Magnetic resonance image positioning diagnosis method based on artificial intelligence

Abstract

The invention discloses a magnetic resonance image positioning diagnosis method based on artificial intelligence, and relates to the technical field of medical image processing and diagnosis. S100, data acquisition and preprocessing, S200, model construction, S300, strategy learning and optimization, S400, personalized scanning path generation, and S500, real-time feedback and adjusting.According to the method, the lesion area of a patient is accurately covered, and scanning of the non-lesion area is reduced, so that the imaging efficiency is improved, the scanning time is shortened, and the imaging accuracy is improved. The imaging quality of a lesion area is remarkably improved, so that the diagnosis accuracy is improved, a traditional magnetic resonance scanning path often adopts a unified standard, individual requirements of different patients are difficult to meet, unclear imaging or omission of the lesion area is easily caused, and the diagnosis result is affected. According to the method, through personalized scanning path design, the defect is overcome, and a more accurate and reliable diagnosis basis is provided for doctors.

Description

Technical Field

[0001] The present invention relates to the field of medical image processing and diagnosis technology, and in particular to a magnetic resonance image positioning diagnosis method based on artificial intelligence. Background Art

[0002] With the rapid development of medical technology, medical image processing and diagnosis technology plays an increasingly important role in clinical practice. Magnetic resonance imaging, as a non-invasive medical imaging technology, has been widely used in the diagnosis of diseases in the nervous system, cardiovascular system, and skeletal muscle system with its advantages of high resolution, multi-parameter imaging, and no radiation damage.

[0003] Traditional MRI scanning path design has many shortcomings. Due to the differences in anatomical structures of different patients and the different locations of lesions, it is often difficult to accurately cover the lesion areas of all patients using a unified scanning path, resulting in poor imaging quality in the lesion area, which in turn affects the accuracy of diagnosis. This includes excessive scanning of non-lesion areas, which not only increases the scanning time, but may also cause patient discomfort and additional radiation exposure. Therefore, it is particularly important to develop an artificial intelligence-based MRI image positioning diagnosis method. Summary of the invention

[0004] The purpose of the present invention is to make up for the deficiencies of the prior art and to provide an artificial intelligence-based magnetic resonance image positioning diagnosis method, which automatically learns and optimizes the scanning path through an intelligent algorithm to meet the patient's personalized needs, improve diagnostic efficiency and accuracy, and reduce unnecessary radiation exposure.

[0005] In order to solve the above technical problems, the present invention provides the following technical solutions: a magnetic resonance image positioning diagnosis method based on artificial intelligence, the specific steps of the method are:

[0006] S100, data collection and preprocessing: collect the patient's medical imaging data and related clinical information, including but not limited to the patient's age, gender, and medical history, preprocess the image data, and extract the patient's anatomical structure and key information on the lesion location;

[0007] S200, model construction: design an optimization model that takes the patient's anatomical structure, lesion location, and imaging requirements as input states, and defines the action space and reward function so that the model can learn the optimal scanning path strategy through continuous exploration;

[0008] S300, strategy learning and optimization: During the training phase, the model continuously adjusts the scanning strategy through multiple trials and errors in simulation and actual environments in order to maximize the reward function value. Through technology, the model's ability to understand complex image features and strategy optimization capabilities are enhanced;

[0009] S400, Personalized Scan Path Generation: After sufficient training, the model can quickly generate a personalized MRI scan path for the patient based on the input patient-specific information. The path can accurately cover the lesion area while minimizing the scanning of non-lesion areas, thereby improving imaging efficiency and diagnostic accuracy.

[0010] S500, real-time feedback and adjustment: During the actual scanning process, the system can fine-tune the scanning path based on the real-time image quality feedback. The collected data can be used to further train and optimize the model, forming a virtuous circle.

[0011] Furthermore, the specific steps of S100 are:

[0012] S101, determine the source of medical images, obtain patient consent and authorization from hospitals and relevant institutions, comply with data protection regulations, use API to download and transmit medical imaging data from data sources, usually including images in DICOM format, design clinical information collection forms according to research needs, including but not limited to basic information about the patient's age, gender, and medical history, collect and enter clinical information through electronic medical record systems and patient interviews, verify the collected clinical information, and ensure the accuracy and completeness of the data;

[0013] S102, converting the image in DICOM format into a format for analysis and processing. If the analysis software supports DICOM, no conversion is required. Remove noise from the image, improve image quality, adjust image brightness and contrast, adjust window width and window position, make the anatomical structure and lesion area in the image clearer, perform image registration for multimodal or multi-phase images to eliminate spatial differences between different images, segment the image using automatic or semi-automatic methods, extract the patient's anatomical structure and lesion location, and annotate them. Check and correct errors, missing values ​​or abnormal values ​​in clinical information, convert text information into numerical codes for subsequent analysis, and associate the processed clinical information with the corresponding medical imaging data to form a complete data set;

[0014] S103, extract quantitative and qualitative features from the preprocessed medical images, combine the extracted image features with the patient's clinical information to form a comprehensive patient portrait, and divide the sorted data set into training set, validation set and test set to prepare for subsequent model training.

[0015] Furthermore, the specific steps of S200 are:

[0016] S201, select the algorithm according to the characteristics of the problem. For complex scanning path optimization problems, DQN and more advanced algorithms are more suitable. Design the overall architecture of the model. For image input, take the patient's medical imaging data as input, extract the key information of the anatomical structure through the preprocessing step, which can be represented in the form of feature maps, and take the annotation information of the lesion location as part of the input. This information can be used to guide the model to focus on the lesion area during the scanning process. Encode the imaging requirements into quantifiable parameters and use them as part of the input state;

[0017] S202, based on the controllable parameters of the magnetic resonance scanning device, define the actions that the model can take, combine all possible actions into an action space, and the model will select an action from the space to execute at each decision step, clarify the goal of model optimization, and build a reward function to evaluate the effect of each action based on the goal setting. The reward function should be able to reflect the performance of the model in achieving the goal, and give positive rewards when the model makes beneficial decisions, and give negative penalties when it makes unfavorable decisions;

[0018] S203, prepare training data, including a series of medical imaging data of patients, lesion location annotations and imaging requirements. These data will be used to simulate the scanning process and generate the input state and reward signal required for training. Through continuous iteration of the training process, the model will learn how to select the optimal action based on the input state to maximize the cumulative reward. In each iteration, the model will perform an action, observe the results, and update its internal parameters to improve future decisions. During the training process, the learning rate, batch size, and training round hyperparameters are adjusted according to the performance of the model to optimize the training effect of the model. Define indicators for evaluating model performance, use the test set to evaluate the model, check its performance on unseen data, and iteratively optimize the model based on the test results.

[0019] Furthermore, the specific steps of S300 are:

[0020] S301, before starting training, the parameters of the model are randomly initialized, which ensures that the model has no bias at the beginning of training and can learn from the data, sets the hyperparameters that affect the model training and performance, and builds a simulation environment based on the characteristics of the magnetic resonance scanning device and the scanning process. This environment can simulate various state changes during the scanning process, including but not limited to scanning angle, layer thickness, and imaging quality. The pre-processed medical imaging data and annotation information are used to generate simulated patient data, which will be used as input for the model in the training phase. In the simulation environment, the model starts to execute the scanning strategy. At each time step, the model selects an action based on the current state, and then the simulation environment generates a new state and reward signal based on this action.

[0021] S302, after receiving the reward signal, the model updates its internal parameters according to the reward value. If the action leads to a higher reward, the model will tend to choose the action again in the future. Conversely, if the action leads to a negative reward, the model will adjust its strategy to avoid making the same mistake again. The model uses technology to extract complex features in the image. These features help the model understand the patient's anatomical structure and lesion location more accurately. The model's strategy is optimized using optimization algorithms. These algorithms can guide the model to find a better solution in the parameter space to maximize the cumulative reward. The experience replay mechanism is used to store and replay past experience, which helps the model break time correlation and learn more stably.

[0022] S303, during the training process, regularly monitor the performance of the model on the validation set, which can be done by calculating the average reward and other evaluation indicators on the validation set. If the performance of the model on the validation set begins to stabilize or no longer improves significantly, this may mean that the model has converged to a local optimal solution or is close to a global optimal solution. At this time, training can be stopped to avoid overfitting. The trained model is deployed in an actual magnetic resonance scanning environment for testing, which involves integrating the model with the scanning equipment and ensuring the correctness of data communication and interfaces, collecting data in the actual environment, and evaluating the performance of the model in real scenarios. Based on the actual test results, the model is further iteratively optimized, including but not limited to adjusting the model architecture, improving the input data representation, and optimizing the reward function.

[0023] Furthermore, the specific steps of S400 are:

[0024] S401, collect patient-specific information from medical records, imaging data, and other related systems, including but not limited to age, gender, weight, height, past medical history, current symptoms, and preliminary diagnosis results, pre-process the medical imaging data provided by the patient, load the fully trained model from the storage medium, ensure that all weights and parameters are correctly loaded, and configure the necessary hardware and software environment to ensure that the model can run efficiently;

[0025] S402, selecting a model to extract features from the preprocessed image data, capturing key features of the morphology, size, and position of the lesion area, fusing the patient's specific information with the image features to construct a more comprehensive patient representation, and applying an optimization algorithm to plan a personalized scanning path based on the extracted and fused features. The path should be able to accurately cover the lesion area and reduce the scanning of non-lesion areas. When planning the path, the physical constraints and imaging requirements of the magnetic resonance scanning equipment are considered to ensure that the generated path is feasible and efficient in actual operation;

[0026] S403, performing a preliminary evaluation on the generated personalized scanning path to check whether it covers all key lesion areas and evaluate the amount of scanning of non-lesion areas. Based on the preliminary evaluation or expert feedback, the model can automatically adjust the parameters of the scanning path to further optimize the scanning efficiency and accuracy, and convert the finalized personalized scanning path into an instruction sequence executable by the magnetic resonance scanning device;

[0027] S404, sending the instruction sequence to the magnetic resonance scanning device to perform the scanning operation. After the scanning is completed, the obtained image data is post-processed to generate high-quality imaging results. Medical experts will diagnose and evaluate the imaging results to verify the effectiveness of the personalized scanning path. Feedback from experts and users is collected to understand the performance and problems of the personalized scanning path in actual application. Based on the feedback and new patient data, the model and optimization algorithm are continuously optimized to improve the generation capability and imaging efficiency of the personalized scanning path.

[0028] Furthermore, the S400 selection model performs feature extraction on the preprocessed image data, and the feature extraction formula is: Among them, F represents the final extracted feature set, M is a morphological weight matrix, which is adjusted according to different lesion morphological characteristics to better capture the morphological characteristics of the lesion area, and I P represents the preprocessed image data, represents a specific morphological feature extraction operation, which is used to extract the morphological features in the image. S is the size feature factor, which is determined by analyzing the pixel distribution range of the lesion area in the image and is used to highlight the size characteristics of the lesion area. C is the fusion coefficient, which is used to balance the degree of fusion between the patient's specific information and the image features. P info represents the patient's specific information vector, G is a position feature generator, which is generated by analyzing the spatial coordinate relationship in the image and is used to extract the key position features of the lesion area;

[0029] Fusion of patient-specific information with image features to construct a more comprehensive patient representation: R = I × W I +P×W P ×F(I,P), where R represents the final more comprehensive patient representation, I is the preprocessed image feature vector, and W I is the image feature weight vector, which is used to adjust the importance of image features in fusion, P is the patient-specific information vector, and W Pis the patient-specific information weight vector, which is used to control the influence of patient-specific information. F(I,P) is a fusion function, which can be a fusion network based on deep learning or a custom mathematical function. It nonlinearly fuses image features and patient-specific information to capture the complex relationship between the two, adding novelty and creativity to the fusion process.

[0030] Based on the extracted and fused features, an optimization algorithm is applied to plan a personalized scanning path: Path = argmin P (D(P)+λ1M(P)+λ2C(P)), where Path represents the planned personalized scanning path, P is a variable set of path parameters, and D(P) is determined by the overlapping area or volume of the domain to ensure that the path can accurately cover the lesion area. M(P) is the influence function on the non-lesion area, which measures the scanning degree of the path P on the non-lesion area. It should be as small as possible and can be determined by counting the number of pixels in the non-lesion area passed by the path or calculating the radiation dose received by the non-lesion area. C(P) is a constraint function, which can be expressed by a series of constraints considering the physical constraints and imaging requirements of the magnetic resonance scanning equipment. λ1 and λ2 are weight coefficients, which are used to balance the importance of the coverage of the lesion area, the impact of the non-lesion area, and the constraints, and can be adjusted according to specific needs.

[0031] Furthermore, the specific steps of S500 are:

[0032] S501, starting the magnetic resonance scanning process according to the preset personalized scanning path, the real-time image quality assessment system and the scanning path adjustment mechanism are in an activated state, ready to receive and process real-time data, and in the scanning process, capturing and processing magnetic resonance imaging data in real time, applying the image quality assessment algorithm to assess the quality of the captured image, and generating a real-time feedback signal based on the assessment result, indicating the advantages and disadvantages of the current image quality and potential improvement directions;

[0033] S502, the system receives and analyzes real-time feedback signals, identifies areas with poor image quality or improper scanning parameter settings, and formulates a scanning path fine-tuning strategy based on the feedback signals. Without affecting the overall progress of the current scanning process, the scanning path is dynamically adjusted to improve image quality. During the scanning process, in addition to image data, all relevant data such as the scanning path, adjustment records, and real-time feedback signals are collected and stored in a safe and reliable database.

[0034] Furthermore, the specific steps of S500 also include:

[0035] S503, preprocessing the collected data, incorporating new data into the model training process, updating model parameters through incremental learning and online learning, improving the accuracy and generalization ability of the model, and after updating the model, using an independent test set to evaluate the model performance to ensure that the updated model is still effective in practical applications;

[0036] S504, feeds back the results of model optimization to the actual scanning process to further verify the effect of model improvement, collects new real-time data, and continuously monitors the real-time feedback and adjustment mechanism of the system to ensure that it can work stably and reliably under various circumstances. According to the feedback and results in actual applications, the various links of the system are continuously iterated and optimized to form a more complete virtuous circle.

[0037] Compared with the prior art, a method for making bricks using waste oil sludge tailings has the following advantages:

[0038] Beneficial effects:

[0039] First, this method can accurately cover the patient's lesion area and reduce the scanning of non-lesion areas. This feature not only improves imaging efficiency and shortens scanning time, but also significantly improves the imaging quality of the lesion area, thereby improving the accuracy of diagnosis. Traditional magnetic resonance scanning paths often adopt a unified standard, which is difficult to meet the personalized needs of different patients, and easily leads to unclear or missed imaging of the lesion area, affecting the diagnosis results. The method of the present invention overcomes this drawback through personalized scanning path design, providing doctors with a more accurate and reliable diagnostic basis.

[0040] 2. This method can accurately control the scanning path and avoid excessive scanning of non-lesion areas, thereby significantly reducing the patient's radiation exposure. Although magnetic resonance imaging itself has no radiation damage, excessive scanning time and unnecessary scanning range may still cause discomfort and potential health risks to patients. By reducing unnecessary scanning, the method of the present invention not only improves the patient's comfort, but also reduces potential health risks, reflecting a high level of concern for patient safety. At the same time, reducing scanning time also helps to improve the utilization efficiency of medical resources and reduce medical costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0042] Figure 1Flow chart of the magnetic resonance image localization diagnosis method based on artificial intelligence. DETAILED DESCRIPTION

[0043] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. 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.

[0044] Embodiment 1

[0045] Download the patient's DICOM format magnetic resonance images from the hospital's PACS system, and ensure that the patient's consent and authorization from relevant institutions have been obtained. Convert the DICOM images into a format suitable for analysis and processing, remove image noise, adjust brightness and contrast, adjust the window width and window position to make the infarct area clearer, and align multimodal or multi-phase images to eliminate spatial differences.

[0046] Use automatic or semi-automatic methods to segment images, extract brain structures and possible infarct areas, and annotate them. Verify and correct errors, missing values, or outliers in clinical information. Convert text information into numerical codes. Associate the processed clinical information with medical imaging data to form a data set. Feature extraction and data set partitioning: Extract quantitative and qualitative features from preprocessed images, combine them with the patient's clinical information to form a comprehensive patient portrait, and divide it into training set, validation set, and test set.

[0047] The DQN algorithm was selected to build the model. The input included the patient's brain structure characteristics, infarct area annotations, and imaging requirements. The action space and reward function were defined: the action space was defined according to the controllable parameters of the magnetic resonance imaging equipment, and the reward function was constructed to minimize the scanning time and maximize the coverage of the infarct area.

[0048] Build a simulation environment to simulate various state changes during the scanning process, use the preprocessed data to generate simulated patient data as input, perform multiple trial and error in the simulation environment, and adjust the model strategy based on the input state and reward signal. Use experience replay and optimization algorithms to optimize model performance. Regularly monitor the performance of the model on the validation set, adjust hyperparameters, optimize the model, and evaluate the final performance on the test set.

[0049] Collect patient-specific information, load the trained model, and extract image features. The feature extraction formula is: Among them, F represents the final extracted feature set, M is a morphological weight matrix, which is adjusted according to different lesion morphological characteristics to better capture the morphological characteristics of the lesion area, and I P represents the preprocessed image data, represents a specific morphological feature extraction operation, which is used to extract the morphological features in the image. S is the size feature factor, which is determined by analyzing the pixel distribution range of the lesion area in the image and is used to highlight the size characteristics of the lesion area. C is the fusion coefficient, which is used to balance the degree of fusion between the patient's specific information and the image features. P info Represents the patient's specific information vector. G is a position feature generator, which is generated by analyzing the spatial coordinate relationship in the image. It is used to extract the key location features of the lesion area, integrate the patient's specific information, and construct a comprehensive patient representation. The formula is: R = I × W I +P×W P ×F(I,P), where R represents the final more comprehensive patient representation, I is the preprocessed image feature vector, and W I is the image feature weight vector, which is used to adjust the importance of image features in fusion, P is the patient-specific information vector, and W P is the patient-specific information weight vector, which is used to control the influence of patient-specific information. F(I,P) is a fusion function, which can be a fusion network based on deep learning or a custom mathematical function. It nonlinearly fuses image features and patient-specific information to capture the complex relationship between the two, adding novelty and creativity to the fusion process. Based on the fused features, a personalized scanning path is planned. Path = argmin P (D(P)+λ1M(P)+λ2C(P)), where Path represents the planned personalized scanning path, P is a variable set of path parameters, and D(P) is determined by the overlapping area or volume of the domain to ensure that the path can accurately cover the lesion area. M(P) is the influence function on the non-lesion area, which measures the scanning degree of the path P on the non-lesion area. It should be as small as possible and can be determined by counting the number of pixels in the non-lesion area passed by the path or calculating the radiation dose received by the non-lesion area. C(P) is a constraint function. Considering the physical constraints and imaging requirements of the magnetic resonance scanning equipment, it can be represented by a series of constraints. λ1 and λ2 are weight coefficients used to balance the importance of the coverage of the lesion area, the impact of the non-lesion area, and the constraints. They can be adjusted according to specific needs, and preliminary evaluation and adjustment are performed to ensure accurate coverage of the infarct area. The scanning path is converted into an executable instruction for the device, the scan is performed, the image is post-processed, and the imaging results are evaluated.

[0050] Capture images during the scanning process, apply image quality assessment algorithms to generate real-time feedback signals, adjust the scanning path based on the feedback signals, improve image quality, collect new data, update the model through incremental learning and online learning to improve accuracy and generalization ability, apply the optimized model to actual scanning, and continuously monitor and iteratively optimize the system.

[0051] Embodiment 2

[0052] Collect patients' DICOM format magnetic resonance images, obtain patients' clinical information including age, gender, history of hypertension and diabetes from the electronic medical record system, convert DICOM images into a format supported by the analysis software, remove image noise, improve image quality, adjust image brightness, contrast, window width and window position to highlight brain structures and lesion areas, align multimodal or multi-phase images to eliminate spatial differences, automatically or semi-automatically segment cerebrovascular areas and annotate them, verify and correct errors, missing values ​​or outliers in clinical information, convert clinical information into numerical codes, and associate them with medical imaging data to form a complete data set, extract quantitative and qualitative features from preprocessed images, combine image features with clinical information to form a patient portrait, and divide the data set into training set, validation set and test set.

[0053] Select a suitable algorithm to build the model, take the preprocessed images and lesion location annotations as input, set the imaging requirements to examine the cerebrovascular structure in detail, ensure high-resolution imaging, define the action space according to the controllable parameters of the magnetic resonance imaging equipment, and construct a reward function to encourage the model to cover the lesion area as detailed as possible and reduce the scanning of non-lesion areas.

[0054] The model is trained using preprocessed data and labeled information in a simulation environment. The model selects actions based on the current state, and the simulation environment generates new states and reward signals. The model's ability to understand complex image features is improved through optimization algorithms and technical means. The experience replay mechanism is used to stabilize learning, monitor the model's performance on the validation set, adjust hyperparameters based on performance, and optimize the model.

[0055] Load the fully trained model, collect patient-specific information, and preprocess the medical imaging data to extract the features of the lesion area in the image. Among them, F represents the final extracted feature set, M is a morphological weight matrix, which is adjusted according to different lesion morphological characteristics to better capture the morphological characteristics of the lesion area, and I P represents the preprocessed image data, represents a specific morphological feature extraction operation, which is used to extract the morphological features in the image. S is the size feature factor, which is determined by analyzing the pixel distribution range of the lesion area in the image and is used to highlight the size characteristics of the lesion area. S is the fusion coefficient, which is used to balance the degree of fusion between the patient's specific information and the image features. P info Represents the patient's specific information vector. G is a position feature generator, which is generated by analyzing the spatial coordinate relationship in the image. It is used to extract the key location features of the lesion area, fuse the patient's specific information with the image features, and construct a comprehensive patient representation. R = I × W I +P×WP ×F(I,P), where R represents the final more comprehensive patient representation, I is the preprocessed image feature vector, and W I is the image feature weight vector, which is used to adjust the importance of image features in fusion, P is the patient-specific information vector, and W P is the patient-specific information weight vector, which is used to control the influence of patient-specific information. F(I,P) is a fusion function, which can be a fusion network based on deep learning or a custom mathematical function. It nonlinearly fuses image features and patient-specific information to capture the complex relationship between the two, adding novelty and creativity to the fusion process. Based on the extracted and fused features, an optimization algorithm is applied to plan a personalized scanning path. Path = argmin P (D(P)+λ1M(P)+λ2C(P)), where Path represents the planned personalized scanning path, P is a variable set of path parameters, and D(P) is determined by the overlapping area or volume of the domain to ensure that the path can accurately cover the lesion area. M(P) is the influence function on the non-lesion area, which measures the scanning degree of the path P on the non-lesion area. It should be as small as possible and can be determined by counting the number of pixels in the non-lesion area passed by the path or calculating the radiation dose received by the non-lesion area. C(P) is a constraint function. Considering the physical constraints and imaging requirements of the magnetic resonance scanning equipment, it can be represented by a series of constraints. λ1 and λ2 are weight coefficients used to balance the importance of the coverage of the lesion area, the influence of the non-lesion area, and the constraints. They can be adjusted according to specific needs. The path should cover the cerebrovascular lesion area and reduce the scanning of the non-lesion area. The scanning path instruction is sent to the magnetic resonance scanning equipment. After the scan is executed, the image data is post-processed to generate high-quality imaging results. The imaging results are evaluated by medical experts to verify the effectiveness of the scanning path.

[0056] During the scanning process, the magnetic resonance imaging data is captured and processed in real time, the image quality assessment algorithm is applied to generate feedback signals, the scanning path is adjusted according to the feedback signals to improve the image quality, the scanning path, adjustment records and real-time feedback signal data are collected, the new data are incorporated into the model training, the model parameters are updated through incremental learning and online learning, the performance of the updated model is evaluated to ensure its effectiveness in practical applications, the optimization results are fed back to the actual scanning process, new real-time data is collected, the real-time feedback and adjustment mechanism of the system is continuously monitored to ensure its stability and reliability, and each link of the system is iteratively optimized according to the feedback and results in actual applications.

[0057] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. A magnetic resonance image localization diagnosis method based on artificial intelligence, characterized in that: The specific steps of this method are: S100, data collection and preprocessing: collect the patient's medical imaging data and related clinical information, including but not limited to the patient's age, gender, and medical history, preprocess the image data, and extract the patient's anatomical structure and key information on the lesion location; S200, model construction: design an optimization model that takes the patient's anatomical structure, lesion location, and imaging requirements as input states, and defines the action space and reward function so that the model can learn the optimal scanning path strategy through continuous exploration; S300, strategy learning and optimization: During the training phase, the model continuously adjusts the scanning strategy through multiple trials and errors in simulation and actual environments in order to maximize the reward function value. Through technology, the model's ability to understand complex image features and strategy optimization capabilities are enhanced; S400, Personalized Scan Path Generation: After sufficient training, the model can quickly generate a personalized MRI scan path for the patient based on the input patient-specific information. The path can accurately cover the lesion area while minimizing the scanning of non-lesion areas, thereby improving imaging efficiency and diagnostic accuracy. S500, real-time feedback and adjustment: During the actual scanning process, the system can fine-tune the scanning path based on the real-time image quality feedback. The collected data can be used to further train and optimize the model, forming a virtuous circle.

2. The magnetic resonance image localization diagnosis method based on artificial intelligence according to claim 1, characterized in that: The specific steps of S100 are: S101, determine the source of medical images, obtain patient consent and authorization from hospitals and relevant institutions, comply with data protection regulations, use API to download and transmit medical imaging data from data sources, including images in DICOM format, design clinical information collection forms according to research needs, including but not limited to basic information about the patient's age, gender, and medical history, collect and enter clinical information through electronic medical record systems and patient interviews, and verify the collected clinical information; S102, converting the image in DICOM format into a format for analysis and processing. If the analysis software supports DICOM, no conversion is required. Remove noise from the image, improve image quality, adjust image brightness and contrast, adjust window width and window position, make the anatomical structure and lesion area in the image clearer, perform image registration for multimodal or multi-phase images to eliminate spatial differences between different images, segment the image using an automatic method, extract the patient's anatomical structure and lesion location, and annotate them. Check and correct errors, missing values, and abnormal values ​​in clinical information, convert text information into numerical codes, and associate the processed clinical information with the corresponding medical imaging data to form a complete data set; S103, extracting quantitative features and qualitative features from the preprocessed medical images, combining the extracted image features with the patient's clinical information to form a comprehensive patient portrait, and dividing the sorted data set into a training set, a validation set, and a test set.

3. The magnetic resonance image localization diagnosis method based on artificial intelligence according to claim 1, characterized in that: The specific steps of S200 are: S201, select the algorithm according to the characteristics of the problem. For complex scanning path optimization problems, DQN and more advanced algorithms are more suitable. Design the overall architecture of the model. For image input, take the patient's medical imaging data as input, extract the key information of the anatomical structure through the preprocessing step, which can be represented in the form of feature maps, and take the annotation information of the lesion location as part of the input. This information can be used to guide the model to focus on the lesion area during the scanning process. Encode the imaging requirements into quantifiable parameters and use them as part of the input state; S202, based on the controllable parameters of the magnetic resonance scanning device, define the actions that the model can take, combine all possible actions into an action space, and the model will select an action from the space to execute at each decision step, clarify the goal of model optimization, and build a reward function to evaluate the effect of each action based on the goal setting. The reward function should be able to reflect the performance of the model in achieving the goal, and give positive rewards when the model makes beneficial decisions, and give negative penalties when it makes unfavorable decisions; S203, prepare training data, which will be used to simulate the scanning process and generate the input state and reward signal required for training. Through continuous iteration of the training process, the model will learn how to select the optimal action based on the input state to maximize the cumulative reward. In each iteration, the model will perform an action, observe the results, and update its internal parameters to improve future decisions. During the training process, adjust the learning rate, batch size, and training round hyperparameters according to the performance of the model to optimize the training effect of the model, define indicators for evaluating model performance, use the test set to evaluate the model, check its performance on unseen data, and iteratively optimize the model according to the test results.

4. The magnetic resonance image localization diagnosis method based on artificial intelligence according to claim 1, characterized in that: The specific steps of S300 are: S301, before starting training, the parameters of the model are randomly initialized, and the model can be learned from the data, and the hyperparameters that affect the model training and performance are set. According to the characteristics of the magnetic resonance scanning device and the scanning process, a simulation environment is constructed. This environment can simulate various state changes in the scanning process, including but not limited to scanning angle, layer thickness, and imaging quality. The pre-processed medical imaging data and annotation information are used to generate simulated patient data. These data will be used as inputs for the model in the training phase. In the simulation environment, the model starts to execute the scanning strategy. At each time step, the model selects an action according to the current state, and then the simulation environment generates a new state and reward signal according to this action; S302, after receiving the reward signal, the model updates its internal parameters according to the reward value. If the action leads to a higher reward, the model will tend to choose the action again in the future. Conversely, if the action leads to a negative reward, the model will adjust its strategy to avoid making the same mistake again. The model uses technology to extract complex features in the image. These features help the model understand the patient's anatomical structure and lesion location more accurately. The model's strategy is optimized using optimization algorithms. These algorithms can guide the model to find a better solution in the parameter space to maximize the cumulative reward. The experience replay mechanism is used to store and replay past experience, which helps the model break time correlation and learn more stably. S303, during the training process, regularly monitor the performance of the model on the validation set, which can be done by calculating the average reward and other evaluation indicators on the validation set, and deploy the trained model to the actual MRI scanning environment for testing, which involves integrating the model with the scanning equipment, collecting data in the actual environment, and evaluating the performance of the model in the real scenario. Based on the actual test results, further iteratively optimize the model, including but not limited to adjusting the model architecture, improving the input data representation, and optimizing the reward function.

5. The magnetic resonance image localization diagnosis method based on artificial intelligence according to claim 1, characterized in that: The specific steps of S400 are: S401, collect patient-specific information from medical records, imaging data and other related systems, including but not limited to age, gender, weight, height, medical history, current symptoms, and preliminary diagnosis results, pre-process the medical imaging data provided by the patient, load a fully trained model from the storage medium, and configure the necessary hardware and software environment; S402, selecting a model to extract features from the preprocessed image data, capturing key features of the morphology, size, and position of the lesion area, fusing the patient's specific information with the image features to construct a more comprehensive patient representation, and applying an optimization algorithm to plan a personalized scanning path based on the extracted and fused features. The path should be able to accurately cover the lesion area and reduce the scanning of non-lesion areas. When planning the path, the physical constraints and imaging requirements of the magnetic resonance scanning equipment are considered; S403, performing a preliminary evaluation on the generated personalized scanning path to check whether it covers all key lesion areas and evaluate the amount of scanning of non-lesion areas. Based on the preliminary evaluation or expert feedback, the model can automatically adjust the parameters of the scanning path to further optimize the scanning efficiency and accuracy, and convert the finalized personalized scanning path into an instruction sequence executable by the magnetic resonance scanning device; S404, sending the instruction sequence to the magnetic resonance scanning device to perform the scanning operation. After the scanning is completed, the obtained image data is post-processed to generate high-quality imaging results. Medical experts will diagnose and evaluate the imaging results to verify the effectiveness of the personalized scanning path. Feedback from experts and users is collected to understand the performance and problems of the personalized scanning path in actual application. Based on the feedback and new patient data, the model and optimization algorithm are continuously optimized to improve the generation capability and imaging efficiency of the personalized scanning path.

6. The magnetic resonance image localization diagnosis method based on artificial intelligence according to claim 5, characterized in that: The S400 selection model performs feature extraction on the preprocessed image data. The feature extraction formula is: Among them, F represents the final extracted feature set, M is a morphological weight matrix, which is adjusted according to different lesion morphological characteristics to better capture the morphological characteristics of the lesion area, and I P represents the preprocessed image data, represents a specific morphological feature extraction operation, which is used to extract the morphological features in the image. S is the size feature factor, which is determined by analyzing the pixel distribution range of the lesion area in the image and is used to highlight the size characteristics of the lesion area. C is the fusion coefficient, which is used to balance the degree of fusion between the patient's specific information and the image features. P info represents the patient's specific information vector, G is a position feature generator, which is generated by analyzing the spatial coordinate relationship in the image and is used to extract the key position features of the lesion area; Fusion of patient-specific information with image features to construct a more comprehensive patient representation: R = I × W I +P×W P ×F(I,P), where R represents the final more comprehensive patient representation, I is the preprocessed image feature vector, and W I is the image feature weight vector, which is used to adjust the importance of image features in fusion, P is the patient-specific information vector, and W P is the patient-specific information weight vector, which is used to control the influence of patient-specific information. F(I,P) is a fusion function, which can be a fusion network based on deep learning or a custom mathematical function. It nonlinearly fuses image features and patient-specific information to capture the complex relationship between the two, adding novelty and creativity to the fusion process. Based on the extracted and fused features, an optimization algorithm is applied to plan a personalized scanning path: Path = argmin P (D(P)+λ1M(P)+λ2C(P)), where Path represents the planned personalized scanning path, P is a variable set of path parameters, and D(P) is determined by the overlapping area or volume of the domain to ensure that the path can accurately cover the lesion area. M(P) is the influence function on the non-lesion area, which measures the scanning degree of the path P on the non-lesion area. It should be as small as possible and can be determined by counting the number of pixels in the non-lesion area passed by the path or calculating the radiation dose received by the non-lesion area. C(P) is a constraint function, which can be expressed by a series of constraints considering the physical constraints and imaging requirements of the magnetic resonance scanning equipment. λ1 and λ2 are weight coefficients, which are used to balance the importance of the coverage of the lesion area, the impact of the non-lesion area, and the constraints, and can be adjusted according to specific needs.

7. The magnetic resonance image localization diagnosis method based on artificial intelligence according to claim 1, characterized in that: The specific steps of S500 are: S501, starting the magnetic resonance scanning process according to the preset personalized scanning path, the real-time image quality assessment system and the scanning path adjustment mechanism are in an activated state, ready to receive and process real-time data, and in the scanning process, capturing and processing magnetic resonance imaging data in real time, applying the image quality assessment algorithm to assess the quality of the captured image, and generating a real-time feedback signal based on the assessment result, indicating the advantages and disadvantages of the current image quality and potential improvement directions; S502, the system receives and analyzes real-time feedback signals, identifies areas with poor image quality and problems with improper scanning parameter settings, formulates a scanning path fine-tuning strategy based on the feedback signals, and dynamically adjusts the scanning path to improve image quality without affecting the overall progress of the current scanning process. During the scanning process, in addition to image data, all relevant data such as the scanning path, adjustment records, and real-time feedback signals are collected and stored in a safe and reliable database.

8. The magnetic resonance image localization diagnosis method based on artificial intelligence according to claim 7, characterized in that: The specific steps of S500 also include: S503, preprocessing the collected data, incorporating new data into the model training process, updating model parameters through incremental learning and online learning, improving the accuracy and generalization ability of the model, and after updating the model, using an independent test set to evaluate the model performance; S504, feed back the results of model optimization to the actual scanning process, further verify the effect of model improvement, and collect new real-time data to continuously monitor the real-time feedback and adjustment mechanism of the system according to the feedback and results in actual application.

Images