Medical image recognition system based on adaptive heterogeneous integration and few-sample learning

Abstract

The invention belongs to the technical field of medical image analysis, and particularly relates to a medical image recognition system based on adaptive heterogeneous integration and few-sample learning, and the recognition system comprises the following steps: S1, collecting medical image data; s2, after original data is obtained, data cleaning is carried out to remove unqualified or repeated data points, and comprehensive preprocessing is carried out on the data; s3, designing a self-adaptive heterogeneous integrated network as a backbone architecture for extracting features, and enhancing the initial performance and adaptability of the model by adopting a dual pre-training strategy; and S4, designing a few-sample learning classification module based on the features extracted by the adaptive heterogeneous integrated network. According to the method, the workload of doctors can be relieved, subjective differences are reduced, the system supports the doctors to make more intelligent clinical decisions through efficient and accurate classification and recognition, and the reliability and consistency of diagnosis are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image analysis, and in particular relates to a medical image recognition system based on adaptive heterogeneous integration and few-sample learning. Background Art

[0002] In recent years, with the development of deep learning, medical image analysis technology has made significant progress. Deep learning models, such as convolutional neural networks, can automatically learn high-level feature representations from large amounts of data, thereby improving the accuracy of diagnosis. Through end-to-end learning, deep learning models can directly predict outputs from raw inputs, simplifying the multi-step processing flow in traditional methods and demonstrating strong generalization capabilities. However, despite the excellent performance of deep learning in many aspects, it still faces some challenges. First, deep learning models usually require a large amount of labeled data for training, which is difficult to meet in some medical fields (such as rare diseases). Second, few-sample learning remains a difficult problem. When encountering new diseases or small sample sets, deep learning models may not be able to adapt quickly, resulting in performance degradation. In addition, medical imaging data often presents a long-tail distribution, that is, the samples of the minority class are far less than the majority class. In this case, traditional deep learning models may be biased towards the majority class and ignore the samples of the minority class, thus affecting the overall diagnostic effect.

[0003] Adaptive heterogeneous integration refers to the integration of models with different architectures, parameter spaces, or training strategies to form a comprehensive system that can dynamically adjust its composition and weights according to the characteristics of the input data. This approach not only takes advantage of the unique advantages of each single model, but also enhances the generalization and robustness of the overall system through adaptive mechanisms, thereby more effectively responding to complex and changing task requirements and data distribution. The combination of model diversity in heterogeneous integration and adaptive adjustment mechanisms can significantly improve prediction performance and stability without adding too much computational burden.

[0004] Few-shot learning classification refers to a learning paradigm that can perform efficient classification with only a small number of labeled samples. This method relies on the model's ability to generalize quickly from a limited number of examples, and usually uses prototype learning to solve classification problems. By constructing a prototype for each category, the distance or similarity between the new sample and the prototype of each category is calculated, and the closest category is determined as the prediction result. This method emphasizes the importance of effective feature representation and measurement mechanism, so that the model can show strong generalization ability and adaptability even when facing unseen categories, and is suitable for processing emerging categories or tasks with high labeling costs.

[0005] Continuous learning refers to the ability of a model to dynamically adapt to new tasks and new data without significantly forgetting previously learned knowledge. This process simulates the mechanism by which humans continuously accumulate knowledge and apply it to new situations. It aims to address the limitations of traditional machine learning models when faced with non-stationary distributed data, especially the problem of catastrophic forgetting. By adopting methods such as elastic weight consolidation, generative replay, and incremental learning strategies, we build an intelligent system that can efficiently transfer existing knowledge, update parameters online, and continuously improve performance over time.

[0006] However, traditional medical impact analysis still has the following problems:

[0007] Traditional medical imaging diagnosis mainly relies on the experience and expertise of doctors, but doctors have a huge workload and long working hours may lead to fatigue and increased risk of misdiagnosis. In addition, there may be subjective differences in the diagnosis results between different doctors, especially when facing complex or rare diseases, which may affect the final diagnosis accuracy;

[0008] Medical imaging data is complex and diverse. The data distribution of different imaging modalities varies greatly. It is difficult for a single model to fully capture all features. The hand-designed features of traditional methods are difficult to capture complex medical imaging patterns, especially when faced with different manifestations of different diseases. There are large differences in the accuracy and data captured.

[0009] Traditional machine learning algorithms and deep learning models usually require a large amount of data for training, but obtaining large-scale, high-quality medical imaging data is both time-consuming and expensive, and these data often present a long-tail distribution, that is, the number of samples in the diseased category is far less than that in the normal category. Traditional deep learning models may be biased towards the majority category and ignore the samples in the minority category, thus affecting the overall diagnostic effect;

[0010] When faced with new clinical needs or emerging diseases, traditional methods and deep learning models need to retrain the models, have slow adaptability, and are unable to respond quickly, resulting in decreased performance. Summary of the invention

[0011] The purpose of the present invention is to provide a medical image recognition system based on adaptive heterogeneous integration and few-sample learning, which can reduce the workload of doctors and reduce subjective differences. Through efficient and accurate classification and recognition, the system supports doctors to make more informed clinical decisions and improve the reliability and consistency of diagnosis.

[0012] The technical solution adopted by the present invention is as follows:

[0013] First, the system collects medical data from public medical data sets, hospital or research institution data sets, and cleans, standardizes and converts the format to ensure data consistency and availability. Next, using the multi-branch structure of adaptive heterogeneous integration, each expert network focuses on a specific type of feature extraction, and dynamically adjusts its working mode and weight distribution according to the characteristics of the input data to achieve multimodal feature fusion and comprehensively capture the data distribution of different imaging modalities. On this basis, the system applies the prototype network in the few-shot learning classification to generate a feature center for each known category in the training phase. In the test phase, by calculating the distance between the new sample and all category prototypes, it is assigned to the category corresponding to the nearest prototype. Throughout the process, the system simultaneously trains the adaptive heterogeneous integration and few-shot learning classification modules in the same framework, and continuously adjusts and optimizes the model parameters through a real-time feedback continuous learning mechanism to enhance the adaptability and responsiveness of the system. In addition, the system also uses transfer learning technology to accelerate model training in new fields based on knowledge in existing fields, ensuring that the model can quickly adapt to new clinical needs. Finally, the trained and optimized model is applied to the automatic recognition and classification of medical images to assist doctors in diagnosis, and the model performance is continuously updated and improved through a closed-loop feedback mechanism.

[0014] A medical image recognition system based on adaptive heterogeneous integration and few-sample learning, the recognition system comprising the following steps:

[0015] S1: Collect medical imaging data;

[0016] In S1, the main channels for collecting medical imaging data include public medical data sets, hospital clinical diagnosis and treatment data sets, and medical research project-related data sets, to ensure that medical images of different modalities such as CT scans, MRI, and X-rays covering a wide range of disease types are obtained. Medical images should contain detailed metadata and label information.

[0017] S2: After obtaining the raw data, data cleaning is performed to remove unqualified or duplicate data points, and the data is fully preprocessed;

[0018] In S2, the data preprocessing specifically includes the following steps:

[0019] S21: adjust all images to a uniform standard resolution to meet the input requirements of the model;

[0020] S22: Normalize the pixel values ​​so that they fall within the interval [0,1] to accelerate the training process and stabilize the values. For three-dimensional images, the voxel size needs to be standardized to ensure spatial consistency.

[0021] S23: Use geometric transformation, color transformation, affine transformation and elastic deformation methods to simulate different shooting angles, environmental changes and natural deformations to increase the diversity of image changes;

[0022] S24: Use non-local mean filtering, BM3D algorithm, adaptive filter or deep learning-based denoising network to reduce noise.

[0023] S3: Design an adaptive heterogeneous ensemble network as the backbone architecture for feature extraction, and adopt a dual pre-training strategy to enhance the initial performance and adaptability of the model;

[0024] In S3, the architecture of the adaptive heterogeneous integrated network includes the following steps:

[0025] S31: When initializing the network, the large model weights of the large-scale general dataset and the pre-trained weights on the medical imaging dataset are loaded simultaneously. For the shared layers, the weighted average method is adopted to ensure that the model can learn common features from a wide range of data while focusing on the subtle differences between different categories of medical images.

[0026] S32: When there are multiple types of image data, design a multimodal fusion module that can integrate information from different modalities at the feature level;

[0027] S33: In order to make the network better adapt to different types of inputs, an adaptive mechanism is introduced to dynamically adjust the learning rate and weight decay coefficient hyperparameters of each branch network. At the same time, the attention mechanism is used to allow the network to automatically learn which areas or features are more important.

[0028] By introducing knowledge distillation model compression technology to increase overall performance, the expert network with good performance acts as a "teachers" to pass its learned knowledge to other expert networks, which act as "students" to imitate the teacher's behavior to accelerate convergence;

[0029] The Kullback-Leibler divergence is introduced as the distillation loss function to encourage the probability distribution of the student model output to be close to the soft label of the teacher model.

[0030] S4: Design a few-shot learning classification module based on features extracted by an adaptive heterogeneous ensemble network;

[0031] In S4, designing a few-sample learning classification module includes the following steps:

[0032] S41: Create a metric space, in which the distance between samples of the same type is small, while the distance between samples of different types is large;

[0033] S42: Aiming at the goal in S41, a metric learning algorithm is introduced, which helps the model learn more discriminative feature representations;

[0034] S43: In the training phase, for each known category, the mean of the embedding vectors of all its member samples in the feature space is calculated as the prototype of the category. These prototypes form the basis of the metric space. The support set, i.e., a small number of labeled samples in each category, is used to calculate the category prototype. For each sample in the query set in the same round, the distance between it and all category prototypes is calculated, and the category to which it belongs is determined according to the nearest neighbor principle. At the same time, the parameters of the embedding function are updated according to the prediction results to reduce the loss function and optimize the model.

[0035] S44: In the testing phase, new samples are first mapped to the same feature space through the adaptive heterogeneous integrated network, and then the Euclidean distance between their feature vectors and all pre-stored category prototypes is calculated. According to the nearest neighbor principle, the new samples are assigned to the category corresponding to the nearest prototype. In order to improve efficiency, prototypes of all categories are pre-calculated and cached to ensure fast retrieval;

[0036] S45: Considering the situation of unknown categories, a reasonable threshold mechanism is set. If the distance from a new sample to the closest prototype exceeds the set threshold, the sample is considered not to belong to any known category. The threshold is dynamically adjusted using the Bayesian confidence interval estimation statistical method to deal with unseen category problems.

[0037] S5: In order to further optimize the model, a joint loss function is designed, which combines classification loss, prototype loss and distillation loss;

[0038] In S5, the classification loss uses a cross entropy loss function to measure the difference between the predicted category and the true label;

[0039] The prototype loss uses cosine similarity-based and Euclidean distance-based losses to ensure that the sample feature vector of each category is close to the corresponding prototype and far away from the prototypes of other categories;

[0040] The distillation loss uses the Kullback-Leibler divergence to encourage the student model to imitate the behavior of the teacher model.

[0041] S6: The system is used for real-time feedback of continuous learning mechanisms, ensuring that the model can maintain its memory of previous knowledge while continuously receiving new data and cope with the problem of data imbalance.

[0042] In the S6, the specific contents include the following steps:

[0043] S61: The system extracts features from the adaptive heterogeneous integration network and feeds them into a randomly initialized linear buffer layer, which maps the features to a higher-dimensional space through a ReLU activation function to adapt to subsequent classification tasks;

[0044] S62: When facing the continuously arriving data stream, the system introduces the free online continuous learning algorithm of analysis examples. The free online continuous learning algorithm of analysis examples does not need to save any past samples and directly calculates the analysis using the recursive least squares method;

[0045] S63: In order to maintain a balance between old knowledge and new knowledge, the free online continuous learning algorithm recursively updates the weights so that the relationship between features and labels under the current task is accurately captured, making the overall model performance close to the result of joint learning, that is, assuming that all data are provided to the model for training at one time.

[0046] To address the class imbalance problem in medical imaging data, the system is equipped with a pseudo-feature generator module. The pseudo-feature generator recursively estimates the mean and variance of the true feature distribution of each category and synthesizes additional pseudo-features based on these statistical information.

[0047] These pseudo-features come from the same normal distribution as the actual features, and are used to compensate for the problem of insufficient sample size in the minority category so that the model will not be biased towards the majority category during training. The synthesized pseudo-features only affect the balanced classifier used in the inference phase and will not participate in the update process of the iterative classifier, providing a fair learning environment without destroying the original learning process.

[0048] The technical effects achieved by the present invention are:

[0049] The medical image recognition system based on adaptive heterogeneous integration and few-sample learning of the present invention can reduce the workload of doctors and reduce subjective differences. Through efficient and accurate classification and recognition, the system supports doctors to make more informed clinical decisions and improves the reliability and consistency of diagnosis.

[0050] The medical image recognition system based on adaptive heterogeneous integration and few-sample learning of the present invention utilizes the multi-branch structure of adaptive heterogeneous integration. Each expert network focuses on feature extraction of a specific type and can dynamically adjust its working mode and weight distribution according to the characteristics of the input data. Through multimodal feature fusion, the system can comprehensively capture complex and changeable data distribution from different angles, thereby enhancing the robustness and generalization ability of the system.

[0051] The present invention introduces a medical image recognition system based on adaptive heterogeneous integration and few-sample learning into the few-sample learning classification. When inputting data, only a small number of labeled samples are needed for each category, which effectively avoids the need for a large amount of data and solves the problem of imbalanced long-tail distribution of data. By using these few samples to build prototypes for each category, the model can efficiently learn the characteristics of each category during the training process. New samples are classified by comparing with these prototypes to ensure that minority categories are treated fairly even when the number of samples is extremely unequal. This approach not only greatly simplifies data requirements and reduces the cost of acquiring and processing large-scale data, but also improves the recognition accuracy of minority categories, thereby improving the performance and reliability of the overall diagnostic system.

[0052] The medical image recognition system based on adaptive heterogeneous integration and few-sample learning of the present invention introduces real-time feedback and continuous learning mechanisms. The system can adjust and optimize instantly according to the latest data, enhance the adaptability of the model to rapidly changing environments, and improve the accuracy and personalization of diagnosis and treatment. Continuous learning allows the model to dynamically update its parameters when receiving new data without forgetting existing knowledge, thereby solving the problem of catastrophic forgetting. Through online learning, the system can update the model immediately after receiving each new data sample to ensure that it is always in the optimal state. This mechanism not only enables the system to provide services efficiently and reliably, but also ensures the consistency and reliability of the quality of medical services, and promotes the immediate application of the latest medical knowledge to better meet complex clinical needs and the challenges of emerging diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is an overall workflow diagram of an embodiment of the present invention;

[0054] Figure 2 is a flowchart of preprocessing in an embodiment of the present invention;

[0055] Figure 3 It is a schematic diagram of the structure of an adaptive heterogeneous integrated network according to an embodiment of the present invention;

[0056] Figure 4 is a schematic diagram of a few-sample learning classification according to an embodiment of the present invention;

[0057] Figure 5 It is a schematic diagram of an online learning mechanism according to an embodiment of the present invention. DETAILED DESCRIPTION

[0058] In order to make the purpose and advantages of the present invention more clearly understood, the present invention is specifically described below in conjunction with embodiments. It should be understood that the following text is only used to describe one or several specific embodiments of the present invention, and does not strictly limit the scope of protection of the specific claims of the present invention.

[0059] like Figure 1-Figure 5 As shown, a medical image recognition system based on adaptive heterogeneous integration and few-sample learning, the recognition system comprises the following steps:

[0060] S1: Collect medical imaging data;

[0061] In S1, the main channels for collecting medical imaging data include public medical data sets, hospital clinical diagnosis and treatment data sets, and medical research project-related data sets, to ensure that medical images of different modalities such as CT scans, MRI, and X-rays covering a wide range of disease types are obtained. Medical images should contain detailed metadata and label information.

[0062] S2: After obtaining the raw data, data cleaning is performed to remove unqualified or duplicate data points, and the data is fully preprocessed;

[0063] like Figure 2 As shown, in S2, the data preprocessing specifically includes the following steps:

[0064] S21: adjust all images to a uniform standard resolution to meet the input requirements of the model;

[0065] S22: Normalize the pixel values ​​so that they fall within the interval [0,1] to accelerate the training process and stabilize the values. For three-dimensional images, such as CT and MRI, the voxel size needs to be standardized to ensure spatial consistency.

[0066] S23: Use geometric transformation, color transformation, affine transformation and elastic deformation methods to simulate different shooting angles, environmental changes and natural deformations to increase the diversity of image changes;

[0067] S24: Use non-local mean filtering, BM3D algorithm, adaptive filter or deep learning based denoising network to reduce noise. Class imbalance is a common problem in medical image classification tasks. To solve this problem, oversampling minority class, undersampling majority class, synthesis method to generate new samples, as well as cost-sensitive learning and ensemble learning strategies are used to ensure that the distribution of samples between categories is as balanced as possible.

[0068] The system integrates multiple expert networks with different structures, each of which focuses on a specific type of feature extraction. These expert networks dynamically adjust their working methods and weight distribution according to the characteristics of the input data to cope with complex and changing data distributions. For example, when processing medical images, some expert networks may be better at capturing texture features, while other expert networks are better at capturing shape or color information. In this way, the system can comprehensively capture data features from multiple imaging modalities from different angles, enhancing the robustness and generalization ability of the system. Especially when dealing with data with large distribution differences and more complex data, this mechanism can effectively improve the recognition accuracy of images of different categories.

[0069] S3: Design an adaptive heterogeneous ensemble network as the backbone architecture for feature extraction, and adopt a dual pre-training strategy to enhance the initial performance and adaptability of the model;

[0070] like Figure 3 As shown, in S3, the architecture of the adaptive heterogeneous integrated network includes the following steps:

[0071] S31: When initializing the network, the large model weights of the large-scale general dataset and the pre-trained weights on the medical imaging dataset are loaded simultaneously. For the shared layers, the weighted average method is adopted to ensure that the model can learn common features from a wide range of data while focusing on the subtle differences between different categories of medical images.

[0072] S32: When there are multiple types of imaging data, a multimodal fusion module is designed, which can integrate information from different modalities at the feature level to improve diagnostic accuracy;

[0073] S33: In order to make the network better adapt to different types of inputs, an adaptive mechanism is introduced to dynamically adjust the learning rate and weight decay coefficient hyperparameters of each branch network. At the same time, the attention mechanism is used to allow the network to automatically learn which areas or features are more important.

[0074] By introducing knowledge distillation model compression technology to increase overall performance, the expert network with good performance acts as a "teachers" to pass its learned knowledge to other expert networks, which act as "students" to imitate the teacher's behavior to accelerate convergence;

[0075] The Kullback-Leibler divergence is introduced as the distillation loss function to encourage the probability distribution of the student model output to be close to the soft label of the teacher model. This knowledge sharing mechanism not only enhances the robustness and generalization ability of the system, but also promotes collaboration between expert networks. In terms of specific implementation, in addition to traditional soft label distillation, various forms such as hard negative mining, feature distillation, and relationship distillation are explored to transfer high-level semantic information by minimizing the distance between the activation values ​​of the intermediate layers between the teacher and student networks.

[0076] S4: Design a few-shot learning classification module based on features extracted by an adaptive heterogeneous ensemble network;

[0077] like Figure 4 As shown, in S4, designing a few-sample learning classification module includes the following steps:

[0078] S41: Create a metric space, in which the distance between samples of the same type is small, while the distance between samples of different types is large;

[0079] S42: Aiming at the goal in S41, metric learning algorithms such as contrast loss, triple loss and center loss are introduced. Metric learning algorithms help the model learn more discriminative feature representations, thereby improving classification accuracy.

[0080] S43: In the training phase, for each known category, the mean of the embedding vectors of all its member samples in the feature space is calculated as the prototype of the category. These prototypes form the basis of the metric space. The support set, i.e., a small number of labeled samples in each category, is used to calculate the category prototype. For each sample in the query set in the same round, the distance between it and all category prototypes is calculated, and the category to which it belongs is determined according to the nearest neighbor principle. At the same time, the parameters of the embedding function are updated according to the prediction results to reduce the loss function and optimize the model.

[0081] S44: In the testing phase, new samples are first mapped to the same feature space through the adaptive heterogeneous integrated network, and then the Euclidean distance between their feature vectors and all pre-stored category prototypes is calculated. According to the nearest neighbor principle, the new samples are assigned to the category corresponding to the nearest prototype. In order to improve efficiency, prototypes of all categories are pre-calculated and cached to ensure fast retrieval;

[0082] During the training phase, the system generates a feature center for each known category, which represents the average feature vector of all samples in that category. During the testing phase, the distance between the new sample and all category prototypes is calculated, and the new sample is assigned to the category corresponding to the nearest prototype. This approach significantly reduces the need for large amounts of annotated data and is particularly suitable for situations where there is only a small amount of annotated data, such as rare diseases or emerging diseases. In addition, by combining the rich features extracted by adaptive heterogeneous integration, the category prototype calculation method can maintain high accuracy in complex and changing data environments and provide reliable clinical support.

[0083] S45: Considering the situation of unknown categories, a reasonable threshold mechanism is set. If the distance from a new sample to the closest prototype exceeds the set threshold, the sample is considered not to belong to any known category. The threshold is dynamically adjusted using the Bayesian confidence interval estimation statistical method to deal with unseen category problems.

[0084] This method not only significantly reduces the need for large amounts of labeled data, but also enhances the flexibility and generalization capabilities of the system. It is especially suitable for situations where only a small amount of labeled data is available, such as the diagnosis of rare or emerging diseases. Even when data is scarce, the system can maintain efficient and accurate classification performance while providing robust processing capabilities for unknown categories.

[0085] like Figure 5 As shown, S5: In order to further optimize the model, a joint loss function is designed, which combines classification loss, prototype loss and distillation loss;

[0086] In S5, the classification loss uses a cross entropy loss function to measure the difference between the predicted category and the true label;

[0087] The prototype loss uses cosine similarity-based and Euclidean distance-based losses to ensure that the sample feature vector of each category is close to the corresponding prototype and far away from the prototypes of other categories;

[0088] The distillation loss uses the Kullback-Leibler divergence to encourage the student model to imitate the behavior of the teacher model. By using these loss functions together, the system can simultaneously optimize the adaptive heterogeneous integration and prototype network modules within the end-to-end training framework, ensuring that the two modules work together to improve performance.

[0089] S6: The system is used for real-time feedback of continuous learning mechanisms, ensuring that the model can maintain its memory of previous knowledge while continuously receiving new data and cope with the problem of data imbalance.

[0090] By dynamically updating model parameters and keeping the memory of existing knowledge when receiving new data, the "catastrophic forgetting" problem is effectively solved. This mechanism allows the system to adjust and optimize immediately after each new data sample is received, ensuring that it is always in the best state and can quickly adapt to new clinical needs or disease characteristics. To address the problem of class imbalance, the system generates additional pseudo-features to supplement the samples of minority categories, so that all categories can be fully learned and optimized, improving the recognition accuracy of rare or emerging diseases. This real-time feedback continuous learning mechanism allows all categories to be fully learned and optimized, improving the recognition accuracy of rare or emerging diseases.

[0091] In the S6, the specific contents include the following steps:

[0092] S61: The system extracts features from the adaptive heterogeneous integration network and feeds them into a randomly initialized linear buffer layer, which maps the features to a higher-dimensional space through a ReLU activation function to adapt to subsequent classification tasks;

[0093] S62: When faced with a continuously arriving data stream, the system introduces an analysis example free online continuous learning algorithm. The analysis example free online continuous learning algorithm does not need to save any past samples and directly calculates the analysis using the recursive least squares technique, thus avoiding the problem of catastrophic forgetting.

[0094] S63: In order to maintain a balance between old knowledge and new knowledge, the free online continuous learning algorithm recursively updates the weights so that the relationship between features and labels under the current task is accurately captured, making the overall model performance close to the result of joint learning, that is, assuming that all data are provided to the model for training at one time.

[0095] To address the class imbalance problem in medical imaging data, the system is equipped with a pseudo-feature generator module. The pseudo-feature generator recursively estimates the mean and variance of the true feature distribution of each category and synthesizes additional pseudo-features based on these statistical information.

[0096] These pseudo-features come from the same normal distribution as the actual features, and are used to compensate for the problem of insufficient sample size in the minority category so that the model will not be biased towards the majority category during training. The synthesized pseudo-features only affect the balanced classifier used in the inference phase and will not participate in the update process of the iterative classifier, providing a fair learning environment without destroying the original learning process.

[0097] Combining adaptive heterogeneous integration, few-shot learning classification and continuous learning, the expert networks of adaptive heterogeneous integration are used to capture rich features from different imaging modalities. Multiple expert networks can dynamically adjust their working methods and weight distribution according to the characteristics of the input data to effectively handle complex and changeable data distribution, especially in the case of class imbalance, to improve the recognition accuracy of minority categories; based on the features extracted by the adaptive heterogeneous integration network, few-shot learning classification is applied to generate a feature center for each known category. The system calculates the distance between the new sample and all category prototypes and assigns it to the category corresponding to the nearest feature center. The system can quickly adapt to new categories with only a small amount of labeled data, greatly reducing the dependence on large-scale labeled data, saving time and cost.

[0098] Furthermore, the integration of continuous learning mechanisms enables the system to dynamically update its model parameters as it continuously receives new training data, maintaining the memory of old knowledge while learning new knowledge, and avoiding the occurrence of catastrophic forgetting. The combination of the three enables fast and accurate processing of complex medical imaging data, which is particularly suitable for rare or emerging diseases. Through joint optimization, the system not only improves the ability to identify minor lesions and respond quickly to new clinical needs, but can also automatically adapt to new data and disease characteristics, providing timely and reliable diagnostic support. Specifically, it includes:

[0099] Multimodal feature fusion and adaptive adjustment: Combined with the multi-branch structure of adaptive heterogeneous integration, each expert network focuses on a specific type of feature extraction and can dynamically adjust its working mode and weight distribution according to the characteristics of the input data. Through multimodal feature fusion, the system can fully capture complex and changeable data distribution from different angles, enhancing the robustness and generalization ability of the system. Especially when processing long-tail distributed data, this mechanism can effectively improve the recognition accuracy of minority categories;

[0100] Knowledge mining and transfer learning: Using knowledge distillation of adaptive heterogeneous integration, the expert network with better performance can pass its learned features to other expert networks, helping them converge faster and improve overall performance. Through transfer learning technology, the system can quickly adapt in new fields, use knowledge from existing fields to accelerate model training, significantly reduce the need for large-scale labeled data, and save time and cost;

[0101] Class imbalance solution: A class prototype calculation method that combines adaptive heterogeneous ensemble and few-shot learning classification ensures accurate classification even when data is scarce. This method is particularly suitable for the diagnosis of rare diseases or rare lesions in medical image analysis, and can maintain high accuracy in the case of class imbalance, providing reliable clinical support;

[0102] Continuous learning and joint optimization: A joint optimization strategy is adopted to use adaptive heterogeneous integration and few-shot learning classification for training in the same framework, and a real-time feedback mechanism is introduced to support continuous learning. The system first performs predictive classification on the newly incoming data to generate preliminary results. Subsequently, the model parameters are dynamically adjusted based on the real labels provided by the user as supervision signals to continuously optimize performance. This approach allows the model parameters to be continuously updated and self-optimized based on real-time data streams, which not only improves the training efficiency, but also enhances the system's immediate responsiveness and adaptability. Through continuous learning, the system can quickly adapt to new clinical needs, provide timely and reliable diagnostic support, and ensure the improvement of decision-making quality and service efficiency.

[0103] The above is only a preferred embodiment of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications should also be considered as the protection scope of the present invention. The structures, devices and operating methods not specifically described and explained in the present invention shall be implemented according to the conventional means in the art unless otherwise specified and limited.

Claims

1. A medical image recognition system based on adaptive heterogeneous integration and few-sample learning, characterized by: The identification system comprises the following steps: S1: Collect medical imaging data; S2: After obtaining the raw data, data cleaning is performed to remove unqualified or duplicate data points, and the data is fully preprocessed; S3: Design an adaptive heterogeneous ensemble network as the backbone architecture for feature extraction, and adopt a dual pre-training strategy to enhance the initial performance and adaptability of the model; S4: Design a few-shot learning classification module based on features extracted by an adaptive heterogeneous ensemble network; S5: In order to further optimize the model, a joint loss function is designed, which combines classification loss, prototype loss and distillation loss; S6: The system is used for real-time feedback of continuous learning mechanisms, ensuring that the model can maintain its memory of previous knowledge while continuously receiving new data and cope with the problem of data imbalance.

2. The medical image recognition system based on adaptive heterogeneous integration and few-sample learning according to claim 1, characterized in that: In S1, the main channels for collecting medical imaging data include public medical data sets, hospital clinical diagnosis and treatment data sets, and medical research project-related data sets, to ensure that medical images of different modalities such as CT scans, MRI, and X-rays covering a wide range of disease types are obtained. Medical images should contain detailed metadata and label information.

3. The medical image recognition system based on adaptive heterogeneous integration and few-sample learning according to claim 1, characterized in that: In S2, the data preprocessing specifically includes the following steps: S21: adjust all images to a uniform standard resolution to meet the input requirements of the model; S22: Normalize the pixel values ​​so that they fall within the interval [0,1] to accelerate the training process and stabilize the values. For three-dimensional images, the voxel size needs to be standardized to ensure spatial consistency. S23: Use geometric transformation, color transformation, affine transformation and elastic deformation methods to simulate different shooting angles, environmental changes and natural deformations to increase the diversity of image changes; S24: Use non-local mean filtering, BM3D algorithm, adaptive filter or deep learning-based denoising network to reduce noise.

4. The medical image recognition system based on adaptive heterogeneous integration and few-sample learning according to claim 1, characterized in that: In S3, the architecture of the adaptive heterogeneous integrated network includes the following steps: S31: When initializing the network, the large model weights of the large-scale general dataset and the pre-trained weights on the medical imaging dataset are loaded simultaneously. For the shared layers, the weighted average method is adopted to ensure that the model can learn common features from a wide range of data while focusing on the subtle differences between different categories of medical images. S32: When there are multiple types of image data, design a multimodal fusion module that can integrate information from different modalities at the feature level; S33: In order to make the network better adapt to different types of inputs, an adaptive mechanism is introduced to dynamically adjust the learning rate and weight decay coefficient hyperparameters of each branch network. At the same time, the attention mechanism is used to allow the network to automatically learn which areas or features are more important.

5. The medical image recognition system based on adaptive heterogeneous integration and few-sample learning according to claim 4, characterized in that: By introducing knowledge distillation model compression technology to increase overall performance, the expert network with good performance acts as a "teachers" and transfers its learned knowledge to other expert networks, which act as "students" to imitate the teacher's behavior to accelerate convergence; The Kullback-Leibler divergence is introduced as the distillation loss function to encourage the probability distribution of the student model output to be close to the soft label of the teacher model.

6. The medical image recognition system based on adaptive heterogeneous integration and few-sample learning according to claim 1, characterized in that: In S4, designing a few-sample learning classification module includes the following steps: S41: Create a metric space, in which the distance between samples of the same type is small, while the distance between samples of different types is large; S42: Aiming at the goal in S41, a metric learning algorithm is introduced, which helps the model learn more discriminative feature representations; S43: In the training phase, for each known category, the mean of the embedding vectors of all its member samples in the feature space is calculated as the prototype of the category. These prototypes form the basis of the metric space. The support set, i.e., a small number of labeled samples in each category, is used to calculate the category prototype. For each sample in the query set in the same round, the distance between it and all category prototypes is calculated, and the category to which it belongs is determined according to the nearest neighbor principle. At the same time, the parameters of the embedding function are updated according to the prediction results to reduce the loss function and optimize the model. S44: In the testing phase, new samples are first mapped to the same feature space through the adaptive heterogeneous integrated network, and then the Euclidean distance between their feature vectors and all pre-stored category prototypes is calculated. According to the nearest neighbor principle, the new samples are assigned to the category corresponding to the nearest prototype. In order to improve efficiency, prototypes of all categories are pre-calculated and cached to ensure fast retrieval; S45: Considering the situation of unknown categories, a reasonable threshold mechanism is set. If the distance from a new sample to the closest prototype exceeds the set threshold, the sample is considered not to belong to any known category. The threshold is dynamically adjusted using the Bayesian confidence interval estimation statistical method to deal with unseen category problems.

7. The medical image recognition system based on adaptive heterogeneous integration and few-sample learning according to claim 1, characterized in that: In S5, the classification loss uses a cross entropy loss function to measure the difference between the predicted category and the true label; The prototype loss uses cosine similarity-based and Euclidean distance-based losses to ensure that the sample feature vector of each category is close to the corresponding prototype and far away from the prototypes of other categories; The distillation loss uses the Kullback-Leibler divergence to encourage the student model to imitate the behavior of the teacher model.

8. The medical image recognition system based on adaptive heterogeneous integration and few-sample learning according to claim 1, characterized in that: In the S6, the specific contents include the following steps: S61: The system extracts features from the adaptive heterogeneous integration network and feeds them into a randomly initialized linear buffer layer, which maps the features to a higher-dimensional space through a ReLU activation function to adapt to subsequent classification tasks; S62: When facing the continuously arriving data stream, the system introduces the free online continuous learning algorithm of analysis examples. The free online continuous learning algorithm of analysis examples does not need to save any past samples and directly calculates the analysis using the recursive least squares method; S63: In order to maintain a balance between old knowledge and new knowledge, the free online continuous learning algorithm recursively updates the weights so that the relationship between features and labels under the current task is accurately captured, making the overall model performance close to the result of joint learning, that is, assuming that all data are provided to the model for training at one time.

9. The medical image recognition system based on adaptive heterogeneous integration and few-sample learning according to claim 8, characterized in that: To address the class imbalance problem in medical imaging data, the system is equipped with a pseudo-feature generator module. The pseudo-feature generator recursively estimates the mean and variance of the true feature distribution of each category and synthesizes additional pseudo-features based on these statistical information. These pseudo-features come from the same normal distribution as the actual features, and are used to compensate for the problem of insufficient sample size in the minority category so that the model will not be biased towards the majority category during training. The synthesized pseudo-features only affect the balanced classifier used in the inference phase and will not participate in the update process of the iterative classifier, providing a fair learning environment without destroying the original learning process.

Images