusion kernel matrix; In the model training submodule, the central processing unit takes the final fusion kernel matrix as input and the clinical prognosis label as the training target to train the support vector machine classifier. It captures the optimal classification hyperplane that can separate the differential class sample points with the maximum geometric margin by solving a quadratic programming problem, and generates the decision boundary. The risk prediction submodule, wherein the central processing unit uses the decision boundary to perform risk scoring prediction on the current patient data, and classifies the risk level to which the patient belongs based on the risk level classification threshold.

[0011] As a further aspect of the present invention, the process by which the width parameter update submodule updates the width parameter of the radiomics kernel function and the width parameter of the genomics kernel function includes: The alignment value generated by the alignment calculation submodule is obtained. The central processing unit calculates the partial derivative of the alignment value with respect to the width parameter of the image omics kernel function according to the chain rule, and multiplies the partial derivative value with a preset image omics learning rate to obtain the image width parameter update amount. The central processing unit calculates the partial derivative of the alignment value with respect to the genomic kernel function width parameter in the same way, and multiplies the partial derivative value by a preset genomic learning rate to obtain the genomic width parameter update amount; The image width parameter update amount is added to the current image omics kernel function width parameter, and the genomics width parameter update amount is added to the current genomics kernel function width parameter, respectively generating updated image omics kernel function width parameters and genomics kernel function width parameters for the next iteration, until the stopping condition is met.

[0012] As a further aspect of the present invention, the process by which the image kernel calculation submodule and the gene kernel calculation submodule generate the image omics kernel matrix and the genomics kernel matrix is ​​specifically as follows: For any two patient samples in the dataset, the reconfigurable hardware acceleration card calculates the square of their Euclidean distance in parallel in the hardware logic unit. The square of the distance is then negative and divided by twice the square of the corresponding kernel function width parameter. The calculation result is then used as an exponent, and a kernel function value is obtained through exponential function operation. The reconfigurable hardware acceleration card applies the computation process in parallel to all patient sample pairs and writes all the calculated kernel function values ​​into a pre-allocated two-dimensional matrix storage space in real time according to the index of the sample pair. The row index and column index of the matrix correspond to the numbers of the two patient samples, thereby generating the complete radiomics kernel matrix or the genomics kernel matrix after a complete parallel computation.

[0013] As a further aspect of the present invention, the process by which the risk prediction submodule predicts the risk level using the decision boundary includes: Obtain radiomics and genomics feature data of a new patient sample. Using the optimized radiomics kernel function width parameter and the genomics kernel function width parameter, calculate the radiomics kernel function value and the genomics kernel function value between the new sample and all support vector samples in the training set, and perform weighted fusion to obtain a fused kernel vector. The central processing unit performs an inner product operation between the fused kernel vector and the support vector coefficients in the support vector machine parameters, adds a bias term, and obtains a continuous risk prediction raw score based on the decision function of the support vector machine. The central processing unit compares the original risk prediction score with one or more preset risk level classification thresholds. If the score is higher than the high-risk threshold, it is determined to be high-risk; if it is lower than the low-risk threshold, it is determined to be low-risk; or if the score falls into a certain preset range, it is determined to be medium-risk, thus generating the final risk level of the new patient sample.

[0014] As a further aspect of the present invention, the stopping condition is any one of the following: If, during the iteration process, the absolute difference between the alignment values ​​calculated in two consecutive iterations by the parameter optimization module is less than a preset convergence tolerance threshold, it indicates that the parameter optimization process has converged to a local optimum. Alternatively, if the total number of iterations performed by the parameter optimization module reaches a preset maximum iteration limit, the parameter optimization process will be forcibly terminated regardless of whether the alignment value has converged, in order to avoid excessive consumption of computing resources, and the currently obtained radiomics kernel function width parameter and genomics kernel function width parameter will be output as the final optimization result.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, a target kernel matrix reflecting clinical prognostic differences is constructed and used as an optimization benchmark. Hardware-accelerated parallel computing is used to generate radiomics and genomics feature kernel matrices. The kernel function width parameter is dynamically updated based on the alignment value to ensure that the multimodal feature distribution and prognostic label are highly consistent. A multidimensional feature space containing rich pathological information is constructed by combining a weighted fusion mechanism after centralized processing. A support vector machine is then trained to establish a highly robust decision boundary. This significantly reduces the computational latency of high-dimensional data while achieving accurate classification of tumor risk levels, effectively improving the reliability and timeliness of auxiliary diagnosis and treatment results in complex pathological environments. Attached Figure Description

[0016] Figure 1 This is an overall flowchart of the urinary tract tumor big data analysis system of the present invention; Figure 2 This is a flowchart of the target kernel matrix generation process of the present invention; Figure 3 This is a flowchart of the multimodal kernel matrix calculation process of the present invention; Figure 4 This is a flowchart of the kernel function width parameter optimization process of this invention; Figure 5 This is a flowchart illustrating the risk hierarchy classification process based on support vector machines in this 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] All user-related information involved in this invention (including but not limited to biometric information, identity verification information, behavioral data, device information, and other data that can be used for identity verification and personalized services) is collected and processed with the user's full knowledge and voluntary consent. The collection, storage, and use of all information strictly comply with applicable national and regional laws and regulations, and meet relevant data protection standards and policy requirements. The use of data is limited to purposes necessary for providing the technical services of this invention, and reasonable technical and management measures will be taken to ensure the security and confidentiality of users' personal information in terms of information protection and privacy.

[0020] Please see Figure 1 and Figure 2 The present invention provides a technical solution: a urinary tract tumor big data analysis system comprising: The target kernel generation module compares the clinical prognostic labels with the central processing unit. If the labels are the same, the value is set to 1; otherwise, the value is set to 0. The target kernel matrix is ​​then constructed. The target kernel matrix is ​​a symmetric square matrix generated based on whether the clinical prognostic labels of any two patient samples are consistent. Its matrix elements consist of the values ​​1 or 0. The specific functions of the target kernel generation module are as follows: The label vector acquisition submodule retrieves clinical prognostic labels for multiple patient samples in a specified queue from the patient database in batches. The clinical prognostic labels are binary-classified based on whether the patient experienced recurrence or metastasis within the postoperative target follow-up time window, and the clinical prognostic labels of all patient samples are organized into a one-dimensional prognostic label vector. The matrix construction submodule obtains the prognostic label vector. The central processing unit traverses all patient sample pairs in the prognostic label vector. For any pair of patient samples, it indexes and compares the corresponding clinical prognostic label values ​​in the prognostic label vector. If the values ​​are the same, it assigns a value of 1 to the corresponding row and column position of the target kernel matrix. If the values ​​are different, it assigns a value of 0. By performing the comparison operation on all sample pairs until the entire matrix is ​​filled, a complete symmetric square matrix representing the prognostic consistency between samples is generated as the target kernel matrix.

[0021] In practice, the central processing unit (CPU) first executes the function of the label vector acquisition submodule. This submodule accesses a pre-built patient database storing information on patients with urinary tract tumors. In this embodiment, a cohort of five bladder cancer patients is used as an example, all of whom underwent radical cystectomy. The target follow-up time window is set at five years. Clinical prognostic labels are binary-classified based on whether the patient experienced tumor recurrence or distant metastasis within this five-year time window; no event is recorded as 0, and an event is recorded as 1. The CPU sequentially reads the IDs of these five patients and their corresponding prognostic labels from the database: Patient P1, label 0; Patient P2, label 1; Patient P3, label 0; Patient P4, label 1; Patient P5, label 0. Subsequently, the processor organizes these label values ​​into a one-dimensional prognostic label vector. Its specific form is: .

[0022] Next, the central processing unit initiates the matrix construction submodule. This submodule first obtains the prognostic label vector generated in the previous step. The central processing unit initializes a 5x5 zero matrix in memory to store the final target kernel matrix. Then, the processor starts a nested loop, with the outer loop variable... Iterate from 1 to 5, inner loop variable Similarly, iterate from 1 to 5, representing all possible patient sample pairs. In each loop, the processor determines the index... and In prognostic label vector The processor extracts the clinical prognostic label values ​​for each of the two patients. For example, for a sample pair (P1, P2), the processor indexes to... The value is 0. The value is 1. The processor performs a comparison operation; since 0 is not equal to 1, it writes the value 0 to the position of the first row, second column and the second row, first column of the target kernel matrix. For example, for the sample pair (P1, P3), the processor indexes to... The value is 0. The value is 0. The processor performs a comparison operation; since 0 equals 0, it writes the value 1 to the position of the first row, third column, and the third row, first column of the target kernel matrix. This process continues until all 25 sample pairs have been traversed. Diagonal elements Since the label is always the same when compared with itself, all diagonal elements are assigned the value 1. After the entire process is completed, a complete 5x5 symmetric matrix is ​​generated in memory, which is the target kernel matrix representing the prognostic consistency between samples. The details are as follows: .

[0023] Please see Figure 1 and Figure 3 The kernel matrix calculation module is equipped with a central processing unit and a reconfigurable hardware acceleration card, which generates radiomics kernel matrices and genomics kernel matrices. The radiomics kernel matrix is ​​a similarity matrix calculated based on radiomics feature data and the width parameter of the radiomics kernel function, while the genomics kernel matrix is ​​a similarity matrix calculated based on genomics feature data and the width parameter of the genomics kernel function. The specific functions of the kernel matrix calculation module are as follows: The image kernel computation submodule, where the central processing unit (CPU) sends the imageomics feature data extracted from the database and the current imageomics kernel function width parameter to the reconfigurable hardware accelerator card, is based on a Gaussian kernel function, according to the formula: ; The radiomics feature data vectors of any two patient samples are paired to calculate the radiomics similarity between the samples, and a radiomics kernel matrix is ​​generated. in, Representative patient sample With patient samples The radiomics similarity kernel function values ​​between them and Representing patient samples respectively and patient samples Multidimensional image omics feature data vectors This represents the square of the Euclidean distance between two feature data vectors. The width parameter of the radiomics kernel function is used to control the range of the radial action of the Gaussian function; The gene kernel calculation submodule: The central processing unit sends the genomic feature data extracted from the database and the current genomic kernel function width parameter to the reconfigurable hardware accelerator card. The reconfigurable hardware accelerator card performs pairwise calculations on the genomic feature data vectors of any two patient samples based on the Gaussian kernel function to obtain the genomic similarity between the samples and generate the genomic kernel matrix. The calculation of the genomic kernel matrix uses the same formula structure, but the input is genomic feature data and the genomic kernel function width parameter; The specific process by which the image kernel computation submodule and the gene kernel computation submodule generate the radiomics kernel matrix and the genomics kernel matrix is ​​as follows: For any two patient samples in the dataset, the reconfigurable hardware accelerator card can calculate the square of their Euclidean distance in parallel in the hardware logic unit. The square of the distance is then negative and divided by twice the square of the corresponding kernel function width parameter. The result is then used as an exponent and a kernel function value is obtained through an exponential function operation. The reconfigurable hardware accelerator card applies the computation process in parallel to all patient sample pairs and writes all the calculated kernel function values ​​into a pre-allocated two-dimensional matrix storage space in real time according to the index of the sample pair. The row index and column index of the matrix correspond to the numbers of the two patient samples, thereby generating a complete radiomics kernel matrix or genomics kernel matrix after a complete parallel computation.

[0024] In practice, the central processing unit (CPU) first executes the image kernel calculation submodule. The CPU extracts radiomics feature data corresponding to the aforementioned five patients from the patient database and combines this data with the current radiomics kernel function width parameter. Together, they are sent to the configured reconfigurable hardware acceleration card. Radiomics features are extracted from the tumor region in preoperative contrast-enhanced CT images, including quantified features across 117 dimensions such as shape, texture, and wavelet transform. For clarity, a simplified three-dimensional feature vector is used here. Simultaneously, genomics feature data is also extracted. Table 1 lists the clinical prognostic labels, simplified radiomics feature vectors, and simplified genomics feature vectors for the five patients in this embodiment.

[0025] Radiomics kernel function width parameter The initial values ​​were determined through a preliminary validation experiment. The experiment used an independent validation set containing follow-up data from 200 cases. Scan within the interval [0.5, 10.0] with a step size of 0.1. For each... For each value, its image omics kernel matrix on the validation set is calculated, and the centered kernel alignment is calculated between it and the target kernel matrix of the validation set. Experimental data show that when When the value is within the range of [1.8, 2.4], the alignment accuracy is relatively high and stable. When the value is 2.0, the highest average alignment of 0.78 is obtained. Therefore, in the initial stage of this embodiment, the width parameter of the radiomics kernel function is set. .

[0026] After receiving data and parameters, the reconfigurable hardware acceleration card uses the Gaussian kernel function formula. Parallel computation of the kernel matrix. Among them, Representative patient sample and The radiomics similarity kernel function values ​​between them; and Patient samples and The multidimensional image omics feature data vectors, the specific values ​​of which are shown in Table 1; It is the square of the Euclidean distance between two feature data vectors, obtained by summing the squares of the differences between the corresponding dimensions of the vectors. It measures the degree of separation between the two samples in the image feature space. The width parameter of the radiomics kernel function is used to control the range of the radial action of the Gaussian function; As the denominator, the square of the distance is scaled; the outermost exponential function The scaled negative distance values ​​are mapped to the (0,1] interval to obtain the final similarity kernel function value.

[0027] To calculate the radiomics similarity between patients P1 and P2 For example: 1. Obtain the feature vector: , .

[0028] 2. Calculate the square of the Euclidean distance: .

[0029] 3. Substitute the width parameter for calculation: and Substituting the exponent part of the formula, we get .

[0030] 4. Calculate the final kernel function value: .

[0031] The parallel logic unit within the hardware accelerator card performs the above calculations simultaneously on all 10 unique sample pairs and writes the results in real time to the corresponding location in a 5x5 matrix storage space, thereby generating a complete radiomics kernel matrix in one go. .

[0032] Next, the central processing unit (CPU) executes the gene kernel computation submodule. The CPU extracts genomic feature data from the patient database, as shown in Table 1, along with the current genomic kernel function width parameter. Send to the reconfigurable hardware acceleration card. (Refer to...) Validation experiments with similar procedures were conducted. However, due to its higher dimensionality and sparsity, the optimal width parameter range for genomic data differs from that of radiomics. The experiments determined the width parameter of the genomic kernel function. The initial value is 5.0. The reconfigurable hardware accelerator card, based on the Gaussian kernel function, performs pairwise calculations on the genomic feature data vectors of any two patient samples to obtain the genomic similarity between the samples and generate a genomic kernel matrix. .

[0033] The aforementioned reconfigurable hardware acceleration card refers to a hardware processing unit whose internal logic and connections can be dynamically configured and programmed to adapt to specific computing tasks. In this embodiment, it is a field-programmable gate array (FPGA) card.

[0034] Please see Figure 1 and Figure 4 The parameter optimization module sends the radiomics kernel function width parameter and the genomics kernel function width parameter to the reconfigurable hardware acceleration card to calculate the alignment value with the target kernel matrix, and updates the radiomics kernel function width parameter and the genomics kernel function width parameter based on the alignment value. The specific functions of the parameter optimization module are as follows: In the alignment calculation submodule, the central processing unit sends the radiomics kernel function width parameter, the genomics kernel function width parameter, and a preset fusion weight coefficient for the current iteration cycle to the reconfigurable hardware accelerator card. The reconfigurable hardware accelerator card performs a weighted summation of the radiomics kernel matrix and the genomics kernel matrix to generate a temporary fusion kernel matrix, and then applies the formula: ; Calculate the similarity between the temporary fusion kernel matrix and the target kernel matrix, and generate alignment values; in, This represents the alignment value. This represents the temporary fusion kernel matrix after centralized processing. Represents the target kernel matrix. The Frobenius inner product is represented by summing the products of corresponding elements of the two matrices. This represents the Frobenius norm, which is calculated by taking the square root of the sum of the squares of all elements in the matrix. In the width parameter update submodule, the central processing unit receives the alignment value, calculates the partial derivative of the alignment value with respect to the width parameters of the radiomics kernel function and the genomics kernel function using the gradient ascent method, obtains the gradient vector, and iteratively updates the two width parameters with a preset learning rate to generate a set of optimized width parameters for the next round of calculation. The process of updating the width parameters of the radiomics kernel function and the genomics kernel function in the width parameter update submodule includes: The alignment value generated by the alignment calculation submodule is obtained. The central processing unit calculates the partial derivative of the alignment value with respect to the width parameter of the image omics kernel function according to the chain rule, and multiplies the partial derivative value with a preset image omics learning rate to obtain the update amount of the image width parameter. The central processing unit calculates the partial derivative of the alignment value with respect to the genomics kernel function width parameter in the same way, and multiplies the partial derivative value by a preset genomics learning rate to obtain the update amount of the genomics width parameter. The image width parameter update is added to the current image omics kernel function width parameter, and the genomics width parameter update is added to the current genomics kernel function width parameter. This generates updated image omics kernel function width parameters and genomics kernel function width parameters for the next iteration, until the stopping condition is met. The stopping condition is any of the following: If the absolute difference between the alignment values ​​calculated in two consecutive iterations is less than a preset convergence tolerance threshold during the parameter optimization module's iteration process, it indicates that the parameter optimization process has converged to a local optimum. Alternatively, if the total number of iterations performed by the parameter optimization module reaches a preset maximum iteration limit, the parameter optimization process will be forcibly terminated regardless of whether the alignment value has converged, in order to avoid excessive consumption of computing resources, and the currently obtained radiomics kernel function width parameter and genomics kernel function width parameter will be output as the final optimization result.

[0035] In practice, the central processing unit (CPU) initiates the alignment calculation submodule to begin the first iteration. The CPU then calculates the imageomics kernel width parameter for the current cycle. Genomics kernel function width parameter and a preset fusion weight coefficient. Send to the reconfigurable hardware acceleration card. This fusion weighting coefficient. The preset value is an empirical value obtained through the analysis of historical datasets. Setting it to 0.5 indicates that equal importance is given to the two modalities of data in the initial stage.

[0036] After receiving the parameters, the reconfigurable hardware acceleration card first retrieves the image omics kernel matrix calculated and generated by the previous module. and genomics kernel matrix The arithmetic logic unit within the hardware accelerator card performs a weighted summation operation to generate a temporary 5x5 fusion core matrix. Next, the hardware follows the formula. Calculate the temporary fusion kernel matrix and the target kernel matrix. The alignment. Among them, Represents the alignment value; This represents the temporary fusion kernel matrix after centralization, where the centralization operation is... ,in For a centered matrix; This represents the target kernel matrix constructed by the target kernel generation module; Represents the Frobenius inner product, which is calculated by converting the matrix... and Multiply the 25 corresponding elements one by one, and then sum all 25 products. Represents the Frobenius norm, which is calculated by taking a specified matrix ( or The formula calculates the cosine similarity value in the range [-1, 1] by squaring all 25 elements of the matrix, summing all the squared values, and finally taking the square root of the sum. The calculation logic of the whole formula is to calculate the inner product of the two matrices and normalize it by their respective norms.

[0037] The central processing unit receives the alignment value returned by the hardware, for example, the value obtained from the first iteration calculation. The width parameter update submodule then starts. The central processing unit calculates the alignment value according to the chain rule. For radiomics kernel function width parameter The partial derivative values. This process is decomposed into calculating the partial derivative values. , , and The product of four parts. The processor calculates the gradient matrix or gradient value of each part in turn, ultimately obtaining a product of four parts. The total partial derivative value, for example, calculated as follows: The partial derivative value is related to a preset imageomics learning rate. Multiplying these values ​​yields the updated image width parameter. The settings have been verified through experiments: a typical set of... Parameters are used to observe the alignment values ​​at different learning rates (e.g., 0.01, 0.05, 0.1, 0.2), and a value is chosen that ensures convergence speed without easily causing oscillations or divergence. In this example... Set to 0.05. The image width parameter update amount is... .

[0038] The central processing unit calculates the alignment value against the genomic kernel function width parameter in the same way. The partial derivative value. The processor calculates according to the chain rule. , , and The product of the four parts gives us the information about The total partial derivative value, for example, is obtained as follows: The partial derivative value is related to a preset genomics learning rate. Multiplying these yields the update amount of the genomics width parameter. Genomics learning rate. The value was also set to 0.1 in the experiment. The update amount of the genomics width parameter is... .

[0039] The image width parameter update is added to the current radiomics kernel function width parameter, and the genomics width parameter update is added to the current genomics kernel function width parameter, generating the updated radiomics kernel function width parameter and genomics kernel function width parameter for the next iteration, respectively. The updated radiomics kernel function width parameter is: The updated genomics kernel function width parameter is These two updated parameters will be used in the next round of iterations.

[0040] One of the stopping conditions for the iterative process is that the absolute difference between the alignment values ​​calculated in two consecutive iterations is less than a preset convergence tolerance threshold. This threshold is set as follows: This value represents a trade-off between model accuracy and computational cost: testing on the validation set revealed that the improvement in alignment was less than... At this point, the improvement in the final model's predictive performance is negligible. For example, in the 88th iteration, the calculated alignment is 0.8152, and in the 89th iteration, it is 0.8153, a difference of 0.0001, satisfying the stopping condition. Another stopping condition is when the total number of iterations reaches the preset maximum iteration limit of 500. This limit is set based on testing experience on multiple datasets; in most cases, the optimization process converges within 300 iterations, and 500 iterations are set as a termination condition to prevent excessive consumption of computational resources. Once either condition is met, the parameter optimization process terminates.

[0041] The chain rule mentioned above refers to a rule used in calculus to calculate the derivative of a composite function. Here, it is used to decompose the global partial derivative of the alignment function with respect to the width parameter of the kernel function into the product of multiple local partial derivatives of the alignment with the fused kernel, the fused kernel with respect to the fused kernel, the fused kernel with respect to the single-modal kernel, and the single-modal kernel with respect to the width parameter, thus solving the problem step by step.

[0042] Please see Figure 1 and Figure 5 In the risk classification module, if the central processing unit meets the stopping condition, it generates a centralized radiomics kernel matrix and a centralized genomics kernel matrix, which are weighted and summed to form a final fusion kernel matrix. Based on the final fusion kernel matrix and clinical prognostic labels, a support vector machine is trained to construct a decision boundary, and the decision boundary is used to predict the risk level. The final fusion kernel matrix is ​​a weighted combination of the centralized radiomics kernel matrix and the centralized genomics kernel matrix, and the decision boundary is a classification hyperplane constructed based on support vector machine parameters and risk level classification thresholds to divide differentiated risk levels. The specific functions of the risk allocation module are as follows: The matrix centering submodule, after meeting the stopping condition, retrieves the radiomics kernel matrix and genomics kernel matrix generated using the final optimized width parameter from the central processing unit, and then centers them through a predefined, same-dimensional centering matrix. Perform radiomics kernel matrix and genomics kernel matrix analysis The transformation operations generate a centralized radiomics kernel matrix and a centralized genomics kernel matrix; In the fusion kernel generation submodule, the central processing unit performs a weighted summation of the centralized radiomics kernel matrix and the centralized genomics kernel matrix based on the fusion weight coefficient that maximizes the alignment value during the parameter optimization process, to obtain the final fusion kernel matrix. In the model training submodule, the central processing unit takes the final fusion kernel matrix as input and the clinical prognosis label as the training target to train the support vector machine classifier. It captures the optimal classification hyperplane that can separate the differential class sample points with the maximum geometric margin by solving a quadratic programming problem, and generates the decision boundary. The risk prediction submodule uses the decision boundary to perform risk scoring prediction on the current patient data and classifies the risk level according to the risk level classification threshold. The risk prediction submodule uses the decision boundary to predict the risk level, and the process includes: Obtain radiomics and genomics feature data of a new patient sample. Using the optimized radiomics kernel function width parameter and genomics kernel function width parameter, calculate the radiomics kernel function value and genomics kernel function value between the new sample and all support vector samples in the training set, and perform weighted fusion to obtain a fused kernel vector. The central processing unit performs an inner product operation on the fused kernel vector and the support vector coefficients in the support vector machine parameters, adds a bias term, and obtains a continuous raw risk prediction score based on the decision function of the support vector machine. The central processing unit compares the raw risk prediction score with one or more preset risk level classification thresholds. If the score is higher than the high-risk threshold, it is judged as high-risk; if it is lower than the low-risk threshold, it is judged as low-risk; or if the score falls into a certain preset range, it is judged as medium-risk, thus generating the final risk level of the new patient sample.

[0043] After the parameter optimization module meets the stopping condition, the central processing unit obtains the final optimized imageomics kernel function width parameter. and genomic kernel function width parameter Simultaneously, during the optimization process, weighting coefficients are integrated. It was also adjusted as one of the optimization variables, and the optimal value was obtained. .

[0044] The central processing unit initiates the matrix centralization submodule. It first instructs the core matrix calculation module to use the final optimized width parameters. and Regenerate the final radiomics kernel matrix. and genomics kernel matrix The processor then constructs a 5x5 centralized matrix. Its definition ,in It is a 5x5 identity matrix. It is the sample size. It is a 5x5 matrix where all elements are 1. The processor performs matrix multiplication. and Centralized radiomics kernel matrix and centralized genomics kernel matrix are generated respectively.

[0045] Next, the fusion kernel generation submodule is activated. The central processing unit (CPU) then determines the optimal fusion weights. The two centered kernel matrices are weighted and summed to calculate... Specifically, this calculation is as follows: The result It is a 5x5 symmetric matrix.

[0046] Subsequently, the model training submodule is launched. The central processing unit will then finalize the fusion of the kernel matrix. As input, the clinical prognostic label vector (where 0 is mapped to -1, and 1 is mapped to +1, i.e.) The support vector machine classifier is trained using this as the training objective. This process involves calling a quadratic programming solver to find a set of Lagrange multipliers. To maximize the classification margin. After solving, the corresponding value for each training sample is obtained. value and a bias term In this example, the patient P1, P2, and P4 are obtained as support vectors, and their corresponding... The value is not zero, specifically: ,and At the same time, the bias term is obtained. These parameters together form the classification hyperplane, or decision boundary, used to distinguish differentiated risk levels.

[0047] Finally, when risk prediction is needed for a new patient sample (e.g., P_new), the risk prediction submodule is executed. This submodule first acquires the radiomics feature data of P_new. and genomic characteristics data Next, utilizing the optimized... and Calculate the kernel function values ​​between P_new and the three support vectors (P1, P2, P4) in the training set. 1. Calculate the image kernel vector: ,get 2. Calculate gene kernel vectors: ,get 3. Perform weighted fusion to obtain the fusion kernel vector: The central processing unit substitutes this fused kernel vector and support vector machine parameters into the decision function to perform an inner product operation, obtaining the raw risk prediction score: .

[0048] The processor compares this score to preset risk tier classification thresholds. These thresholds are determined based on the risk score distribution of an independent validation cohort of 500 patients. In this cohort, all scores are sorted from low to high, with the lowest third (33.3 percentile) defined as the low-risk threshold (-0.5) and the highest third (66.7 percentile) defined as the high-risk threshold (+0.5). Since the new patient's initial risk prediction score of -0.1379 falls between -0.5 and +0.5, the central processing unit ultimately classifies this new patient sample as "intermediate risk."

[0049] The above-mentioned centralized matrix A kernel matrix is ​​a matrix used to transform a kernel matrix so that the mean of the data in the feature space is zero, which is an inherent requirement of many machine learning algorithms.

[0050] The aforementioned support vector machine classifier is a supervised learning model based on statistical learning theory. Its core idea is to find a hyperplane in the feature space that can separate sample points of different classes with the maximum margin.

[0051] 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 big data analysis system for urinary tract tumors, characterized in that, The system includes: The target kernel generation module compares the clinical prognostic labels with the central processing unit. If the labels are the same, the value is set to 1; otherwise, the value is set to 0. The target kernel matrix is ​​then constructed. The kernel matrix calculation module, wherein the central processing unit is configured with a reconfigurable hardware acceleration card, the reconfigurable hardware acceleration card generates radiomics kernel matrices and genomics kernel matrices; The parameter optimization module sends the radiomics kernel function width parameter and the genomics kernel function width parameter to the reconfigurable hardware acceleration card to calculate the alignment value with the target kernel matrix, and updates the radiomics kernel function width parameter and the genomics kernel function width parameter based on the alignment value. In the risk classification module, if the central processor meets the stopping condition, it generates a centralized radiomics kernel matrix and a centralized genomics kernel matrix, and performs a weighted summation to form a final fusion kernel matrix. Based on the final fusion kernel matrix and the clinical prognostic labels, a support vector machine is trained to construct a decision boundary, and the decision boundary is used to predict the risk level.

2. The urinary tract tumor big data analysis system according to claim 1, characterized in that, The target kernel matrix is ​​a symmetric square matrix generated based on whether the clinical prognostic labels of any two patient samples are consistent. Its matrix elements consist of values ​​1 or 0. The radiomics kernel matrix is ​​a similarity matrix calculated based on radiomics feature data and the width parameter of the radiomics kernel function. The genomics kernel matrix is ​​a similarity matrix calculated based on genomics feature data and the width parameter of the genomics kernel function. The final fusion kernel matrix is ​​a weighted combination of the centralized radiomics kernel matrix and the centralized genomics kernel matrix. The decision boundary is a classification hyperplane constructed based on support vector machine parameters and risk level classification thresholds to divide differentiated risk levels.

3. The urinary tract tumor big data analysis system according to claim 1, characterized in that, The specific functions of the target kernel generation module are as follows: The label vector acquisition submodule retrieves clinical prognostic labels corresponding to multiple patient samples in a specified queue from the patient database in batches. The clinical prognostic labels are binary-classified based on whether the patient has recurrence or metastasis within the postoperative target follow-up time window, and the clinical prognostic labels of all patient samples are organized into a one-dimensional prognostic label vector. The matrix construction submodule obtains the prognostic label vector. The central processing unit traverses all patient sample pairs in the prognostic label vector. For any pair of patient samples, it indexes and compares the corresponding clinical prognostic label values ​​in the prognostic label vector. If the values ​​are the same, a value of 1 is assigned to the corresponding row and column position of the target kernel matrix. If the values ​​are different, a value of 0 is assigned. By performing the comparison operation on all sample pairs until the entire matrix is ​​filled, a complete symmetric square matrix representing the prognostic consistency between samples is generated as the target kernel matrix.

4. The urinary tract tumor big data analysis system according to any one of claims 1-2, characterized in that, The specific functions of the kernel matrix calculation module are as follows: The image kernel computation submodule, where the central processing unit sends the imageomics feature data extracted from the database and the current imageomics kernel function width parameter to the reconfigurable hardware accelerator card, is based on a Gaussian kernel function, according to the formula: ； The radiomics feature data vectors of any two patient samples are paired to calculate the radiomics similarity between the samples, and a radiomics kernel matrix is ​​generated. in, Representative patient sample With patient samples The radiomics similarity kernel function values ​​between them and Representing patient samples respectively and patient samples Multidimensional image omics feature data vectors This represents the square of the Euclidean distance between two feature data vectors. The width parameter of the radiomics kernel function is used to control the range of the radial action of the Gaussian function; The gene kernel calculation submodule, wherein the central processing unit sends the genomic feature data extracted from the database and the current genomic kernel function width parameter to the reconfigurable hardware acceleration card, wherein the reconfigurable hardware acceleration card performs pairwise calculations on the genomic feature data vectors of any two patient samples based on the Gaussian kernel function to obtain the genomic similarity between the samples and generate a genomic kernel matrix; The calculation of the genomic kernel matrix uses the same formula structure, but the inputs are the genomic feature data and the genomic kernel function width parameter.

5. The urinary tract tumor big data analysis system according to claim 1, characterized in that, The specific function of the parameter optimization module is as follows: In the alignment calculation submodule, the central processing unit sends the radiomics kernel function width parameter, the genomics kernel function width parameter, and a preset fusion weight coefficient for the current iteration cycle to the reconfigurable hardware accelerator card. The reconfigurable hardware accelerator card performs a weighted summation of the radiomics kernel matrix and the genomics kernel matrix to generate a temporary fusion kernel matrix, and then applies the formula: ； Calculate the similarity between the temporary fusion kernel matrix and the target kernel matrix to generate an alignment score. in, This represents the alignment value. This represents the temporary fusion kernel matrix after central processing. Represents the target kernel matrix, The Frobenius inner product is represented by summing the products of corresponding elements of the two matrices. This represents the Frobenius norm, which is calculated by taking the square root of the sum of the squares of all elements in the matrix. In the width parameter update submodule, the central processing unit receives the alignment value, calculates the partial derivative of the alignment value with respect to the width parameters of the radiomics kernel function and the genomics kernel function using the gradient ascent method, obtains the gradient vector, and iteratively updates the two width parameters with a preset learning rate to generate a set of optimized width parameters for the next round of calculation.

6. The urinary tract tumor big data analysis system according to claim 1, characterized in that, The specific functions of the risk allocation module are as follows: In the matrix centering submodule, after the stopping condition is met, the central processing unit obtains the radiomics kernel matrix and the genomics kernel matrix generated using the final optimized width parameter, and centers them through a predefined centering matrix with the same dimensions as the kernel matrices. Perform the following on the radiomics kernel matrix and the genomics kernel matrix: The transformation operations generate a centralized radiomics kernel matrix and a centralized genomics kernel matrix; The fusion kernel generation submodule, wherein the central processing unit performs a weighted summation of the centralized radiomics kernel matrix and the centralized genomics kernel matrix based on the fusion weight coefficient corresponding to maximizing the alignment value during the parameter optimization process, to obtain the final fusion kernel matrix; In the model training submodule, the central processing unit takes the final fusion kernel matrix as input and the clinical prognosis label as the training target to train the support vector machine classifier. It captures the optimal classification hyperplane that can separate the differential class sample points with the maximum geometric margin by solving a quadratic programming problem, and generates the decision boundary. The risk prediction submodule, wherein the central processing unit uses the decision boundary to perform risk scoring prediction on the current patient data, and classifies the risk level to which the patient belongs based on the risk level classification threshold.

7. The urinary tract tumor big data analysis system according to claim 5, characterized in that, The process by which the width parameter update submodule updates the width parameters of the radiomics kernel function and the genomics kernel function includes: The alignment value generated by the alignment calculation submodule is obtained. The central processing unit calculates the partial derivative of the alignment value with respect to the width parameter of the image omics kernel function according to the chain rule, and multiplies the partial derivative value with a preset image omics learning rate to obtain the image width parameter update amount. The central processing unit calculates the partial derivative of the alignment value with respect to the genomic kernel function width parameter in the same way, and multiplies the partial derivative value by a preset genomic learning rate to obtain the genomic width parameter update amount; The image width parameter update amount is added to the current image omics kernel function width parameter, and the genomics width parameter update amount is added to the current genomics kernel function width parameter, respectively generating updated image omics kernel function width parameters and genomics kernel function width parameters for the next iteration, until the stopping condition is met.

8. The urinary tract tumor big data analysis system according to claim 4, characterized in that, The specific process by which the image kernel calculation submodule and the gene kernel calculation submodule generate the image omics kernel matrix and the genomics kernel matrix is ​​as follows: For any two patient samples in the dataset, the reconfigurable hardware acceleration card calculates the square of their Euclidean distance in parallel in the hardware logic unit. The square of the distance is then negative and divided by twice the square of the corresponding kernel function width parameter. The calculation result is then used as an exponent, and a kernel function value is obtained through exponential function operation. The reconfigurable hardware acceleration card applies the computation process in parallel to all patient sample pairs and writes all the calculated kernel function values ​​into a pre-allocated two-dimensional matrix storage space in real time according to the index of the sample pair. The row index and column index of the matrix correspond to the numbers of the two patient samples, thereby generating the complete radiomics kernel matrix or the genomics kernel matrix after a complete parallel computation.

9. The urinary tract tumor big data analysis system according to claim 6, characterized in that, The process by which the risk prediction submodule predicts the risk level using the decision boundary includes: Obtain radiomics and genomics feature data of a new patient sample. Using the optimized radiomics kernel function width parameter and the genomics kernel function width parameter, calculate the radiomics kernel function value and the genomics kernel function value between the new sample and all support vector samples in the training set, and perform weighted fusion to obtain a fused kernel vector. The central processing unit performs an inner product operation between the fused kernel vector and the support vector coefficients in the support vector machine parameters, adds a bias term, and obtains a continuous risk prediction raw score based on the decision function of the support vector machine. The central processing unit compares the original risk prediction score with one or more preset risk level classification thresholds. If the score is higher than the high-risk threshold, it is determined to be high-risk; if it is lower than the low-risk threshold, it is determined to be low-risk; or if the score falls into a certain preset range, it is determined to be medium-risk, thus generating the final risk level of the new patient sample.

10. The urinary tract tumor big data analysis system according to claim 1, characterized in that, The stopping condition is any of the following: If, during the iteration process, the absolute difference between the alignment values ​​calculated in two consecutive iterations by the parameter optimization module is less than a preset convergence tolerance threshold, it indicates that the parameter optimization process has converged to a local optimum. Alternatively, if the total number of iterations performed by the parameter optimization module reaches a preset maximum iteration limit, the parameter optimization process will be forcibly terminated regardless of whether the alignment value has converged, in order to avoid excessive consumption of computing resources, and the currently obtained radiomics kernel function width parameter and genomics kernel function width parameter will be output as the final optimization result.

Images