Neural tumor recognition system based on artificial intelligence

Abstract

The invention discloses a neural tumor recognition system based on artificial intelligence in the technical field of medical image analysis. The system comprises a network architecture search module, a diagnosis path optimization module, a data self-supervised learning module, a continuous learning matching module, a dynamic tumor tracking module, an iterative image processing module, a result fusion feedback module and a system evaluation improvement module. According to the invention, network architecture search is carried out through the genetic algorithm and the deep Q network, the system can more accurately identify neural tumors, the diagnosis accuracy is improved, the diagnosis process is more transparent due to the application of the SHAP algorithm and the model local interpretable method, doctors can conveniently understand and judge a trust model, and use of a contrast learning algorithm and a clustering algorithm is facilitated; the method has higher efficiency and precision in the aspect of processing multi-source medical image data, and particularly provides powerful technical support for dynamic tracking of tumors through combination of an adaptive learning model and an optical flow algorithm in the aspect of construction of a heterogeneous data feature set.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image analysis, and in particular to a neural tumor recognition system based on artificial intelligence. Background Art

[0002] Medical image analysis technology uses computer vision, image processing, pattern recognition, and machine learning to conduct in-depth analysis of medical imaging data. The main purpose of this field is to improve the accuracy and efficiency of diagnosis while reducing the workload of medical professionals. It covers a wide range of technologies from basic image processing techniques (such as image enhancement and segmentation) to advanced algorithms (such as feature extraction and classification).

[0003] Among them, the AI-based neuro-tumor recognition system is a medical auxiliary tool that uses AI technology to identify and classify neuro-tumors. Its main purpose is to improve the speed and accuracy of identifying neuro-tumors through automated analysis, thereby assisting doctors in making more effective diagnoses and treatment plans. The system aims to improve the efficiency and reliability of medical imaging diagnosis by reducing dependence on professional radiologists, which is particularly important in environments with limited resources or a shortage of professional doctors.

[0004] Traditional systems lack flexibility and adaptability in the selection and optimization of network architecture, resulting in low efficiency and limited accuracy in processing complex and variable medical imaging data. Traditional systems are poor in transparency and explainability of the diagnostic process, which to some extent affects doctors' trust in and adoption of diagnostic results. Traditional systems lack effective integration and analysis methods for multi-source and heterogeneous data processing, which limits the comprehensiveness and accuracy of diagnosis. Due to the lack of effective performance evaluation and optimization mechanisms, traditional systems are prone to inefficiency and lack of adaptability in long-term operation.

[0005] Based on this, the present invention designs a neural tumor recognition system based on artificial intelligence to solve the above problems. Summary of the invention

[0006] The purpose of the present invention is to provide a neural tumor recognition system based on artificial intelligence to solve the problem that the traditional system proposed in the above background technology lacks flexibility and adaptability in the selection and optimization of network architecture, resulting in low efficiency and limited accuracy in processing complex and changeable medical imaging data. The traditional system is poor in transparency and interpretability of the diagnosis process, which to a certain extent affects the doctors' trust and adoption rate of the diagnosis results. For multi-source and heterogeneous data processing, the traditional system lacks effective integration and analysis methods, which limits the comprehensiveness and accuracy of the diagnosis. Due to the lack of effective performance evaluation and optimization mechanisms, traditional systems are prone to show problems of inefficiency and lack of adaptability in long-term operation.

[0007] To achieve the above-mentioned object, the present invention provides the following technical solutions: a neural tumor recognition system based on artificial intelligence includes a network architecture search module, a diagnostic path optimization module, a data self-supervised learning module, a continuous learning matching module, a dynamic tumor tracking module, an iterative image processing module, a result fusion feedback module and a system evaluation improvement module;

[0008] The network architecture search module uses a genetic algorithm based on a medical image dataset to initially screen the network architecture, uses a deep Q network to perform performance evaluation and parameter adjustment on the screened architecture, captures the network structure that matches the needs of neuro-tumor identification, and performs iterative architecture optimization to generate a neural network optimized architecture;

[0009] The diagnostic path optimization module is based on a neural network optimization architecture and uses the SHAP algorithm to quantitatively analyze model decision factors, use the model local interpretability method to locally analyze the model's predictions, and make the model's decision path transparent to generate an optimized diagnostic path.

[0010] The data self-supervised learning module is based on optimizing the diagnostic path, applying a contrastive learning algorithm, analyzing multi-source medical imaging data including MRI and CT scanning equipment, using a clustering algorithm to classify and fuse data features, and constructing and optimizing feature sets to generate heterogeneous data feature sets;

[0011] The continuous learning and matching module uses a support vector machine based on a heterogeneous data feature set to perform real-time updates and self-matching of the model, combines a recursive neural network to process newly accessed medical imaging data, continuously updates and optimizes the model, and generates an adaptive learning model;

[0012] The dynamic tumor tracking module is based on an adaptive learning model and uses an optical flow algorithm to track the movement trajectory of the tumor in the sequence images. It uses deformation field theory to analyze the morphological changes and growth trends of the tumor, analyzes and predicts the dynamic changes of the tumor, and generates dynamic tumor tracking results.

[0013] The iterative image processing module uses threshold segmentation based on the tumor dynamic tracking results to preliminarily divide the potential tumor area in the image, combines morphological segmentation and convolutional neural network to mine and analyze tumor characteristics, divide and identify the tumor area, and generate targeted tumor recognition results;

[0014] The result fusion feedback module is based on the targeted tumor identification results, combined with the neural network optimization architecture, optimized diagnostic pathway, heterogeneous data feature set, adaptive learning model and tumor dynamic tracking results, and uses the feedback mechanism to adjust the diagnosis and treatment strategies, organize and optimize the results, and generate a comprehensive diagnosis result;

[0015] The system evaluation and improvement module is based on the comprehensive diagnosis results, uses linear regression to analyze the correlation between system performance and parameters, uses hyperparameter optimization to adjust the key parameters of the identification system, optimizes the system configuration and adjusts the parameters, and performs system performance evaluation and optimization processing to generate system performance optimization records.

[0016] Preferably, the neural network optimization architecture includes a hierarchical structure of neural tumor recognition, activation function type selection and connection weight configuration, the optimized diagnostic path includes key feature recognition, decision node positioning and path transparency indicators, the heterogeneous data feature set includes texture features, edge features and shape descriptors of MRI and CT images, the tumor dynamic tracking results include tumor position changes, size changes and predicted growth trends, the targeted tumor recognition results include tumor positioning, boundary division and classification information, the comprehensive diagnosis results include summarized tumor recognition information, diagnostic path certainty and feature set integrity, and the system performance optimization records include performance evaluation indicators, key parameter adjustment records and optimized configuration plans.

[0017] Preferably, the network architecture search module includes an architecture generation submodule, a first performance evaluation submodule, and an architecture optimization submodule;

[0018] The architecture generation submodule uses gene expression programming based on the medical image dataset to simulate the genetic operation in the natural selection process, screen the matching key network architecture as the initial candidate set, and perform preliminary configuration of the architecture to generate the initial architecture candidate set;

[0019] The first performance evaluation submodule uses deep deterministic policy gradient based on the initial architecture candidate set to evaluate the performance of the network architecture by building a proxy model and interacting with the environment, focusing on the recognition accuracy and computational efficiency of the architecture on the target task, and generating an architecture performance evaluation result;

[0020] The architecture optimization submodule uses a simulated annealing algorithm based on the architecture performance evaluation results, randomly selects network parameters, calculates performance changes, optimizes according to a preset cooling plan, and iteratively optimizes potential architectures to generate a neural network optimization architecture.

[0021] Preferably, the diagnostic path optimization module includes a decision path tracking submodule, a key feature analysis submodule, and a path optimization submodule;

[0022] The decision path tracking submodule is based on a neural network optimization architecture and adopts a run test method to analyze the distribution of continuous variables in model prediction, identify statistical laws and trends in the decision path, and parse the model decision path to generate a decision path analysis record;

[0023] The key feature analysis submodule uses principal component analysis based on decision path analysis records, optimizes data dimensions by transforming data into a new coordinate system, highlights key variances in the data set, mines and analyzes key features of the model, and generates key feature analysis results;

[0024] The path optimization submodule uses a decision tree algorithm based on the key feature analysis results to build a tree structure model, map the relationship between features and output variables, and adjust the decision path of the model to optimize the path intuitiveness and feasibility, and generate an optimized diagnostic path.

[0025] Preferably, the data self-supervised learning module includes a data integration submodule, a feature extraction submodule, and a feature fusion submodule;

[0026] The data integration submodule uses a multi-source data fusion algorithm based on an optimized diagnostic pathway to synchronize medical imaging data of different formats and sources, including MRI and CT scan data, to perform standardized processing and quality control, verify data consistency and integrity, and generate a pooled imaging data set;

[0027] The feature extraction submodule uses a convolutional neural network and a contrastive learning algorithm based on the collected image data set to construct a multi-level feature extraction network to extract key features from the image data layer by layer from basic texture to salient pattern, and generate an extracted feature data set;

[0028] The feature fusion submodule uses a Gaussian mixture model based on the extracted feature data set to model the probability distribution of feature points, collect information from multiple perspectives and sources, cluster related features, identify and fuse related features, and generate a heterogeneous data feature set.

[0029] Preferably, the continuous learning and matching module includes an online learning submodule, a model adaptation submodule, and a data processing submodule;

[0030] The online learning submodule adopts online stochastic gradient descent based on heterogeneous data feature sets, receives and analyzes new medical imaging data in real time, performs model update processing, and dynamically adjusts the learning model to match the changes and new features of the data stream, thereby generating an online updated learning model;

[0031] The model adaptation submodule uses a support vector machine based on an online updated learning model to monitor the difference between data features and model outputs in real time, adjust the hyperparameters of the support vector machine, automatically adjust the model parameters to match the new data characteristics, and generate an adaptively adjusted model;

[0032] The data processing submodule uses a recursive neural network based on an adaptively adjusted model to serialize newly accessed medical imaging data, extract time-related features, optimize data format and quality, and generate an adaptive learning model.

[0033] Preferably, the dynamic tumor tracking module includes a deformation analysis submodule, a growth prediction submodule, and a tracking result integration submodule;

[0034] The deformation analysis submodule is based on an adaptive learning model and uses an optical flow algorithm to track the spatial position change of the tumor by calculating the displacement vector of the pixel points between consecutive image frames, analyze the time dimension change of the pixel points, determine the movement trajectory of the tumor, and generate the tumor trajectory analysis result;

[0035] The growth prediction submodule uses deformation field theory based on the tumor trajectory analysis results to construct a morphological change model of the tumor in continuous image frames, analyzes the deformation and growth trend of the tumor, and performs trend prediction to generate tumor change and growth prediction results;

[0036] The tracking result integration submodule extracts tumor movement trajectory and morphological change data based on tumor change and growth prediction results, uses data fusion technology to summarize the extracted information, analyzes the dynamic changes of the tumor, and establishes a tumor change view to generate tumor dynamic tracking results.

[0037] Preferably, the iterative image processing module includes a preliminary segmentation submodule, a refinement analysis submodule, and an iterative recognition submodule;

[0038] The preliminary segmentation submodule adopts threshold segmentation based on the dynamic tracking result of the tumor, analyzes the pixel distribution in the image, sets a matching grayscale threshold, distinguishes the tumor area from the background, and identifies the tumor area to generate a preliminary tumor segmentation result;

[0039] The refinement analysis submodule adopts morphological segmentation based on the preliminary segmentation result of the tumor, analyzes and defines the edge and structure of the tumor by applying morphological expansion and corrosion operations, optimizes the segmentation accuracy, and generates a refined segmentation result of the tumor;

[0040] The iterative recognition submodule uses a convolutional neural network based on the tumor segmentation results to identify and classify tumor features through the network structure, iteratively optimizes the recognition accuracy, and extracts the features of the tumor area layer by layer to generate targeted tumor recognition results.

[0041] Preferably, the result fusion feedback module includes a result fusion submodule, a strategy adjustment submodule, and a feedback optimization submodule;

[0042] The result fusion submodule is based on the targeted tumor recognition results, adopts a multi-dimensional data fusion algorithm, combines the neural network optimization architecture, optimized diagnostic pathways, heterogeneous data feature sets, adaptive learning models and tumor dynamic tracking results, and uses data alignment, synchronization and association analysis to verify the consistency of multi-source data in time and space, and generate preliminary comprehensive diagnosis results;

[0043] The strategy adjustment submodule uses a classification algorithm based on machine learning based on the preliminary comprehensive diagnosis results, analyzes the feasibility and accuracy of the current diagnosis results, combines rule-based logical inference and historical case comparison, evaluates the existing diagnosis results and proposes a strategy adjustment plan to generate an optimized diagnosis strategy;

[0044] The feedback optimization submodule applies an adaptive control algorithm based on the optimized diagnostic strategy, monitors system performance in real time, dynamically adjusts system parameters according to strategy evaluation results, optimizes diagnostic performance, and matches it to dynamically changing diagnostic requirements to generate comprehensive diagnostic results.

[0045] Preferably, the system evaluation and improvement module includes a second performance evaluation submodule, a result analysis submodule, and a system optimization submodule;

[0046] The second performance evaluation submodule uses linear regression based on the comprehensive diagnosis results to perform statistical analysis on the data between system performance and key parameters, mine patterns and trends in the data, identify key factors affecting system performance, and generate performance parameter correlation analysis results;

[0047] The result analysis submodule uses the K-means algorithm based on the performance parameter correlation analysis results to mine the influence of multi-parameter combinations on system performance by dividing the parameters into multiple clusters, analyze the relationship between key parameters and system performance, and generate system performance analysis results;

[0048] Based on the system performance analysis results, the system optimization submodule uses grid search and Bayesian optimization to adjust key parameters in the system, explore parameter space, adjust search strategies, optimize parameter configuration and system performance, and generate system performance optimization records.

[0049] Compared with the prior art, the beneficial effects of the present invention are as follows: by adopting genetic algorithms and deep Q networks to search network architectures, the system can more accurately identify neural tumors and improve the accuracy of diagnosis. The application of the SHAP algorithm and the local interpretability method of the model makes the diagnosis process more transparent, which is convenient for doctors to understand and trust the judgment of the model. The use of comparative learning algorithms and clustering algorithms shows higher efficiency and accuracy in processing multi-source medical imaging data, especially in the construction of heterogeneous data feature sets. The combination of adaptive learning models and optical flow algorithms provides strong technical support for dynamic tracking of tumors. The system effectively improves the overall operating efficiency and performance and reduces resource consumption through linear regression and hyperparameter optimization methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0051] Figure 1 A module diagram of a neural tumor recognition system based on artificial intelligence is proposed for the present invention;

[0052] Figure 2 A system framework diagram of a neural tumor recognition system based on artificial intelligence is proposed for the present invention;

[0053] Figure 3 A schematic diagram of a network architecture search module in a neural tumor recognition system based on artificial intelligence is proposed for the present invention;

[0054] Figure 4 A schematic diagram of a diagnostic path optimization module in a neural tumor recognition system based on artificial intelligence is proposed for the present invention;

[0055] Figure 5 The present invention proposes a schematic diagram of a data self-supervised learning module in a neural tumor recognition system based on artificial intelligence;

[0056] Figure 6 The present invention proposes a schematic diagram of a continuous learning and matching module in a neural tumor recognition system based on artificial intelligence;

[0057] Figure 7 A schematic diagram of a dynamic tumor tracking module in a neural tumor recognition system based on artificial intelligence is proposed for the present invention;

[0058] Figure 8 The present invention proposes a schematic diagram of an iterative image processing module in a neural tumor recognition system based on artificial intelligence;

[0059] Fig. 9 The present invention proposes a schematic diagram of a result fusion feedback module in a neural tumor recognition system based on artificial intelligence;

[0060] Fig.10 The present invention proposes a schematic diagram of a system evaluation improvement module in a neural tumor recognition system based on artificial intelligence. DETAILED DESCRIPTION

[0061] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0062] See also Figure 1 , the present invention provides a technical solution: a neural tumor recognition system based on artificial intelligence includes a network architecture search module, a diagnostic path optimization module, a data self-supervised learning module, a continuous learning matching module, a dynamic tumor tracking module, an iterative image processing module, a result fusion feedback module and a system evaluation improvement module;

[0063] The network architecture search module uses a genetic algorithm to initially screen the network architecture based on the medical image dataset, and uses a deep Q network to perform performance evaluation and parameter adjustment on the screened architecture, capturing the network structure that matches the needs of neuro-tumor identification, and iteratively optimizes the architecture to generate a neural network optimized architecture.

[0064] The diagnostic path optimization module is based on a neural network optimization architecture and uses the SHAP algorithm to quantitatively analyze the model's decision factors. It uses the model's local interpretability method to locally analyze the model's predictions and make the model's decision path transparent to generate an optimized diagnostic path.

[0065] The data self-supervised learning module is based on optimizing the diagnostic path and applying the contrastive learning algorithm to analyze multi-source medical imaging data including MRI and CT scanning equipment. It uses the clustering algorithm to classify and fuse data features, and constructs and optimizes feature sets to generate heterogeneous data feature sets.

[0066] The continuous learning matching module uses support vector machines based on heterogeneous data feature sets to update and self-match the model in real time. It combines recursive neural networks to process newly accessed medical imaging data, continuously update and optimize the model, and generate an adaptive learning model.

[0067] The dynamic tumor tracking module is based on an adaptive learning model and uses an optical flow algorithm to track the movement trajectory of the tumor in the sequence image. It uses deformation field theory to analyze the morphological changes and growth trends of the tumor, analyze and predict the dynamic changes of the tumor, and generate dynamic tumor tracking results.

[0068] The iterative image processing module uses threshold segmentation based on the dynamic tumor tracking results to preliminarily divide the potential tumor areas in the image. It combines morphological segmentation and convolutional neural networks to mine and analyze tumor characteristics, divide and identify tumor areas, and generate targeted tumor recognition results.

[0069] The result fusion feedback module is based on targeted tumor identification results, combined with neural network optimization architecture, optimized diagnostic pathways, heterogeneous data feature sets, adaptive learning models and tumor dynamic tracking results, and uses feedback mechanisms to adjust diagnosis and treatment strategies, organize and optimize results, and generate comprehensive diagnostic results.

[0070] Based on the comprehensive diagnosis results, the system evaluation and improvement module uses linear regression to analyze the correlation between system performance and parameters, and uses hyperparameter optimization to adjust the key parameters of the identification system, optimize the system configuration and adjust the parameters, and perform system performance evaluation and optimization processing to generate system performance optimization records.

[0071] The neural network optimization architecture includes the hierarchical structure of neural tumor recognition, activation function type selection and connection weight configuration. The optimized diagnostic path includes key feature identification, decision node positioning and path transparency indicators. The heterogeneous data feature set includes texture features, edge features and shape descriptors of MRI and CT images. The dynamic tumor tracking results include tumor position changes, size changes and predicted growth trends. The targeted tumor recognition results include tumor localization, boundary demarcation and classification information. The comprehensive diagnostic results include summarized tumor recognition information, diagnostic path certainty and feature set integrity. The system performance optimization records include performance evaluation indicators, key parameter adjustment records and optimized configuration plans.

[0072] In the network architecture search module, genetic algorithms and deep Q networks are used to optimize the neural network architecture for the specific needs of neuro-tumor recognition. In the initial stage, the genetic algorithm generates a variety of neural network architectures based on the medical image dataset. These architectures are regarded as individuals in the genetic algorithm, and their fitness is based on the recognition performance of the architecture. Through crossover and mutation operations, the genetic algorithm continuously evolves these network architectures to generate new architecture variants. Subsequently, the deep Q network performs in-depth performance evaluation and parameter adjustment on these architectures. In the reinforcement learning environment of the deep Q network, each network architecture is evaluated for its performance on specific tasks, such as the ability to accurately identify neuro-tumors. Based on performance feedback, the deep Q network adjusts architecture parameters, such as hierarchical structure, activation function type, and connection weights, to improve recognition results. This module ultimately generates an optimized neural network architecture specifically for the efficient recognition of neuro-tumors.

[0073] In the diagnostic pathway optimization module, the SHAP algorithm and the model local interpretability method are used to analyze and optimize the decision path of the neural network model. The role of the SHAP algorithm here is to clarify the contribution of each feature in the model decision through quantitative analysis, and assign a SHAP value to each feature to indicate the degree of influence of the feature on the model prediction results. This method reveals which features play a decisive role in the identification of neural tumors. In addition, the model local interpretability method is used to explain in detail why the model makes predictions for specific cases, which improves the transparency of the model decision path and enables doctors to understand and trust the model's diagnosis. Finally, the module generates an optimized diagnostic pathway, which helps to improve the accuracy and interpretability of the diagnosis.

[0074] In the data self-supervised learning module, the contrastive learning algorithm and clustering algorithm are combined to process multi-source medical imaging data, such as MRI and CT scans. The contrastive learning algorithm first learns the intrinsic structure of the data by comparing similar and different features in the medical imaging data, which helps to improve the robustness and accuracy of the model when processing complex medical images. Then, the clustering algorithm is used to classify and fuse these features to enhance the model's ability to recognize different types of neural tumor features. Through this process, the module constructs and optimizes a heterogeneous data feature set containing rich texture features, edge features, and shape descriptors, providing strong data support for the accurate identification of neural tumors.

[0075] In the continuous learning and matching module, support vector machines and recursive neural networks are used to achieve real-time updating and adaptive matching of the neural tumor recognition model. The support vector machine classifies and processes the heterogeneous data feature sets to ensure that the model can effectively distinguish different types of neural tumors. The recursive neural network is used to process newly added medical imaging data, allowing the model to self-adjust and optimize according to the latest data. This process ensures that the model can maintain high accuracy and flexibility while continuously receiving new data, generating an adaptive learning model that can adapt to new situations and data.

[0076] In the dynamic tumor tracking module, the movement and morphological changes of neural tumors in continuous medical images are tracked through the optical flow algorithm and deformation field theory. The optical flow algorithm accurately tracks the position and movement trajectory of the tumor in the sequence of images, while the deformation field theory analyzes the morphological changes and growth trends of the tumor. The key to this module is the ability to dynamically capture the changes in the tumor, providing valuable information for understanding the development of the tumor. The generated tumor dynamic tracking results include the changes in the tumor's position, size and its predicted growth trend, which are of great significance for formulating personalized treatment plans.

[0077] In the iterative image processing module, threshold segmentation, morphological segmentation and convolutional neural network are combined to conduct in-depth analysis and identification of tumor characteristics. Threshold segmentation first preliminarily divides the potential tumor area in the image, morphological segmentation then refines the shape and boundary of the tumor, and convolutional neural network further mines and analyzes the deep characteristics of the tumor. This comprehensive processing method enables the module to accurately divide and identify tumor areas, and generate targeted tumor recognition results that include tumor localization, boundary division and classification information, providing an important basis for subsequent treatment.

[0078] In the result fusion feedback module, the output results of all previous modules are combined, and adjustments and optimizations are made through the feedback mechanism to improve the efficiency and accuracy of the overall diagnosis. This module comprehensively considers the neural network optimization architecture, optimized diagnostic path, heterogeneous data feature set, adaptive learning model and tumor dynamic tracking results, and optimizes the diagnosis and treatment strategies through the feedback adjustment mechanism. The final comprehensive diagnostic results include summarized tumor identification information, the certainty of the diagnostic path and the completeness of the feature set, providing clinicians with comprehensive and detailed diagnostic information to help make more accurate treatment decisions.

[0079] In the system evaluation and improvement module, the overall performance of the system is analyzed and adjusted through linear regression and hyperparameter optimization. Linear regression is used to analyze the relationship between system performance and key parameters, revealing which parameters have a significant impact on system performance. Hyperparameter optimization further adjusts these key parameters to improve the overall performance and efficiency of the system. Through these analysis and adjustment processes, the module generates a system performance optimization record, including performance evaluation indicators, adjustment records of key parameters, and optimized configuration plans. This record not only provides a detailed evaluation of system performance, but also guides the continuous improvement and optimization of the system, ensuring the efficiency and reliability of the neuro-tumor recognition system in clinical applications.

[0080] See also Figure 2 and Figure 3 ,The network architecture search module includes an architecture generation submodule, a first performance evaluation submodule, and an architecture optimization submodule;

[0081] The architecture generation submodule uses gene expression programming based on medical image datasets to simulate genetic operations in the natural selection process, screen the key matching network architectures as the initial candidate set, and perform preliminary configuration of the architecture to generate the initial architecture candidate set;

[0082] The first performance evaluation submodule uses deep deterministic policy gradient based on the initial architecture candidate set to evaluate the performance of the network architecture by building a proxy model and interacting with the environment. It focuses on the recognition accuracy and computational efficiency of the architecture on the target task and generates architecture performance evaluation results.

[0083] The architecture optimization submodule uses a simulated annealing algorithm based on the architecture performance evaluation results, randomly selects network parameters, calculates performance changes, optimizes according to a preset cooling plan, and iteratively optimizes potential architectures to generate a neural network optimized architecture.

[0084] In the architecture generation submodule, an initial architecture candidate set is generated based on the medical image dataset through gene expression programming (GEP). First, the medical image dataset is converted into a format that can be processed by GEP, for example, MRI and CT scan images are converted into feature vectors or image matrices. Next, GEP simulates genetic operations in the natural selection process, such as crossover, mutation, and selection, to generate and optimize the network architecture. In this process, each network architecture is regarded as an "individual" whose fitness is based on how to effectively process and recognize medical image data. GEP decides which architectures should be retained and which should be modified by evaluating the efficiency and accuracy of these architectures in processing data. This process is repeated until a set of efficient initial network architectures are found as candidate sets. These candidate sets include basic configurations such as the hierarchical structure of the neural network and the type of activation function. The generated initial architecture candidate set provides a network structure that efficiently processes neuro-tumor image data.

[0085] In the first performance evaluation submodule, the deep deterministic policy gradient (DDPG) is used to evaluate the performance of the initial architecture candidate set. DDPG evaluates the performance of the network architecture by building a proxy model and interacting with the environment. In this process, each architecture candidate is input into the proxy model, and the model evaluates its performance in processing medical image data by interacting with the simulation environment. The main focus includes the accuracy and computational efficiency of the architecture in identifying neuro-tumor tasks. The DDPG algorithm optimizes the performance evaluation process by continuously and iteratively updating the strategy to ensure a more accurate evaluation of the actual performance of each architecture. Through this process, the generated architecture performance evaluation results describe in detail the performance indicators of each architecture, including recognition accuracy and processing speed, providing a basis for selecting the best architecture.

[0086] In the architecture optimization submodule, the simulated annealing algorithm is used to optimize the network architecture based on the performance evaluation results. The simulated annealing algorithm finds the optimal solution by randomly selecting network parameters and calculating performance changes. Specifically, the algorithm starts from the current solution and randomly changes certain parameters of the network architecture (such as the number of layers, the type of activation function) each time, and then calculates the impact of these changes on performance (such as accuracy and efficiency). If the change leads to performance improvement, the new solution is accepted; if the performance decreases, the new solution is also accepted according to a probability value, which decreases as the "temperature" parameter decreases. The cooling plan is the core of the algorithm, which allows the algorithm to explore a wider solution space in the initial stage, and then gradually focus on high-performance solutions. After many iterations, the algorithm generates an optimized neural network architecture that can more effectively handle neural tumor recognition tasks.

[0087] Assume that the input is a series of MRI and CT scan images, which are first converted into feature vectors. For example, each image is converted into a 1000-dimensional vector. In the architecture generation submodule, the GEP algorithm generates multiple neural network architectures with different numbers of layers and activation function types. For example, one architecture is a three-layer convolutional neural network, each layer is followed by a ReLU activation function. In the first performance evaluation submodule, the DDPG algorithm evaluates the accuracy and computational efficiency of these architectures when processing a set of image data containing labeled neural tumors. Suppose that the algorithm finds that a specific five-layer network structure performs best on this batch of data, with a recognition accuracy of 90% and a processing time of 0.5 seconds per image. Finally, in the architecture optimization submodule, the simulated annealing algorithm fine-tunes the five-layer network, adjusts the configuration and parameters of the layers, and finally obtains an optimized architecture with a recognition accuracy of 92% and a processing time of 0.4 seconds. The optimized architecture is the model used by the neural tumor recognition system.

[0088] See also Figure 2 and Figure 4 ,The diagnosis path optimization module includes the decision path tracking submodule, the key feature analysis submodule, and the path optimization submodule;

[0089] The decision path tracking submodule is based on the neural network optimization architecture and adopts the run test method. It analyzes the distribution of continuous variables in the model prediction, identifies the statistical laws and trends in the decision path, and parses the model decision path to generate a decision path analysis record.

[0090] The key feature analysis submodule uses principal component analysis based on decision path analysis records, optimizes data dimensions by transforming data into a new coordinate system, highlights key variances in the data set, mines and analyzes the key features of the model, and generates key feature analysis results;

[0091] The path optimization submodule is based on the key feature analysis results and adopts the decision tree algorithm to build a tree structure model, map the relationship between features and output variables, adjust the decision path of the model, optimize the path intuitiveness and feasibility, and generate an optimized diagnostic path.

[0092] In the decision path tracing submodule, operations are performed through the neural network optimization architecture and the runs test method. First, the neural network optimization architecture trains the decision model. The data format used is usually multidimensional numerical data, which are patients' medical records, laboratory test results, etc. The runs test method is used to analyze the distribution of continuous variables in the neural network model prediction and identify the statistical laws and trends therein. Specifically, by calculating and comparing the length and frequency of the sequence (run) of continuous similar results, the randomness of the model prediction is judged. Through this analysis, key patterns in the model decision path can be identified and parsed, such as the association between specific parameters or input data and specific prediction results. The generated decision path analysis record shows in detail the statistical characteristics and decision path of the model prediction, which plays an important role in understanding the decision-making mechanism of the model and improving the interpretability of the model.

[0093] In the key feature analysis submodule, based on the decision path analysis records, the data is processed by principal component analysis (PCA). PCA is a data dimensionality reduction technology that converts data to a new coordinate system through linear transformation so that the data variance on the first coordinate axis after transformation is maximized, and so on. In this process, the dimension of the original data set is optimized, highlighting the most important variance in the data, namely the key features. Through this transformation, the submodule is able to mine and analyze the key features predicted by the model, including the medical indicators that have the greatest impact on the predicted results. The generated key feature analysis results provide in-depth insights into the key drivers of model predictions, which are critical for improving the accuracy of model predictions and providing support for clinical decision-making.

[0094] In the path optimization submodule, the decision tree algorithm is used based on the key feature analysis results. The decision tree is a tree structure model that maps the relationship between features and output variables by constructing a tree. First, the branching conditions of the tree are determined based on the key feature analysis results, such as using key medical indicators as the basis for node splitting. Then, the relationship between these key features and the final diagnosis results is mapped through the decision tree model. In this process, the decision path of the model is also adjusted to optimize the intuitiveness and feasibility of the path. The generated optimized diagnostic path provides doctors with a clearer and more reasonable decision-making framework to help them make more accurate diagnoses in complex clinical situations.

[0095] Assume that the system collects the following data items and their simulated values: magnetic resonance imaging (MRI) scan results, patient age (45 years old), and symptom descriptions (headache, blurred vision). In the decision path tracing submodule, these data are input into the neural network model, and the statistical characteristics of the model prediction are analyzed by the runs test method. The key feature analysis submodule uses principal component analysis to reveal the features of MRI results and symptom descriptions that have the greatest impact on the diagnostic results. Finally, in the path optimization submodule, the decision tree algorithm generates an optimized diagnostic path based on these key features. For example, first determine whether a tumor exists based on the MRI results, and then further refine the diagnosis based on the symptom description. This process makes the identification of neural tumors more accurate and improves the efficiency of diagnosis.

[0096] See also Figure 2 and Figure 5 ,The data self-supervised learning module includes data integration submodule, feature extraction submodule, and feature fusion submodule;

[0097] The data integration submodule uses a multi-source data fusion algorithm based on the optimized diagnostic pathway to synchronize medical imaging data of different formats and sources, including MRI and CT scan data, perform standardized processing and quality control, verify data consistency and integrity, and generate a pooled imaging data set;

[0098] The feature extraction submodule uses convolutional neural networks and contrastive learning algorithms based on the collected image data sets to construct a multi-level feature extraction network to extract key features from the image data layer by layer from basic textures to salient patterns, and generate an extracted feature data set;

[0099] The feature fusion submodule uses a Gaussian mixture model based on the extracted feature data set to model the probability distribution of feature points, aggregate information from multiple perspectives and sources, cluster related features, identify and fuse related features, and generate a heterogeneous data feature set.

[0100] In the data integration submodule, medical imaging data are synchronized and standardized through a multi-source data fusion algorithm. This process involves medical imaging data in various formats and sources, such as MRI and CT scans. First, these data in different formats are converted into a unified format for easy processing and analysis. Then, standardization processing is performed, such as adjusting the image size and pixel intensity to ensure comparability between data. In the quality control link, the consistency and integrity of the data are checked, such as confirming that the image data is not missing or damaged. This series of operations ensures that medical imaging data from different sources and formats can be accurately and effectively fused. The generated aggregated image data set provides a standardized and high-quality data foundation for subsequent feature extraction and analysis.

[0101] In the feature extraction submodule, based on the collected image data set, convolutional neural network (CNN) and contrastive learning algorithm are used. The convolutional neural network constructs a multi-level feature extraction network to extract key features in the image data layer by layer, from basic texture to prominent pattern. Each layer in the CNN hierarchical structure can extract features of different levels, such as edges, textures, etc. The contrastive learning algorithm further enhances the feature extraction capability. By comparing the differences between different images, the model's ability to recognize key features is strengthened. The feature data set generated by this process contains the key feature information in the image data, providing a rich source of information for subsequent feature fusion and disease diagnosis.

[0102] In the feature fusion submodule, the Gaussian mixture model is used for feature fusion based on the extracted feature data set. The Gaussian mixture model can effectively aggregate information from multiple perspectives and different sources by modeling the probability distribution of feature points. In this process, the probability distribution of each feature point is analyzed, similar or related feature clusters are identified, and they are fused through clustering methods. This operation enables the information obtained from different images and perspectives to be comprehensively considered, enhancing the comprehensiveness and representativeness of the data. The generated heterogeneous data feature set provides more comprehensive and detailed feature information for in-depth analysis and diagnosis.

[0103] Assume that the system collects MRI and CT scan data, including information such as tumor size (3 cm in diameter) and location (right side of the brain). In the data integration submodule, these data are standardized and quality controlled to form a unified image data set. In the feature extraction submodule, detailed features about the tumor, such as edge clarity and texture features, are extracted from the image data through CNN and contrastive learning algorithms. Finally, in the feature fusion submodule, these features are fused through the Gaussian mixture model to generate a detailed heterogeneous data feature set, which provides comprehensive feature information for the diagnosis of neural tumors.

[0104] See also Figure 2 and Figure 6,The continuous learning and matching module includes an online learning submodule, a model adaptation submodule, and a data processing submodule;

[0105] The online learning submodule uses online stochastic gradient descent based on heterogeneous data feature sets, receives and analyzes new medical imaging data in real time, updates the model, and dynamically adjusts the learning model to match the changes in data streams and new features, generating an online updated learning model.

[0106] The model adaptation submodule uses a support vector machine based on an online updated learning model to monitor the differences between data features and model outputs in real time, adjust the hyperparameters of the support vector machine, automatically adjust the model parameters to match the new data characteristics, and generate an adaptively adjusted model.

[0107] The data processing submodule uses a recursive neural network based on an adaptively adjusted model to serialize newly acquired medical imaging data, extract time-related features, optimize data format and quality, and generate an adaptive learning model.

[0108] In the online learning submodule, the online stochastic gradient descent algorithm is used to process heterogeneous data feature sets to achieve real-time updating of the learning model. This submodule targets a variety of medical imaging data, such as MRI and CT scans. These data are first converted into feature vector form for algorithm processing. The online stochastic gradient descent algorithm processes new medical imaging data immediately upon receiving it, and continuously adjusts and optimizes the learning model. The algorithm gradually adjusts model parameters, such as weights and biases, to reduce prediction errors by calculating the gradient of new data relative to the current model. This real-time update mechanism enables the model to quickly adapt to changes in data streams and the emergence of new features. The online updated learning model generated by this process can continuously adapt to new data and maintain its accuracy and efficiency in neuro-tumor recognition.

[0109] In the model adaptation submodule, support vector machine (SVM) is used to achieve adaptive adjustment of the model. Based on the online updated learning model, this module adjusts the hyperparameters of SVM by real-time monitoring the differences between data features and model outputs. Specifically, the algorithm analyzes the difference between the feature distribution of newly accessed data and the output of the current model, and improves the performance of the model by adjusting hyperparameters such as kernel function type and penalty coefficient. This automatic adjustment mechanism ensures that the model can still maintain efficient recognition capabilities when facing new or changing data characteristics. The generated adaptive adjustment model can therefore better adapt to the changing data characteristics and improve the accuracy and reliability of neural tumor identification.

[0110] In the data processing submodule, a recurrent neural network (RNN) is used to serialize the newly received medical imaging data. This process focuses on extracting time-related features, which is particularly critical for tracking the development of tumors. For example, RNN can analyze a series of continuous MRI scan images and identify the changing trends of tumor size and morphology over time. In this process, the recurrent neural network processes serialized data through its unique loop structure and continuously updates its internal state to reflect the characteristics of the latest data. The adaptive learning model finally generated by this module optimizes the data format and quality, and improves the ability of the neural tumor recognition system to process time series medical images.

[0111] Assuming that the input is a series of time-series MRI scan images, and assuming that each MRI image is converted into a 2000-dimensional feature vector containing specific texture and shape features, the online learning submodule immediately uses the online stochastic gradient descent algorithm to update the model parameters after receiving the new MRI image (for example, the feature vector value is [0.8, 0.3, ..., 0.5]) to better adapt to these new data. The model adaptation submodule detects that the feature distribution of the new data is different from that of the old data, and adjusts the hyperparameters of the SVM (such as adjusting the kernel function type from a linear kernel to a radial basis kernel) to better match the characteristics of the new data. The data processing submodule analyzes the time series of these images through RNN to identify the characteristics of the tumor changing over time (such as the gradual increase in tumor size). Finally, the system generates an optimized adaptive learning model that can not only accurately identify current neural tumors, but also predict their future development trends.

[0112] See also Figure 2 and Figure 7 ,The dynamic tumor tracking module includes a deformation analysis submodule, a growth prediction submodule, and a tracking result integration submodule;

[0113] The deformation analysis submodule is based on an adaptive learning model and uses an optical flow algorithm to track the spatial position changes of the tumor by calculating the displacement vectors of pixels between consecutive image frames, analyze the temporal changes of pixels, determine the movement trajectory of the tumor, and generate tumor trajectory analysis results;

[0114] The growth prediction submodule uses deformation field theory based on the tumor trajectory analysis results to construct a morphological change model of the tumor in continuous image frames, analyze the deformation and growth trend of the tumor, and make trend predictions to generate tumor change and growth prediction results;

[0115] The tracking result integration submodule extracts the tumor movement trajectory and morphological change data based on the tumor change and growth prediction results, uses data fusion technology to summarize the extracted information, analyzes the dynamic changes of the tumor, and establishes a tumor change view to generate tumor dynamic tracking results.

[0116] In the deformation analysis submodule, the optical flow algorithm is used to analyze the medical imaging data provided by the adaptive learning model to track and analyze the spatial position changes of the tumor. The main task of the optical flow algorithm in this process is to calculate the displacement vectors of pixels between consecutive image frames. First, the algorithm receives a series of time series medical imaging data from the adaptive learning model, such as continuous MRI or CT scan images. Each image frame is converted into a pixel matrix. The optical flow algorithm then calculates the movement of corresponding pixels in adjacent image frames. This process involves estimating the movement speed and direction of each pixel between consecutive frames. By analyzing these displacement vectors, the algorithm can determine the movement trajectory and speed of the tumor in space. The generated tumor trajectory analysis results include not only the precise position of the tumor in each image frame, but also its movement speed and direction, providing key information for further growth prediction and treatment planning.

[0117] In the growth prediction submodule, deformation field theory is used to predict tumor growth trends based on the tumor trajectory analysis results generated by the optical flow algorithm. The deformation field theory is used to construct a morphological change model of the tumor in continuous image frames. Specifically, the algorithm first analyzes the shape and size changes of the tumor in each image frame, and then constructs a mathematical model to describe these changes. The model takes into account the deformation mode of the tumor, such as expansion, contraction or shape distortion, and predicts its future change trend. Through this analysis, the generated tumor change and growth prediction results provide important information about the future development of the tumor, such as the expected size growth rate and shape change trend, which is crucial for planning patient treatment strategies.

[0118] In the tracking result integration submodule, the motion trajectory and morphological change information of the tumor are integrated through data fusion technology. The task of this module is to extract and summarize the motion and deformation data of the tumor in continuous image frames. The algorithm first receives the morphological change data of the tumor, such as changes in size and shape, from the growth prediction submodule, and then combines it with the motion trajectory data provided by the deformation analysis submodule. Through data fusion, the algorithm can construct a comprehensive view of the dynamic changes of the tumor, including the movement trajectory, speed, morphological changes and predicted growth trends of the tumor. The generated dynamic tumor tracking results provide doctors with an intuitive overview of tumor behavior, help better understand the development process of the tumor, and provide support for subsequent treatment decisions.

[0119] Suppose a set of MRI scan sequences, which includes scans of the same patient every day for one month. Each MRI image is converted into a 5000*5000 pixel matrix, and each pixel represents the different density of the tumor and surrounding tissue. The deformation analysis submodule applies the optical flow algorithm to calculate the displacement of pixels between adjacent image frames, and finds that the tumor gradually moves to the upper right in space, with an average speed of 0.5 mm per day. The growth prediction submodule uses deformation field theory to analyze the morphological changes of the tumor and predicts that the tumor will grow at a rate of 2 mm per week. The tracking result integration submodule summarizes these data and constructs a dynamic view that includes tumor movement trajectory, morphological changes and growth predictions, providing doctors with a comprehensive tumor behavior analysis to assist in formulating more effective treatment plans.

[0120] See also Figure 2 and Figure 8 ,The iterative image processing module includes a preliminary segmentation submodule, a refinement analysis submodule, and an iterative recognition submodule;

[0121] The preliminary segmentation submodule uses threshold segmentation based on the dynamic tracking results of the tumor. It analyzes the pixel distribution in the image, sets a matching grayscale threshold, distinguishes the tumor area from the background, and identifies the tumor area to generate preliminary tumor segmentation results.

[0122] The refinement analysis submodule uses morphological segmentation based on the preliminary tumor segmentation results. By applying morphological dilation and erosion operations, it analyzes and defines the edge and structure of the tumor, optimizes the segmentation accuracy, and generates a refined tumor segmentation result.

[0123] The iterative recognition submodule uses a convolutional neural network based on the tumor segmentation results to identify and classify tumor features through the network structure, iteratively optimizes the recognition accuracy, and extracts the features of the tumor area layer by layer to generate targeted tumor recognition results.

[0124] In the preliminary segmentation submodule, the tumor dynamic tracking results are processed by threshold segmentation technology to identify and separate the tumor area. This step first receives the medical image data from the tumor dynamic tracking module. These data are usually grayscale images, in which the pixel intensities of the tumor and surrounding tissue are different. The core of threshold segmentation is to set a grayscale threshold to divide the pixels in the image into two categories: one is the tumor area, and the other is the non-tumor background. The algorithm determines an optimal threshold by analyzing the pixel distribution in the image, so that the tumor area and the background are best separated. This process involves multiple threshold attempts to find the best cutting point to distinguish the tumor from the background. The generated tumor preliminary segmentation result is a preliminary identification of the tumor location and approximate shape, providing a basis for subsequent more precise segmentation.

[0125] In the refinement analysis submodule, morphological segmentation technology is used to further process the preliminary segmentation results of the tumor. Morphological segmentation mainly uses morphological operations, such as dilation and erosion, to improve the recognition and definition of the tumor edge. The dilation operation helps to connect adjacent pixels, while the erosion operation is used to remove small spikes on the segmentation edge, thereby clearly defining the edge and structure of the tumor. In this step, the algorithm carefully analyzes the relationship between each pixel and its neighborhood, and gradually optimizes the edge and shape of the tumor area by iteratively applying dilation and erosion operations. The generated tumor refinement segmentation results provide more accurate and detailed tumor edge and structure information, providing a basis for accurate identification and subsequent treatment.

[0126] In the iterative recognition submodule, a convolutional neural network (CNN) is used to perform in-depth analysis and recognition of the tumor segmentation results. Through its multi-layer structure, CNN effectively extracts and recognizes the complex features of the tumor area, such as texture, shape, and size. In this process, CNN learns the tumor features in the image layer by layer and optimizes the network weights and biases through the back-propagation algorithm to improve recognition accuracy. Through iterative optimization, CNN can gradually improve its ability to recognize tumor features and accurately classify tumor types. The generated targeted tumor recognition results contain a detailed description of the tumor type, size, shape, and malignancy, which is of great value for clinical diagnosis and treatment planning.

[0127] Suppose a set of continuous MRI image sequences containing dynamic tracking information of tumors. These images first pass through the preliminary segmentation submodule, in which each image is processed into a grayscale image. By setting the threshold, the tumor area is successfully identified and distinguished from the background. For example, the grayscale threshold of an MRI image is set to 120, the grayscale value of the tumor area is higher than this threshold, and the grayscale value of the background area is lower than this threshold. In the refinement analysis submodule, the tumor edge is more clearly defined through morphological dilation and erosion operations. Finally, in the iterative recognition submodule, CNN conducts an in-depth analysis of the refined tumor area and successfully identifies the specific type and characteristics of the tumor, such as the degree of malignancy and potential growth trend. The targeted tumor recognition results finally generated provide detailed information about the tumor and provide an important reference for doctors to formulate treatment plans.

[0128] See also Figure 2 and Fig. 9 ,The result fusion feedback module includes a result fusion submodule, a strategy adjustment submodule, and a feedback optimization submodule;

[0129] The result fusion submodule is based on the targeted tumor recognition results, adopts a multi-dimensional data fusion algorithm, combines the neural network optimization architecture, optimized diagnostic pathways, heterogeneous data feature sets, adaptive learning models and tumor dynamic tracking results, and uses data alignment, synchronization and association analysis to verify the consistency of multi-source data in time and space, and generate preliminary comprehensive diagnostic results;

[0130] The strategy adjustment submodule uses a machine learning-based classification algorithm based on the preliminary comprehensive diagnosis results, analyzes the feasibility and accuracy of the current diagnosis results, combines rule-based logical inference and historical case comparison, evaluates the existing diagnosis results, proposes a strategy adjustment plan, and generates an optimized diagnosis strategy.

[0131] The feedback optimization submodule applies an adaptive control algorithm based on the optimized diagnostic strategy, monitors system performance in real time, dynamically adjusts system parameters according to strategy evaluation results, optimizes diagnostic performance, and matches it to dynamically changing diagnostic requirements to generate comprehensive diagnostic results.

[0132] In the result fusion submodule, the targeted tumor identification results and other related data are integrated through a multidimensional data fusion algorithm to generate a preliminary comprehensive diagnosis result. This submodule first receives data from different sources, including data generated by the neural network optimization architecture, information on optimized diagnostic pathways, heterogeneous data feature sets, outputs of adaptive learning models, and tumor dynamic tracking results. These data usually exist in structured formats, such as tables, time series, or image matrices. The key task of the multidimensional data fusion algorithm is to effectively combine data from these different formats and sources. The algorithm first aligns and synchronizes the data to ensure the consistency of different data in time and space. Subsequently, it performs association analysis to explore potential connections between different data, such as the relationship between diagnostic pathways and tumor feature sets. Through this process, the algorithm can integrate information from different sources to generate preliminary comprehensive diagnosis results. The results provide a basis for subsequent strategy adjustments and system optimization.

[0133] In the strategy adjustment submodule, a machine learning-based classification algorithm is used to analyze and evaluate the preliminary comprehensive diagnostic results. The purpose of this submodule is to verify the feasibility and accuracy of the current diagnostic results and propose an estimated strategy adjustment plan. The machine learning classification algorithm is used here to process the comprehensive diagnostic results and identify patterns and anomalies. In addition, combined with rule-based logical inference and historical case comparison, the algorithm evaluates the effectiveness of existing diagnostic results. For example, the algorithm identifies the similarity between certain diagnostic results and historical successful cases, or finds that certain results do not match the known diagnostic logic. Based on these analyses, the generated optimized diagnostic strategy provides clear guidance for adjusting diagnostic methods and processes, which helps to improve diagnostic accuracy and efficiency.

[0134] In the feedback optimization submodule, an adaptive control algorithm is used to implement and optimize the optimized diagnostic strategy. This submodule adjusts system parameters by real-time monitoring of system performance to adapt to changing diagnostic needs. The adaptive control algorithm dynamically adjusts key parameters such as algorithm configuration and hardware resource allocation based on the strategy evaluation results. For example, if it is found that the system performance degrades when processing a specific type of tumor image, the algorithm will increase the resource allocation for processing this type of image. This real-time and dynamic optimization mechanism enables the system to continuously adapt to new diagnostic challenges and needs. The generated comprehensive diagnostic results not only include the latest tumor diagnostic information, but also reflect the latest status of system performance and optimization adjustments.

[0135] Consider a complex diagnostic case involving multiple data, including a set of MRI and CT scan images, each with a resolution of 512x512 pixels, and corresponding diagnostic pathway information provided by a neural network optimization architecture. After receiving this data, the result fusion submodule combines the image data with the diagnostic pathway information using a multidimensional data fusion algorithm to generate a preliminary comprehensive diagnostic report showing the location, size, and estimated type of the tumor. The strategy adjustment submodule then analyzes this report using a machine learning algorithm, compares it with historical cases, and finds patterns similar to certain successful treatment cases, and adjusts the diagnostic strategy accordingly. Finally, the feedback optimization submodule optimizes the system configuration in real time based on these adjustments, such as increasing image processing resources for specific types of tumors. The final comprehensive diagnostic result includes detailed tumor information and adjusted system performance status.

[0136] See also Figure 2 and Fig.10 ,The system evaluation and improvement module includes a second performance evaluation submodule, a result analysis submodule, and a system optimization submodule;

[0137] The second performance evaluation submodule uses linear regression based on the comprehensive diagnosis results to perform statistical analysis on the data between system performance and key parameters, mine patterns and trends in the data, identify key factors affecting system performance, and generate performance parameter correlation analysis results;

[0138] The result analysis submodule uses the K-means algorithm to divide the parameters into multiple clusters based on the performance parameter correlation analysis results, explores the impact of multi-parameter combinations on system performance, analyzes the relationship between key parameters and system performance, and generates system performance analysis results;

[0139] Based on the system performance analysis results, the system optimization submodule uses grid search and Bayesian optimization to adjust key parameters in the system, explore the parameter space, adjust the search strategy, optimize the parameter configuration and system performance, and generate system performance optimization records.

[0140] In the second performance evaluation submodule, linear regression is used to perform statistical analysis on the comprehensive diagnostic results of the neuro-tumor recognition system to identify key factors affecting system performance. This process first involves collecting and organizing data from the comprehensive diagnostic results. These data include key performance indicators such as the running time of each module of the system, recognition accuracy, and the amount of image data processed. The linear regression model is used here to analyze the relationship between these performance indicators and the key parameters of the system. By constructing a linear relationship between parameters and performance indicators, the algorithm can identify which parameters have a significant impact on system performance. For example, the algorithm finds that the image processing speed is strongly correlated with the configuration parameters of a specific algorithm. The generated performance parameter correlation analysis results provide an important basis for system performance optimization and point out the direction of improvement.

[0141] In the result analysis submodule, the K-means algorithm is used to further mine and analyze the performance parameter association analysis results. The K-means algorithm identifies the impact of different parameter combinations on system performance by dividing the parameters into several clusters. The process begins by inputting the data points in the performance parameter association analysis results (i.e., various parameter configurations and their corresponding performance indicators) into the K-means algorithm. The algorithm attempts to group these data points into different clusters, each cluster representing a specific parameter combination pattern. By analyzing the characteristics of the data points in each cluster, K-means reveals which parameter combinations have a significant impact on improving or degrading system performance. The generated system performance analysis results provide in-depth insights and guide how to adjust parameter configurations to optimize overall system performance.

[0142] In the system optimization submodule, grid search and Bayesian optimization are combined to fine-tune key parameters to optimize system performance. First, grid search systematically explores different parameter combinations in a predefined parameter space and evaluates the impact of each combination on system performance. Although this method is time-consuming, it can comprehensively evaluate all estimated parameter combinations. Then, Bayesian optimization further optimizes parameter selection based on the results of grid search. Bayesian optimization uses a probabilistic model to predict the performance of different parameter combinations and selects the combination that can best improve performance for testing. This method is more efficient and can quickly find the optimal solution. Through the combination of these two methods, the system optimization submodule can accurately adjust parameters. The generated system performance optimization record not only includes the optimized parameter configuration, but also detailed guidelines on how to adjust parameters for optimal performance.

[0143] Assuming the data contained in the comprehensive diagnosis results, such as image processing time (0.8 seconds per image on average), recognition accuracy (92%) and the amount of image data processed (about 1,000 images per day), the second performance evaluation submodule uses linear regression to analyze the relationship between these indicators and parameters (such as algorithm configuration, hardware resource allocation). The results show that the processing time is closely related to the algorithm configuration. In the result analysis submodule, the K-means algorithm divides the parameter configuration into three clusters, each cluster showing a different performance pattern. The system optimization submodule adjusts the parameters through grid search and Bayesian optimization, and finally finds a set of optimized parameter configurations, which reduces the image processing time to 0.6 seconds while maintaining the recognition accuracy above 92%. The generated system performance optimization record records the parameter adjustment process and the final configuration in detail, providing a basis for continuous improvement of the system.

[0144] In the description of this specification, the description with reference to the terms "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0145] The preferred embodiments of the present invention disclosed above are only used to help illustrate the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific implementation methods described. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A neural tumor recognition system based on artificial intelligence, characterized in that: The system includes a network architecture search module, a diagnostic path optimization module, a data self-supervised learning module, a continuous learning matching module, a dynamic tumor tracking module, an iterative image processing module, a result fusion feedback module and a system evaluation and improvement module; The network architecture search module uses a genetic algorithm based on a medical image dataset to initially screen the network architecture, uses a deep Q network to perform performance evaluation and parameter adjustment on the screened architecture, captures the network structure that matches the needs of neuro-tumor identification, and performs iterative architecture optimization to generate a neural network optimized architecture; The diagnostic path optimization module is based on a neural network optimization architecture and uses the SHAP algorithm to quantitatively analyze model decision factors, use the model local interpretability method to locally analyze the model's predictions, and make the model's decision path transparent to generate an optimized diagnostic path. The data self-supervised learning module is based on optimizing the diagnostic path, applying a contrastive learning algorithm, analyzing multi-source medical imaging data including MRI and CT scanning equipment, using a clustering algorithm to classify and fuse data features, and constructing and optimizing feature sets to generate heterogeneous data feature sets; The continuous learning and matching module uses a support vector machine based on a heterogeneous data feature set to perform real-time updates and self-matching of the model, combines a recursive neural network to process newly accessed medical imaging data, continuously updates and optimizes the model, and generates an adaptive learning model; The dynamic tumor tracking module is based on an adaptive learning model and uses an optical flow algorithm to track the movement trajectory of the tumor in the sequence images. It uses deformation field theory to analyze the morphological changes and growth trends of the tumor, analyze and predict the dynamic changes of the tumor, and generate dynamic tumor tracking results. The iterative image processing module uses threshold segmentation based on the tumor dynamic tracking results to preliminarily divide the potential tumor area in the image, combines morphological segmentation and convolutional neural network to mine and analyze tumor characteristics, divide and identify the tumor area, and generate targeted tumor recognition results; The result fusion feedback module is based on the targeted tumor identification results, combined with the neural network optimization architecture, optimized diagnostic pathway, heterogeneous data feature set, adaptive learning model and tumor dynamic tracking results, and uses the feedback mechanism to adjust the diagnosis and treatment strategies, organize and optimize the results, and generate a comprehensive diagnosis result; The system evaluation and improvement module is based on the comprehensive diagnosis results, uses linear regression to analyze the correlation between system performance and parameters, uses hyperparameter optimization to adjust the key parameters of the identification system, optimizes the system configuration and adjusts the parameters, and performs system performance evaluation and optimization processing to generate system performance optimization records.

2. The artificial intelligence-based neural tumor recognition system according to claim 1, characterized in that: The neural network optimization architecture includes a hierarchical structure of neural tumor recognition, activation function type selection and connection weight configuration, the optimized diagnostic path includes key feature recognition, decision node positioning and path transparency indicators, the heterogeneous data feature set includes texture features, edge features and shape descriptors of MRI and CT images, the tumor dynamic tracking results include tumor position changes, size changes and predicted growth trends, the targeted tumor recognition results include tumor positioning, boundary division and classification information, the comprehensive diagnosis results include summarized tumor recognition information, diagnostic path certainty and feature set integrity, and the system performance optimization records include performance evaluation indicators, key parameter adjustment records and optimized configuration plans.

3. The artificial intelligence-based neural tumor recognition system according to claim 1, characterized in that: The network architecture search module includes an architecture generation submodule, a first performance evaluation submodule, and an architecture optimization submodule; The architecture generation submodule uses gene expression programming based on the medical image dataset to simulate the genetic operation in the natural selection process, screen the matching key network architecture as the initial candidate set, and perform preliminary configuration of the architecture to generate the initial architecture candidate set; The first performance evaluation submodule uses deep deterministic policy gradient based on the initial architecture candidate set to evaluate the performance of the network architecture by building a proxy model and interacting with the environment, focusing on the recognition accuracy and computational efficiency of the architecture on the target task, and generating an architecture performance evaluation result; The architecture optimization submodule uses a simulated annealing algorithm based on the architecture performance evaluation results, randomly selects network parameters, calculates performance changes, optimizes according to a preset cooling plan, and iteratively optimizes potential architectures to generate a neural network optimization architecture.

4. The artificial intelligence-based neural tumor recognition system according to claim 1, characterized in that: The diagnostic path optimization module includes a decision path tracking submodule, a key feature analysis submodule, and a path optimization submodule; The decision path tracking submodule is based on a neural network optimization architecture and adopts a run test method to analyze the distribution of continuous variables in model prediction, identify statistical laws and trends in the decision path, and parse the model decision path to generate a decision path analysis record; The key feature analysis submodule uses principal component analysis based on decision path analysis records, optimizes data dimensions by transforming data into a new coordinate system, highlights key variances in the data set, mines and analyzes key features of the model, and generates key feature analysis results; The path optimization submodule uses a decision tree algorithm based on the key feature analysis results to build a tree structure model, map the relationship between features and output variables, and adjust the decision path of the model to optimize the path intuitiveness and feasibility, and generate an optimized diagnostic path.

5. The artificial intelligence-based neural tumor recognition system according to claim 1, characterized in that: The data self-supervised learning module includes a data integration submodule, a feature extraction submodule, and a feature fusion submodule; The data integration submodule uses a multi-source data fusion algorithm based on an optimized diagnostic pathway to synchronize medical imaging data of different formats and sources, including MRI and CT scan data, to perform standardized processing and quality control, verify data consistency and integrity, and generate a pooled imaging data set; The feature extraction submodule uses a convolutional neural network and a contrastive learning algorithm based on the collected image data set to construct a multi-level feature extraction network to extract key features from the image data layer by layer from basic texture to salient pattern, and generate an extracted feature data set; The feature fusion submodule uses a Gaussian mixture model based on the extracted feature data set to model the probability distribution of feature points, collect information from multiple perspectives and sources, cluster related features, identify and fuse related features, and generate a heterogeneous data feature set.

6. The artificial intelligence-based neural tumor recognition system according to claim 1, characterized in that: The continuous learning and matching module includes an online learning submodule, a model adaptation submodule, and a data processing submodule; The online learning submodule adopts online stochastic gradient descent based on heterogeneous data feature sets, receives and analyzes new medical imaging data in real time, performs model update processing, and dynamically adjusts the learning model to match the changes and new features of the data stream, thereby generating an online updated learning model; The model adaptation submodule uses a support vector machine based on an online updated learning model to monitor the difference between data features and model outputs in real time, adjust the hyperparameters of the support vector machine, automatically adjust the model parameters to match the new data characteristics, and generate an adaptively adjusted model; The data processing submodule uses a recursive neural network based on an adaptively adjusted model to serialize newly accessed medical imaging data, extract time-related features, optimize data format and quality, and generate an adaptive learning model.

7. The artificial intelligence-based neural tumor recognition system according to claim 1, characterized in that: The dynamic tumor tracking module includes a deformation analysis submodule, a growth prediction submodule, and a tracking result integration submodule; The deformation analysis submodule is based on an adaptive learning model and uses an optical flow algorithm to track the spatial position change of the tumor by calculating the displacement vector of the pixel points between consecutive image frames, analyze the time dimension change of the pixel points, determine the movement trajectory of the tumor, and generate the tumor trajectory analysis result; The growth prediction submodule uses deformation field theory based on the tumor trajectory analysis results to construct a morphological change model of the tumor in continuous image frames, analyzes the deformation and growth trend of the tumor, and performs trend prediction to generate tumor change and growth prediction results; The tracking result integration submodule extracts tumor movement trajectory and morphological change data based on tumor change and growth prediction results, uses data fusion technology to summarize the extracted information, analyzes the dynamic changes of the tumor, and establishes a tumor change view to generate tumor dynamic tracking results.

8. The artificial intelligence-based neural tumor recognition system according to claim 1, characterized in that: The iterative image processing module includes a preliminary segmentation submodule, a refinement analysis submodule, and an iterative recognition submodule; The preliminary segmentation submodule adopts threshold segmentation based on the dynamic tracking result of the tumor, analyzes the pixel distribution in the image, sets a matching grayscale threshold, distinguishes the tumor area from the background, and identifies the tumor area to generate a preliminary tumor segmentation result; The refinement analysis submodule adopts morphological segmentation based on the preliminary segmentation result of the tumor, analyzes and defines the edge and structure of the tumor by applying morphological expansion and corrosion operations, optimizes the segmentation accuracy, and generates a refined segmentation result of the tumor; The iterative recognition submodule uses a convolutional neural network based on the tumor segmentation results to identify and classify tumor features through the network structure, iteratively optimizes the recognition accuracy, and extracts the features of the tumor area layer by layer to generate targeted tumor recognition results.

9. The artificial intelligence-based neural tumor recognition system according to claim 1, characterized in that: The result fusion feedback module includes a result fusion submodule, a strategy adjustment submodule, and a feedback optimization submodule; The result fusion submodule is based on the targeted tumor recognition results, adopts a multi-dimensional data fusion algorithm, combines the neural network optimization architecture, optimized diagnostic pathways, heterogeneous data feature sets, adaptive learning models and tumor dynamic tracking results, and uses data alignment, synchronization and association analysis to verify the consistency of multi-source data in time and space, and generate preliminary comprehensive diagnosis results; The strategy adjustment submodule uses a classification algorithm based on machine learning based on the preliminary comprehensive diagnosis results, analyzes the feasibility and accuracy of the current diagnosis results, combines rule-based logical inference and historical case comparison, evaluates the existing diagnosis results and proposes a strategy adjustment plan to generate an optimized diagnosis strategy; The feedback optimization submodule applies an adaptive control algorithm based on the optimized diagnostic strategy, monitors system performance in real time, dynamically adjusts system parameters according to strategy evaluation results, optimizes diagnostic performance, and matches it to dynamically changing diagnostic requirements to generate comprehensive diagnostic results.

10. The artificial intelligence-based neural tumor recognition system according to claim 1, characterized in that: The system evaluation and improvement module includes a second performance evaluation submodule, a result analysis submodule, and a system optimization submodule; The second performance evaluation submodule uses linear regression based on the comprehensive diagnosis results to perform statistical analysis on the data between system performance and key parameters, mine patterns and trends in the data, identify key factors affecting system performance, and generate performance parameter correlation analysis results; The result analysis submodule uses the K-means algorithm based on the performance parameter correlation analysis results to mine the influence of multi-parameter combinations on system performance by dividing the parameters into multiple clusters, analyze the relationship between key parameters and system performance, and generate system performance analysis results; Based on the system performance analysis results, the system optimization submodule uses grid search and Bayesian optimization to adjust key parameters in the system, explore parameter space, adjust search strategies, optimize parameter configuration and system performance, and generate system performance optimization records.

Images