AI-based mobile medical ultrasound contrast imaging analysis and assistance system
Through improved combined heterogeneous image analysis method and visualization technology of CNN models, medical data security and interpretability problems are solved, and efficient ultrasound contrast image analysis and auxiliary diagnosis are achieved to ensure data privacy and improve model generalization capabilities and accuracy sex.
Patent Information
- Application Number
- CN202411020191.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-29
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2044-07-29
AI Technical Summary
Medical data is sensitive and involves personal privacy. The existing methods cannot train efficient AI models on the premise of ensuring data security and privacy, and the heterogeneity of medical image data affects the accuracy of the model; at the same time, the lack of interpretability of medical-assisted diagnostic systems leads to doctors It is difficult for patients to trust their reasoning logic.
The ultrasonic contrast image analysis model is generated using an improved joint heterogeneous image analysis method, ensuring data security through advanced encryption technology and differential privacy policies, and enhancing the model's generalization ability of heterogeneous data; at the same time, training is carried out based on the CNN model and visualizing it using PDPs. Technology improves model interpretability and uses improved optimization algorithms to improve the accuracy of auxiliary diagnostic results.
Without exposing the original patient data, model training is carried out in conjunction with multicenter heterogeneous medical image data to ensure personal information security, improve the accuracy and clinical practicality of ultrasound contrast image analysis, and shorten the doctor's experience by improving model interpretability Judge time and improve work efficiency.
Smart Images

Figure CN118588245B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of intelligent AI, and specifically refers to an AI-based mobile medical ultrasound contrast image analysis and assistance system. Background Art
[0002] The application of AI in the field of medical ultrasound contrast imaging analysis is experiencing rapid development, especially in auxiliary diagnosis. However, this field still faces several challenges: medical data is sensitive and involves personal privacy, and existing methods cannot train efficient AI models under the premise of ensuring data security and privacy; medical image data needs to be highly standardized to ensure the generalization ability of the model, but in reality, data often has heterogeneity, which affects the accuracy of the model; when providing decision support, the current medical auxiliary diagnosis system often lacks sufficient explainability, making it difficult for doctors and patients to establish full trust in the reasoning logic behind it. Summary of the invention
[0003] In view of the above situation, the present invention provides an AI-based mobile medical ultrasound contrast image analysis and auxiliary system. In view of the problem that medical data is sensitive and involves personal privacy, and the existing methods cannot train efficient AI models under the premise of ensuring data security and privacy, an improved joint heterogeneous image analysis method is used to generate an ultrasound contrast image analysis model. Without exposing the patient's original data, the model is trained by coordinating heterogeneous medical image data from multiple centers. By integrating advanced encryption technology and differential privacy strategies, the security of personal information is ensured. At the same time, in view of the diversity and quality heterogeneity of medical image data, this method can enhance the model's generalization ability and Robustness: on the basis of maintaining privacy and security, the accuracy and clinical practicality of ultrasound contrast image analysis are improved; in view of the problem that the current medical auxiliary diagnosis system often lacks sufficient explainability when providing decision support, which makes it difficult for doctors and patients to establish sufficient trust in the reasoning logic behind it, training is based on the CNN model, and PDPs visualization technology is used to improve the explainability of the model, so as to intuitively understand how each feature affects the prediction result. At the same time, an improved optimization algorithm is used to improve the model performance, improve the accuracy of auxiliary diagnosis results, shorten the judgment time of doctors, improve work efficiency, and inject stronger scientific and technological power into medical practice.
[0004] The AI-based mobile medical ultrasound contrast imaging analysis and auxiliary system provided by the present invention includes an image acquisition and processing module, an image analysis and privacy security module, an auxiliary diagnosis and report generation module, a user interface and interaction module, and a data management and storage module;
[0005] The image acquisition and processing module uses a highly sensitive portable ultrasound device to acquire ultrasound contrast images, and uses wireless communication technology to transmit them in real time to the hospital's cloud and local workstations, and performs denoising, image enhancement and image registration operations on the ultrasound contrast images to obtain image data;
[0006] The image analysis and privacy security module uses an improved joint heterogeneous image analysis method to analyze image data and optimizes the security and privacy during image data transmission to obtain image analysis results;
[0007] The auxiliary diagnosis and report generation module is trained based on the CNN model according to the image analysis results, and optimized using PDPs visualization technology and improved optimization algorithms to generate a structured auxiliary diagnosis report to provide decision support for doctors;
[0008] The user interface and interaction module provide an intuitive and easy-to-operate interface, allowing users to view and verify image analysis results, adjust analysis parameters, and perform necessary manual annotations and corrections;
[0009] The data management and storage module ensures that all collected and analyzed data can be stored securely and in compliance with regulations, and supports data retrieval to facilitate long-term tracking of patient conditions or scientific research reviews.
[0010] Furthermore, the image analysis and privacy security module uses an improved joint heterogeneous image analysis method to analyze the image data. The improved joint heterogeneous medical image analysis method specifically includes the following steps:
[0011] Step S1: Data set construction and processing, collecting anonymized ultrasound contrast imaging data from medical institutions, performing standardization processing, and obtaining standardized image data;
[0012] Step S2: Model construction, using the PeleeNet model and the personalized federated learning model to build an ultrasound contrast imaging image analysis model, analyze the standardized image data, and obtain image data analysis results;
[0013] Step S3: CSAHE mechanism, taking each medical institution as a user, and using encryption-summation method to provide protection for all users;
[0014] Step S4: Data analysis and model evaluation, evaluating the performance of the ultrasound contrast image analysis model, and regularly conducting privacy impact assessments to verify whether the CSAHE mechanism effectively prevents the risk of image data leakage through gradient information reverse engineering;
[0015] Step S5: Practical application: deploying the trained ultrasound contrast imaging analysis model to medical institutions to assist doctors in image analysis in real time;
[0016] Step S6: Model iteration and update: Based on the feedback from actual applications, new ultrasound contrast images are continuously collected to iterate the training and optimization of the ultrasound contrast image analysis model. Meanwhile, security measures are continuously strengthened to ensure the accuracy and privacy protection capability of the ultrasound contrast image analysis model.
[0017] Furthermore, in step S2, the model is constructed, specifically including the following steps:
[0018] Step S21: Main model construction, using depth-wise separable convolution and pyramid structure units to build the main part of the PeleeNet model and generate a primary model. Each pyramid unit contains K different numbers of convolutional layers and channel rearrangement operations.
[0019] Step S22: transition layer design, using transition layers between adjacent pyramid structure units to reduce the dimension and size of the primary model, the transition layer includes BN, ReLU activation function, 1x1 convolution and 2x2 average pooling operations;
[0020] Step S23: Model compression and optimization, using model compression and optimization techniques to optimize the storage requirements of the primary model;
[0021] Step S24: Initialize the meta-model, and use the personalized FedAvg algorithm to initialize a global model on the server as the starting point of the meta-model;
[0022] Step S25: Adaptive training, using the MAML-inspired method to perform adaptive training, obtain a high-performance initial meta-model, and record meta-model parameter information;
[0023] Step S26: metamodel update, transmitting updated metamodel parameter information back to the server, and the server updates the metamodel according to the metamodel parameter information;
[0024] Furthermore, in step S3, the CSAHE mechanism specifically includes the following steps:
[0025] Step S31: Initialize and select the initiator, send the global model to all users, and at the end of each round of meta-model training, randomly select a user as the initiator of this round of aggregation and set the parameters required for encryption;
[0026] Step S32: local model training, each user takes the received global model as a starting point, uses private ultrasound contrast imaging data, performs training locally, and generates a local model;
[0027] Step S32: Preparation before encryption, calculating the gradient difference between the local model and the global model, the initiator adds random noise in the calculation of the gradient to generate a noisy gradient;
[0028] Step S33: Homomorphic encryption: encrypt the noisy gradient using a fully homomorphic encryption scheme to generate encrypted tensor data, and use TenSEAL to process the encrypted tensor data to obtain an encrypted noisy gradient. Homomorphic encryption includes a public key and a decryption key.
[0029] Step S34: Circular encryption aggregation, the initiator sends the encrypted and noisy gradient to the next user, and each user uses the encrypted and noisy gradient accumulation scheme to perform privacy-preserving gradient updates to obtain aggregated encrypted gradients;
[0030] Step S35: The initiator decrypts and aggregates. The initiator uses the public keys and decryption keys of all users to decrypt the aggregated encrypted gradients to obtain the aggregated gradients of the global model.
[0031] Step S36: After decryption, remove the previously added random noise from the aggregated gradient of the global model and process the result to obtain an updated global gradient.
[0032] Step S37: Apply gradient update. The server uses the updated global gradient to update the global model, and then broadcasts the updated model to all users to start the next round of training.
[0033] Furthermore, the auxiliary diagnosis and report generation module is optimized using PDPs visualization technology and an improved optimization algorithm, and the PDPs visualization technology and the improved optimization algorithm specifically include the following steps:
[0034] Step P1: CNN model training, using the CNN model to train and learn the image analysis results, extract features, and generate a preliminary auxiliary diagnosis report;
[0035] Step P2: PDPs optimization, use PDPs visualization technology to optimize the CNN model and draw the PDP image;
[0036] Step P3: Model parameter optimization, using an improved optimization algorithm to optimize the parameters of the CNN model;
[0037] Step P4: Iterative optimization: after 100 iterations of CNN network training and parameter optimization, the optimization is terminated and the optimal CNN model parameters are selected as the final CNN model to generate an auxiliary diagnosis report;
[0038] Further, in step P3, the model parameters are optimized, specifically including the following steps:
[0039] Step P31: Population initialization, use logical chaos mapping to initialize the population, set the population size, the number of individuals in the population, the maximum number of iterations, the search space boundary and the population role, the formula used is as follows:
[0040] ;
[0041] ;
[0042] In the formula, is a natural number, and Respectively represent Second and The chaos value of the iteration, is a control parameter in the interval [0,4], and are the upper and lower boundaries of the search space, respectively. Chaos value The values mapped to the search space;
[0043] Step P32: In the exploration phase, all individuals use a unified Levy distribution and jumping strategy to search for the optimal solution of the individual in the search space. The formula used is as follows:
[0044] ;
[0045] In the formula, is the adaptation function, Represents the current iteration number, is the maximum number of iterations, Represents the problem dimension, represents the Levy distribution, is the jump coefficient, and Represent the current individual position and the updated position respectively, represents the best historical position, represents the average position of the current population, To round down, is a random number in the range [0,1];
[0046] Step P33: Step P33: In the development stage, the population roles include leaders, explorers, followers and losers. Individuals with high fitness are defined as leaders to explore better areas and obtain successful strategies for successful exploration. The formula used is as follows:
[0047] ;
[0048] In the formula, represents the ease coefficient, is an exponential function, is a random number between [0,1], is the assessment value of the situation, Representative basis A basic adjustment amount is determined. To generate a dimension The row vector of , the values of the row vector are all 1;
[0049] Define the individual who explores the unknown area as the explorer, looking for the global optimal solution and staying away from the current worst solution;
[0050] Define individuals near the leader as followers, learn and adopt successful strategies;
[0051] Individuals with low fitness are defined as losers and eliminated, and the population is supplemented using a reproduction mechanism.
[0052] The beneficial effects achieved by the present invention using the above scheme are as follows:
[0053] (1) As medical data is sensitive and involves personal privacy, existing methods cannot train efficient AI models while ensuring data security and privacy. An improved joint heterogeneous image analysis method is used to generate an ultrasound contrast image analysis model. This model can train models with heterogeneous medical image data from multiple centers without exposing the original patient data. By integrating advanced encryption technology and differential privacy strategies, the security of personal information is ensured. At the same time, given the diversity and quality heterogeneity of medical image data, this method can enhance the generalization and robustness of the model for heterogeneous data, thereby improving the accuracy and clinical practicality of ultrasound contrast image analysis while maintaining privacy security.
[0054] (2) In view of the problem that the current medical auxiliary diagnosis system often lacks sufficient interpretability when providing decision support, which makes it difficult for doctors and patients to establish sufficient trust in the reasoning logic behind it, we train it based on the CNN model and use PDPs visualization technology to improve the interpretability of the model, so as to intuitively understand how each feature affects the prediction result, shorten the doctor's judgment time, improve work efficiency, and inject stronger scientific and technological power into medical practice. At the same time, we use an improved optimization algorithm to improve the model performance. The improved optimization algorithm improves the accuracy of auxiliary diagnosis results, initializes the population through logical chaos mapping, combines the Levy flight strategy for global exploration, and distinguishes individual roles, including leaders guiding the search, explorers broadening the solution space, followers borrowing successful strategies, and the elimination and supplement mechanism of losers, so as to achieve efficient optimization and promote the rapid convergence and optimization of model parameters. BRIEF DESCRIPTION OF THE DRAWINGS
[0055] Figure 1A schematic diagram of the AI-based mobile medical ultrasound contrast imaging analysis and assistance system provided by the present invention;
[0056] Figure 2 is a schematic diagram of the process of step S3;
[0057] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION
[0058] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.
[0059] Example 1, see Figure 1 The AI mobile medical ultrasound contrast imaging analysis and auxiliary system provided by the present invention includes an image acquisition and processing module, an image analysis and privacy security module, an auxiliary diagnosis and report generation module, a user interface and interaction module, and a data management and storage module;
[0060] The image acquisition and processing module uses a highly sensitive portable ultrasound device to acquire ultrasound contrast images, and uses wireless communication technology to transmit them in real time to the hospital's cloud and local workstations, and performs denoising, image enhancement and image registration operations on the ultrasound contrast images to obtain image data;
[0061] The image analysis and privacy security module uses an improved joint heterogeneous image analysis method to analyze image data and optimizes the security and privacy during image data transmission to obtain image analysis results;
[0062] The auxiliary diagnosis and report generation module is trained based on the CNN model according to the image analysis results, and optimized using PDPs visualization technology and improved optimization algorithms to generate a structured auxiliary diagnosis report to provide decision support for doctors;
[0063] The user interface and interaction module provide an intuitive and easy-to-operate interface, allowing users to view and verify image analysis results, adjust analysis parameters, and make necessary manual annotations and corrections. The user interface only allows doctors and medical institutions with viewing permissions to log in and view, and to annotate and modify image analysis results. Different medical institutions can view the information of other medical institutions only when they are granted relevant permissions.
[0064] The data management and storage module ensures that all collected and analyzed data can be stored securely and in compliance with regulations, and supports data retrieval to facilitate long-term tracking of patient conditions or scientific research reviews. The data management and storage module saves and manages all data in the image acquisition and processing module, image analysis and privacy security module, auxiliary diagnosis and report generation module, and user interface and interaction module, and is updated according to the data of each module.
[0065] Embodiment 2: This embodiment is based on the above embodiment. The image analysis and privacy security module uses an improved joint heterogeneous image analysis method to analyze image data. The improved joint heterogeneous medical image analysis method specifically includes the following steps:
[0066] Step S1: Data set construction and processing, collecting anonymized ultrasound contrast imaging data from medical institutions, performing standardization processing, and obtaining standardized image data;
[0067] Step S2: Model construction, using the PeleeNet model and the personalized federated learning model to build an ultrasound contrast imaging image analysis model, analyze the standardized image data, and obtain image data analysis results;
[0068] Step S3: CSAHE mechanism, taking each medical institution as a user, and using encryption-summation method to provide protection for all users;
[0069] Step S4: Data analysis and model evaluation, evaluating the performance of the ultrasound contrast image analysis model, and regularly conducting privacy impact assessments to verify whether the CSAHE mechanism effectively prevents the risk of image data leakage through gradient information reverse engineering;
[0070] Step S5: Practical application: deploying the trained ultrasound contrast imaging analysis model to medical institutions to assist doctors in image analysis in real time;
[0071] Step S6: Model iteration and update. Based on feedback from actual applications, new ultrasound contrast imaging images are continuously collected to iteratively train and optimize the ultrasound contrast imaging image analysis model. Meanwhile, security measures are continuously strengthened to ensure the accuracy and privacy protection capabilities of the ultrasound contrast imaging image analysis model.
[0072] Embodiment 3: This embodiment is based on the above embodiment. In step S2, the model is constructed, which specifically includes the following steps:
[0073] Step S21: Main model construction, using depth-wise separable convolution and pyramid structure units to build the main part of the PeleeNet model and generate a primary model. Each pyramid unit contains K different numbers of convolutional layers and channel rearrangement operations.
[0074] Step S22: transition layer design, using transition layers between adjacent pyramid structure units to reduce the dimension and size of the primary model, the transition layer includes BN, ReLU activation function, 1x1 convolution and 2x2 average pooling operations;
[0075] Step S23: Model compression and optimization, using model compression and optimization techniques to optimize the storage requirements of the primary model;
[0076] Step S24: Initialize the meta-model, and use the personalized FedAvg algorithm to initialize a global model on the server as the starting point of the meta-model;
[0077] Step S25: Adaptive training, using the MAML-inspired method to perform adaptive training, obtain a high-performance initial meta-model, and record meta-model parameter information;
[0078] Step S26: meta-model update, transmitting updated meta-model parameter information back to the server, and the server updates the meta-model according to the meta-model parameter information.
[0079] Example 4, see Figure 2 This embodiment is based on the above embodiment. Figure 2 In the demonstrated process, the global model is distributed to the user end via a broadcast mechanism, involving user 1, user 2, and user 3. These three users represent three independent medical institutions. The three users use their own private image data 1, image data 2, and image data 3 to train the received global model, generating three local models with different parameters, namely local model 1, local model 2, and local model 3. The gradients of the trained local models are transmitted to the global model using homomorphic encryption, driving the iterative update and upgrade of the global model parameters, enhancing the comprehensiveness of the global model, and improving the cross-domain adaptability and efficiency of the global model.
[0080] In step S3, the CSAHE mechanism specifically includes the following steps:
[0081] Step S31: Initialize and select the initiator, send the global model to all users, and at the end of each round of meta-model training, randomly select a user as the initiator of this round of aggregation and set the parameters required for encryption;
[0082] Step S32: local model training, each user takes the received global model as a starting point, uses private ultrasound contrast imaging data, performs training locally, and generates a local model;
[0083] Step S32: Preparation before encryption, calculating the gradient difference between the local model and the global model, the initiator adds random noise in the calculation of the gradient to generate a noisy gradient;
[0084] Step S33: Homomorphic encryption: encrypt the noisy gradient using a fully homomorphic encryption scheme to generate encrypted tensor data, and use TenSEAL to process the encrypted tensor data to obtain an encrypted noisy gradient. Homomorphic encryption includes a public key and a decryption key.
[0085] Step S34: Circular encryption aggregation, the initiator sends the encrypted and noisy gradient to the next user, and each user uses the encrypted and noisy gradient accumulation scheme to perform privacy-preserving gradient updates to obtain aggregated encrypted gradients;
[0086] Step S35: The initiator decrypts and aggregates. The initiator uses the public keys and decryption keys of all users to decrypt the aggregated encrypted gradients to obtain the aggregated gradients of the global model.
[0087] Step S36: After decryption, remove the previously added random noise from the aggregated gradient of the global model and process the result to obtain an updated global gradient.
[0088] Step S37: Apply gradient update. The server uses the updated global gradient to update the global model, and then broadcasts the updated model to all users to start the next round of training.
[0089] Medical data is sensitive and involves personal privacy. Existing methods cannot train efficient AI models under the premise of ensuring data security and privacy. An improved joint heterogeneous image analysis method is used to generate an ultrasound contrast image analysis model. This model can train models with heterogeneous medical image data from multiple centers without exposing the original data of patients. By integrating advanced encryption technology and differential privacy strategies, the security of personal information is ensured. At the same time, in view of the diversity and quality heterogeneity of medical image data, this method can enhance the generalization ability and robustness of the model for heterogeneous data, and improve the accuracy and clinical practicality of ultrasound contrast image analysis on the basis of maintaining privacy security.
[0090] Embodiment 5, based on the above embodiment, the auxiliary diagnosis and report generation module is optimized using PDPs visualization technology and improved optimization algorithm, and the PDPs visualization technology and improved optimization algorithm specifically include the following steps:
[0091] Step P1: CNN model training, using the CNN model to train and learn the image analysis results, extract features, and generate a preliminary auxiliary diagnosis report;
[0092] Step P2: PDPs optimization, use PDPs visualization technology to optimize the CNN model and draw the PDP image;
[0093] Step P3: Model parameter optimization, using an improved optimization algorithm to optimize the parameters of the CNN model;
[0094] Step P4: Iterative optimization: after 100 iterations of CNN network training and parameter optimization, the optimization is terminated and the optimal CNN model parameters are selected as the final CNN model to generate an auxiliary diagnosis report;
[0095] In step P3, the model parameters are optimized, specifically including the following steps:
[0096] Step P31: Population initialization, use logical chaos mapping to initialize the population, set the population size, the number of individuals in the population, the maximum number of iterations, the search space boundary and the population role, the formula used is as follows:
[0097] ;
[0098] ;
[0099] In the formula, is a natural number, and Respectively represent Second and The chaos value of the iteration, is a control parameter in the interval [0,4], and are the upper and lower boundaries of the search space, respectively. Chaos value Mapping to the location of the search space;
[0100] Step P32: In the exploration phase, all individuals use a unified Levy distribution and jumping strategy to search for the optimal solution of the individual in the search space. The formula used is as follows:
[0101] ;
[0102] In the formula, is the adaptation function, Represents the current iteration number, is the maximum number of iterations, Represents the problem dimension, represents the Levy distribution, is the jump coefficient, and Represent the current individual position and the updated position respectively, represents the best historical position, represents the average position of the current population, To round down, is a random number in the range [0,1];
[0103] Step P33: In the development stage, the population roles include leaders, explorers, followers, and losers. Individuals with high fitness are defined as leaders, who explore better areas and obtain successful strategies for successful exploration. The formula used is as follows:
[0104] ;
[0105] In the formula, represents the ease coefficient, is an exponential function, is a random number between [0,1], is the assessment value of the situation, Representative basis A basic adjustment amount is determined. To generate a dimension The row vector of , the values of the row vector are all 1;
[0106] Define the individual who explores the unknown area as the explorer, looking for the global optimal solution and staying away from the current worst solution;
[0107] Define individuals near the leader as followers, learn and adopt successful strategies;
[0108] Individuals with low fitness are defined as losers and eliminated, and the population is supplemented using a reproduction mechanism.
[0109] Embodiment 6: This embodiment is based on the above embodiment. In step P33, the specific definitions and formulas of explorers, followers and losers in the development stage are as follows:
[0110] Explorers: Individuals ranked in the top 30% to 70% of the population in terms of fitness are designated as explorers and are explored according to the following formula:
[0111] ;
[0112] In the formula, Represents the position of the worst solution in the current iteration;
[0113] Followers: Individuals follow the guidance of the most adaptive leader and follow the leader’s exploration strategy. 70% to 90% of individuals are assigned the role of followers. The formula used is as follows:
[0114] ;
[0115] In the formula, represents the optimal position in the current iteration number, is a random number in the range [1, dim];
[0116] Losers: Individuals with low fitness are losers and eliminated. After elimination, the number of individuals in the population is supplemented. The formula used is as follows:
[0117] ;
[0118] In view of the problem that the current medical auxiliary diagnosis system often lacks sufficient interpretability when providing decision support, which makes it difficult for doctors and patients to establish sufficient trust in the reasoning logic behind it, we train it based on the CNN model and use PDPs visualization technology to improve the interpretability of the model, intuitively understand how each feature affects the prediction result, shorten the doctor's judgment time, improve work efficiency, and inject stronger scientific and technological power into medical practice. At the same time, we use an improved optimization algorithm to improve the model performance, improve the accuracy of auxiliary diagnosis results, initialize the population through logical chaos mapping, combine the Levy flight strategy for global exploration, and distinguish individual roles, including leaders guiding search, explorers broadening solution space, followers borrowing successful strategies, and losers elimination and supplement mechanism, to achieve efficient optimization and promote rapid convergence and optimization of model parameters.
[0119] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.
[0120] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should belong to the protection scope of the present invention.
Claims
1. AI-based mobile medical ultrasound contrast imaging analysis and assistance system, characterized by: It includes image acquisition and processing module, image analysis and privacy security module, auxiliary diagnosis and report generation module, user interface and interaction module and data management and storage module; The image acquisition and processing module uses a highly sensitive portable ultrasound device to acquire ultrasound contrast images, and uses wireless communication technology to transmit them in real time to the hospital's cloud and local workstations, and performs denoising, image enhancement and image registration operations on the ultrasound contrast images to obtain image data; The image analysis and privacy security module uses an improved joint heterogeneous image analysis method to analyze image data and optimizes the security and privacy during image data transmission to obtain image analysis results; The auxiliary diagnosis and report generation module is trained based on the CNN model according to the image analysis results, and optimized using PDPs visualization technology and improved optimization algorithm to generate a structured auxiliary diagnosis report; The user interface and interaction module provide an intuitive and easy-to-operate interface; The data management and storage module ensures that all collected and analyzed data can be stored securely and in compliance with regulations, and supports data retrieval; The improved combined heterogeneous medical image analysis method specifically comprises the following steps: Step S1: Data set construction and processing, collecting anonymized ultrasound contrast imaging data from medical institutions, performing standardization processing, and obtaining standardized image data; Step S2: Model construction, using the PeleeNet model and the personalized federated learning model to build an ultrasound contrast imaging image analysis model, analyze the standardized image data, and obtain image data analysis results; Step S3: CSAHE mechanism, taking each medical institution as a user, and using encryption-summation method to provide protection for all users; Step S4: Data analysis and model evaluation, evaluating the performance of the ultrasound contrast image analysis model, and regularly conducting privacy impact assessments to verify whether the CSAHE mechanism effectively prevents the risk of image data leakage through gradient information reverse engineering; Step S5: practical application; Step S6: model iteration and update; In step S2, the model is constructed, which specifically includes the following steps: Step S21: Main body model construction, using depth-separable convolution and pyramid structure units to build the main part of the PeleeNet model and generate a primary model; Step S22: transition layer design, using transition layers between adjacent pyramid structure units to reduce the dimension and size of the primary model; Step S23: Model compression and optimization, using model compression and optimization techniques to optimize the storage requirements of the primary model; Step S24: Initialize the meta-model, and use the personalized FedAvg algorithm to initialize a global model on the server as the starting point of the meta-model; Step S25: Adaptive training, using the MAML-inspired method to perform adaptive training, obtain a high-performance initial meta-model, and record meta-model parameter information; Step S26: metamodel update, transmitting updated metamodel parameter information back to the server, and the server updates the metamodel according to the metamodel parameter information; In step S3, the CSAHE mechanism specifically includes the following steps: Step S31: Initialize and select the initiator, send the global model to all users, and at the end of each round of meta-model training, randomly select a user as the initiator of this round of aggregation and set the parameters required for encryption; Step S32: local model training, each user takes the received global model as a starting point, uses private ultrasound contrast imaging data, performs training locally, and generates a local model; Step S32: Preparation before encryption, calculating the gradient difference between the local model and the global model, the initiator adds random noise in the calculation of the gradient to generate a noisy gradient; Step S33: Homomorphic encryption: encrypt the noisy gradient using a fully homomorphic encryption scheme to generate encrypted tensor data, and use TenSEAL to process the encrypted tensor data to obtain an encrypted noisy gradient. Homomorphic encryption includes a public key and a decryption key. Step S34: Circular encryption aggregation, the initiator sends the encrypted and noisy gradient to the next user, and each user uses the encrypted and noisy gradient accumulation scheme to perform privacy-preserving gradient updates to obtain aggregated encrypted gradients; Step S35: The initiator decrypts and aggregates. The initiator uses the public keys and decryption keys of all users to decrypt the aggregated encrypted gradients to obtain the aggregated gradients of the global model. Step S36: After decryption, remove the previously added random noise from the aggregated gradient of the global model and process the result to obtain an updated global gradient. Step S37: Apply gradient update; The auxiliary diagnosis and report generation module uses PDPs visualization technology and improved optimization algorithm for optimization, and the PDPs visualization technology and improved optimization algorithm specifically include the following steps: Step P1: CNN model training, using the CNN model to train and learn the image analysis results, extract features, and generate a preliminary auxiliary diagnosis report; Step P2: PDPs optimization, use PDPs visualization technology to optimize the CNN model and draw the PDP image; Step P3: Model parameter optimization, using an improved optimization algorithm to optimize the parameters of the CNN model; Step P4: Iterative optimization: after 100 iterations of CNN network training and parameter optimization, the optimization is terminated and the optimal CNN model parameters are selected as the final CNN model to generate an auxiliary diagnosis report; The auxiliary diagnosis and report generation module uses PDPs visualization technology and improved optimization algorithm for optimization, and the PDPs visualization technology and improved optimization algorithm specifically include the following steps: Step P1: CNN model training, using the CNN model to train and learn the image analysis results, extract features, and generate a preliminary auxiliary diagnosis report; Step P2: PDPs optimization, use PDPs visualization technology to optimize the CNN model and draw the PDP image; Step P3: Model parameter optimization, using an improved optimization algorithm to optimize the parameters of the CNN model; Step P4: Iterative optimization: after 100 iterations of CNN network training and parameter optimization, the optimization is terminated and the optimal CNN model parameters are selected as the final CNN model to generate an auxiliary diagnosis report; In step P3, the model parameters are optimized, specifically including the following steps: Step P31: Population initialization, use logical chaos mapping to initialize the population, set the population size, the number of individuals in the population, the maximum number of iterations, the search space boundary and the population role, the formula used is as follows: ; ; In the formula, is a natural number, and Respectively represent Second and The chaos value of the iteration, is a control parameter in the interval [0,4], and are the upper and lower boundaries of the search space, respectively. Chaos value The values mapped to the search space; Step P32: In the exploration phase, all individuals use a unified Levy distribution and jumping strategy to search for the optimal solution of the individual in the search space. The formula used is as follows: ; In the formula, is the adaptation function, Represents the current iteration number, is the maximum number of iterations, Represents the problem dimension, represents the Levy distribution, is the jump coefficient, and Represent the current individual position and the updated position respectively, represents the best historical position, represents the average position of the current population, To round down, is a random number in the range [0,1]; Step P33: In the development stage, the population roles include leaders, explorers, followers, and losers. Individuals with high fitness are defined as leaders, who explore better areas and obtain successful strategies for successful exploration. The formula used is as follows: ; In the formula, represents the ease coefficient, is an exponential function, is a random number between [0,1], is the assessment value of the situation, Representative basis A basic adjustment is determined. To generate a dimension The row vector of , the values of the row vector are all 1; Define the individual who explores the unknown area as the explorer, looking for the global optimal solution and staying away from the current worst solution; Define individuals near the leader as followers, learn and adopt successful strategies; Individuals with low fitness are defined as losers and eliminated, and the population is supplemented using a reproduction mechanism.
