Medical image recognition method and system based on distribution perception hybrid expert model

Abstract

The application provides a medical image recognition method and system based on a distribution perception hybrid expert model, belongs to the technical field of medical image recognition, and acquires medical image data to be recognized; a pre-trained distribution perception hybrid expert model is used to process the acquired medical image data to be recognized, and a medical image recognition result is obtained; a two-stage learning framework is provided, the framework can be used in combination with any basic model, and can be used in combination with class rebalancing technology and information enhancement technology in long-tail distribution learning; rare disease samples can also obtain stable and specialized feature representation, so that the diagnosis accuracy of rare diseases is improved; for highly similar diseases, class confusion can be reduced, so that the discrimination ability for similar diseases is improved.

Description

Technical Field

[0001] This invention relates to the field of medical image recognition technology, specifically to a medical image recognition method and system based on a distributed sensing hybrid expert model. Background Technology

[0002] Medical imaging, with its high image quality and resolution, intuitive visualization, and ability to clearly present human anatomical structures and lesion details, has been widely used in various aspects of clinical diagnosis, disease screening, treatment evaluation, and prognostic analysis. In the modern medical system, the continuous development and updating of various imaging technologies, such as CT (computed tomography), MRI (magnetic resonance imaging), X-rays, and ultrasound, enables hospitals to obtain more comprehensive and accurate patient imaging information. This imaging data not only provides doctors with important diagnostic evidence but also lays the foundation for early disease detection and precision treatment.

[0003] Against this backdrop, computer-aided diagnostic technology has gradually become an important research direction in the field of medical imaging. In recent years, with the improvement of computing power and the expansion of data scale, deep learning technology has developed rapidly and has become a core method in medical image classification and recognition tasks. Deep learning can automatically learn high-level, abstract feature representations from large amounts of image data through multi-layer neural networks, thereby effectively improving the model's discriminative ability and generalization performance. Related studies have shown that deep learning-based methods have demonstrated good performance in tasks such as common disease identification, lesion detection, and category classification, providing strong support for achieving more intelligent and accurate medical image analysis.

[0004] Meanwhile, medical image recognition also faces several challenges, one of which is the extreme long-tail distribution of data. Unlike idealized balanced datasets, medical image data collected in real clinical settings often exhibits a significant long-tail distribution. Influenced by disease epidemiology, rare diseases have extremely low incidence rates in the general population, resulting in an inherently insufficient sample size for medical image data collection. The Chest X-ray14 dataset, open-sourced by the National Institutes of Health Clinical Center, contains over 100,000 chest X-ray images. The official training dataset includes 68,000 images, of which 45,439 are normal, disease-free images. Among the abnormal, diseased images, the most numerous are 6,941, while the fewest are only 7, resulting in an imbalance ratio of 6,491. This imbalance is even more pronounced in real clinical settings. This long-tail distribution characteristic causes deep learning models to tend to learn the discriminative features of head categories during training, while their recognition performance on tail categories significantly decreases.

[0005] Furthermore, many diseases exhibit remarkably similar morphological and appearance features in medical imaging, with subtle and difficult-to-distinguish differences. This high degree of morphological consistency further complicates learning, making it difficult for models to learn stable and discriminative specific features. For example... Figure 1 , Figure 2 As shown, medical images exhibit a more extreme long-tail distribution compared to natural images, and the high similarity between classes presents greater challenges in feature learning. These two phenomena are prevalent in medical image datasets, severely limiting the application value and reliability of models in real-world clinical scenarios.

[0006] Researchers in the field have conducted extensive studies on the problem of learning long-tailed distributions in medical images, which can be mainly divided into three categories: classifier optimization techniques, feature representation enhancement techniques, and network structure improvement techniques.

[0007] In terms of classifier optimization techniques, the bias problem caused by long-tail distribution is mainly mitigated by adjusting the loss function, classifier structure, or decision boundary. Since the number of head class samples in long-tail data far exceeds that of tail classes, the model is more easily dominated by the gradient of the head class during training, thus forming a decision boundary biased towards the head class. To address this, related techniques typically focus on tail class samples during training by assigning different loss weights to different classes, introducing auxiliary classification tasks, or calibrating the predicted logit results, without changing the feature extractor structure. Furthermore, some methods introduce additional discriminative constraints, enabling the model to learn more stable discriminative features while classifying diseases, thereby alleviating the problem of insufficient tail class recognition ability to some extent.

[0008] Feature representation enhancement techniques primarily focus on representation learning to improve a model's ability to model features for a small number of sample classes. Because different diseases in medical images exhibit subtle visual differences and a long-tailed distribution, insufficient inter-class distance or insufficient compactness of intra-class features in the feature space can hinder the model's ability to accurately distinguish between similar diseases. Common approaches include class resampling and data augmentation to balance and enhance feature learning across different classes, or contrastive learning to improve the discriminative power of mid-tail classes and alleviate representational conflicts. Other methods utilize multi-scale feature extraction, attention mechanisms, or feature fusion strategies to enhance the model's perception of overall impact, key lesion regions, and subtle features, thereby reducing performance degradation caused by feature compression or information loss.

[0009] Network architecture improvement methods primarily mitigate the impact of long-tailed distributions by designing novel model structures or training frameworks. Some methods introduce dual-branch or multi-branch structures based on a shared feature extractor, enabling the model to learn feature representations for different categories along different paths. Other methods employ knowledge distillation mechanisms, using a teacher model to guide the student model to obtain more stable supervision information on low-frequency categories. Still other methods utilize multi-expert structures or modular designs, allowing different sub-networks to undertake learning tasks for different categories or features during training. Furthermore, some methods optimize the feature extraction and classification decision-making processes at different stages through phased training or decoupling strategies.

[0010] Most of the methods mentioned above are still optimized within a unified feature and decision space. Although they alleviate the long-tail distribution problem to some extent, they are still largely optimized within a unified feature and decision space. Furthermore, when the inter-class similarity is extremely high, in this unified space, samples from different classes share the same representation learning path and classification boundary, and the learning between different classes may still influence each other. Summary of the Invention

[0011] The purpose of this invention is to provide a medical image recognition method and system based on a distributed sensing hybrid expert model, so as to solve at least one of the technical problems existing in the background art.

[0012] To achieve the above objectives, the present invention adopts the following technical solution:

[0013] In a first aspect, the present invention provides a medical image recognition method based on a distributed-aware hybrid expert model, comprising:

[0014] Acquire the medical image data to be identified;

[0015] A pre-trained distribution-aware hybrid expert model is used to process the acquired medical image data to be identified, resulting in medical image recognition. The learning objective of the distribution-aware hybrid expert model is to learn a mapping that yields a more discriminative representation in long-tail classification tasks. The distribution-aware hybrid expert model includes a distribution-aware expert knowledge learning module and an expert knowledge-guided classifier calibration module. In the distribution-aware expert knowledge learning module, multiple parallel expert encoders are used to learn distribution-specific knowledge and to learn an initial global representation for the entire dataset. In the expert knowledge-guided classifier calibration module, all expert encoders are frozen, and the global classifier head is optimized through adaptive routing and fusion of distribution-aware expert features, thereby calibrating the classifier to reduce head class dominance and improve tail class recognition ability.

[0016] As a further limitation of the first aspect of the present invention, in the distribution-aware expert knowledge learning module, the original dataset is first divided into disjoint subsets according to the number of samples in each category, and each subset contains categories with similar number of samples; for all subsets, encoders with the same structure and no shared singular parameters are used to extract features, and each encoder has an independent classifier, so that the encoder learns the representation of a specific distribution.

[0017] As a further limitation of the first aspect of the invention, in order to enhance the learning effect of expert knowledge for a specific distribution and the learning effect of global representation, a progressive reweighting training strategy is adopted for all encoders, so that all encoders can first learn a stable and universal representation, and then progressively learn low-frequency categories.

[0018] As a further limitation of the first aspect of the present invention, in the expert knowledge-guided classifier calibration module, the learned expert knowledge is connected with the long-tail decision boundary, thereby effectively solving the problem of unclear classification of the subsequent categories; in this module, all expert encoders are frozen, while only the head classifier is optimized and calibrated, thereby enabling stable knowledge transfer from experts with distribution perception capabilities to the global classifier.

[0019] As a further limitation of the first aspect of the present invention, in the expert knowledge-guided classifier calibration module, for all image patch-level features extracted by the encoder, in order to avoid the image-level features being tightly coupled with features of a specific distribution, they are averaged to obtain image-level features, and then these image-level features are projected into a unified representation space through a learnable linear mapping.

[0020] In order to effectively integrate knowledge of a specific distribution, the global image representation is regarded as a benchmark, and a routing function is designed to measure the relevance of each expert to the benchmark.

[0021] Specifically, global features are used to generate routing weights for experts of all specific distributions; routing scores are obtained through linear projection, thereby generating adaptive fusion weights for each sample;

[0022] Then, the weighted expert features and global features are fused through residual connections, and finally fed into the global classifier for prediction.

[0023] As a further limitation of the first aspect of the present invention, during the classifier calibration process, a loss consisting of three parts—classification level, prediction level, and feature level—is designed to ensure the stability of the training process.

[0024] At the prediction level, a knowledge distillation loss based on KL divergence is introduced. The purpose of this loss is to perform discriminative knowledge transfer from experts with distribution awareness to global experts, ensuring that the global prediction is consistent with the experts' predictions on common categories.

[0025] At the feature level, consistency constraints are set to preserve global features as much as possible to prevent large deviations in feature learning during the adaptive expert routing process.

[0026] Secondly, the present invention provides a medical image recognition system based on a distributed sensing hybrid expert model, comprising:

[0027] The acquisition module is used to acquire the medical image data to be identified;

[0028] The processing module is used to process the acquired medical image data to be identified using a pre-trained distribution-aware hybrid expert model to obtain medical image recognition results. The learning objective of the distribution-aware hybrid expert model is to learn a mapping that yields a more discriminative representation in long-tail classification tasks. The distribution-aware hybrid expert model includes a distribution-aware expert knowledge learning module and an expert knowledge-guided classifier calibration module. In the distribution-aware expert knowledge learning module, multiple parallel expert encoders are used to learn distribution-specific knowledge and to learn an initial global representation for the entire dataset. In the expert knowledge-guided classifier calibration module, all expert encoders are frozen, and the global classification head is optimized through adaptive routing and fusion of distribution-aware expert features, thereby calibrating the classifier to reduce head class dominance and improve tail class recognition ability.

[0029] Thirdly, the present invention provides a non-transitory computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the medical image recognition method based on a distributed-aware hybrid expert model as described in the first aspect.

[0030] Fourthly, the present invention provides a computer device including a memory and a processor, wherein the processor and the memory communicate with each other, the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the medical image recognition method based on a distributed sensing hybrid expert model as described in the first aspect.

[0031] Fifthly, the present invention provides an electronic device, comprising: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions for implementing the medical image recognition method based on a distributed sensing hybrid expert model as described in the first aspect.

[0032] The beneficial effects of this invention are as follows: The proposed method is independent of specific feature extraction modules and presents a two-stage learning framework that can be combined with any base model. This framework can also be used in conjunction with class rebalancing and information augmentation techniques in long-tailed distribution learning. Designed for extreme long-tailed distributions, it enables rare disease samples to obtain stable and specialized feature representations, thereby improving the diagnostic accuracy of rare diseases. For highly similar diseases, this invention reduces class confusion, thus improving the ability to distinguish between similar diseases. The implementation of this invention has good theoretical interpretability, which helps reduce false positives in model predictions, decreases misdiagnosis or missed diagnosis in disease-aided diagnosis, and thus improves the accuracy of medical image recognition.

[0033] The advantages of additional aspects of the invention will be set forth more clearly in the following description or will be learned by practice of the invention. Attached Figure Description

[0034] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This diagram illustrates the category distribution and inter-class similarity of natural image datasets in existing technologies.

[0036] Figure 2 This diagram illustrates the category distribution and inter-class similarity of medical image datasets in existing technologies.

[0037] Figure 3 This is a functional framework diagram of the distributed sensing hybrid expert model described in an embodiment of the present invention. Detailed Implementation

[0038] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0039] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0040] It should also be understood that terms such as those defined in general dictionaries should be understood to have meanings consistent with their meanings in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as here.

[0041] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, and / or groups thereof.

[0042] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0043] To facilitate understanding of the present invention, the present invention will be further explained and described below with reference to the accompanying drawings and specific embodiments. However, the specific embodiments do not constitute a limitation on the embodiments of the present invention.

[0044] Those skilled in the art should understand that the accompanying drawings are merely schematic diagrams of embodiments, and the components in the drawings are not necessarily essential for implementing the present invention.

[0045] Hybrid expert models have proven highly effective in learning from long-tailed data. They typically employ a shared encoder to extract global representations from all samples, followed by a gating network that generates expert selection weights based on input features. Only the top-k most relevant experts are activated for computation, and the outputs of the selected experts are then fused. This structure preserves a unified feature space and achieves functional division among experts through conditional computation. However, in medical image recognition, due to extreme long-tailed distributions and high class similarity, the shared encoder is easily dominated by the head class, leading to unbalanced expert utilization. Routing based on global features also struggles to capture differences between samples, resulting in expert mapping errors. This suggests that we can learn global features and features from different class subsets separately, and improve the predictive performance of long-tailed categories in medical images by using feature fusion to simultaneously consider both global and specific information.

[0046] This invention analyzes the advantages of hybrid expert models and, addressing the long-tail distribution phenomenon and high class similarity in medical image recognition, designs a novel long-tail distribution learning method for medical image recognition. This method fully leverages the data distribution characteristics of the long-tail distribution in medical images, introducing a decoupled learning approach between features and classifiers to promote the learning of discriminative features and decision boundaries. It's important to note that this method does not completely replace the methods mentioned above, but can be combined with them. For example, data augmentation can be performed while using this method. This invention primarily addresses the problem of low accuracy in disease image recognition for deep learning models trained when existing medical image data exhibits an extreme long-tail distribution and high class similarity. This invention proposes a distribution-aware hybrid expert model framework that, while retaining global feature extraction capabilities, performs expert feature learning for data with different distributions, separately acquiring discriminative representations of the head and tail distributions. Unlike traditional shared encoders or fixed expert structures, this invention explicitly separates features from different distributions during the representation learning stage and designs an expert knowledge-guided classifier calibration mechanism during the classification stage. By employing the strategy of decoupling the training of features and classifiers as described above, the diagnostic performance of medical images can be effectively improved.

[0047] Example 1

[0048] This embodiment first proposes a medical image recognition system based on a distribution-aware hybrid expert model. The system includes: an acquisition module for acquiring medical image data to be recognized; and a processing module for processing the acquired medical image data using a pre-trained distribution-aware hybrid expert model to obtain medical image recognition results. The learning objective of the distribution-aware hybrid expert model is to learn a more discriminative representation mapped to a long-tail classification task. The distribution-aware hybrid expert model includes a distribution-aware expert knowledge learning module and an expert knowledge-guided classifier calibration module. In the distribution-aware expert knowledge learning module, multiple parallel expert encoders are used to learn distribution-specific knowledge and learn an initial global representation for the entire dataset. In the expert knowledge-guided classifier calibration module, all expert encoders are frozen, and the global classifier head is optimized through adaptive routing and fusion of distribution-aware expert features, thereby calibrating the classifier to reduce head class dominance and improve tail class recognition ability.

[0049] In this embodiment, the above system is used to implement a medical image recognition method based on a distributed sensing hybrid expert model, including: using an acquisition module to acquire medical image data to be recognized; and then using a processing module to process the acquired medical image data to be recognized using a pre-trained distributed sensing hybrid expert model to obtain medical image recognition results.

[0050] In this embodiment, the above-mentioned distributed perceptual expert hybrid model network architecture was trained using a two-stage feature representation and classifier decoupling training strategy. This model can be used to learn more discriminative feature representations in long-tailed medical image classification tasks.

[0051] like Figure 3 As shown, the distributed-aware expert hybrid model network architecture includes two modules: a distributed-aware expert knowledge learning module and an expert knowledge-guided classifier calibration module. In the distributed-aware expert knowledge learning module, multiple expert encoders in parallel are used to learn distribution-specific knowledge and learn an initial global representation for the entire dataset. In the expert knowledge-guided classifier calibration module, all expert encoders are frozen, and the global classifier head is optimized through adaptive routing and fusion of distributed-aware expert features, thereby calibrating the classifier to reduce head class advantage and improve tail class recognition ability.

[0052] In the distribution-aware expert knowledge learning module, the original dataset is first divided into disjoint subsets based on the number of samples in each category. Each subset contains categories with similar sample numbers. For all subsets, encoders with the same structure and no shared singular parameters are used to extract features. Each encoder has an independent classifier, enabling the encoder to learn the representation of a specific distribution.

[0053] Specifically, given a dataset ,in Let represent the images and their corresponding labels in the dataset, respectively, and N represent the number of samples in the dataset. First, the dataset is sorted by the number of samples per category. Assume... Each category The number of samples then has The goal of learning is to learn a The mapping can learn more discriminative representations in long-tail classification tasks, where R represents the real space in which the image is converted into a tensor, H and W represent the height and width of the input image, respectively, X represents the input image sample, Y represents the output label, and C represents the total number of categories.

[0054] In the distribution-aware expert knowledge learning module, the original dataset is first divided into K disjoint subsets based on the number of samples in each category. ,in Each subset Both categories contain similar numbers of samples.

[0055] For each subset, use an encoder with the same structure but not shared singular parameters. To extract features: ;in It is a global feature of the image. This represents the features of P image patches obtained under the Transformer architecture (deep learning); where P represents the number of image patches, and d represents the feature dimension of each image patch. Furthermore, each encoder... Each has an independent classifier This allows the encoder to learn a representation of a specific distribution; in addition, there are encoders for the entire dataset. and classifier Used to extract global features .

[0056] While dividing the entire dataset alleviates the problem of imbalanced data distribution to some extent, there may still be imbalances in the categories within each subset.

[0057] Therefore, in this embodiment, in order to enhance the learning effect of expert knowledge for a specific distribution and the learning effect of global representation, a progressive reweighting training strategy is adopted for all encoders, so that all encoders can first learn a stable and universal representation, and then progressively learn low-frequency categories.

[0058] Specifically, the training loss is calculated independently for each subset and the entire dataset:

[0059] ;

[0060] ;

[0061] Here It is the cross-entropy function. It is an unbalanced proportion. This indicates that the loss function applies to subsets The cross-entropy loss of all samples in the subset is weighted and summed, and then averaged over the number of samples to obtain the training loss of each sample of the encoder for that subset. Extracting global feature representations of input samples using a global encoder Then through a classifier Obtain the prediction results and compare them with the true labels. The cross-entropy loss is calculated to obtain the training loss for samples on the global encoder. Weight coefficients are introduced in this calculation. This is used to adjust the contribution of different classes of samples during the training process, especially in the early stages of training. Set to 1, and use it after the predefined training rounds. This strategy enables all encoders to first learn stable and generalized representations, and then progressively learn low-frequency categories.

[0062] In the expert knowledge-guided classifier calibration module, the learned expert knowledge is connected to the long-tail decision boundary, thereby effectively solving the problem of unclear classification of the posterior categories. In this module, all expert encoders are frozen, while only the head classifier is optimized and calibrated, thus enabling stable knowledge transfer from experts with distribution perception capabilities to the global classifier.

[0063] Specifically, in extreme long-tail distribution scenarios, directly optimizing the expert encoder can lead to a bias towards the head class and disrupt the training process. To mitigate this issue, the expert knowledge-guided classifier calibration module proposed in this embodiment connects the learned expert knowledge with the long-tail decision boundary, thereby effectively solving the problem of unclear classification of subsequent classes. In this module, all expert encoders are frozen, while only the head classifier is optimized and calibrated, thus enabling stable knowledge transfer from experts with distribution awareness to the global classifier.

[0064] In the implementation process, each image is first... The input is fed into all encoders to extract features at the image patch level. and To avoid the tight coupling between image-level features and features of a specific distribution, they are averaged to obtain image-level features, which are then projected into a unified representation space through a learnable linear mapping.

[0065] For each encoder, these image-level features are computed as follows:

[0066] ;

[0067] ;

[0068] in , and , These are the parameters of the linear mapping. , Let represent the representation features of the i-th sample and p-th image patch extracted from the k-th expert encoder and the global encoder, respectively.

[0069] To effectively integrate knowledge of a specific distribution, the global representation of the image is treated as a benchmark, and a routing function is designed to measure the relevance of each expert to this benchmark.

[0070] Specifically, global features It is used to generate routing weights for experts across all specific distributions. Routing scores are obtained through linear projection, thereby generating adaptive fusion weights for each sample. :

[0071] ;

[0072] in, and These represent the weight matrix and bias parameters of the linear mapping of the route, respectively. Indicates the number of expert encoders, Indicates the first Each expert encoder, through the above calculations, can generate normalized weight coefficients for different experts based on the global features of the input samples. This enables adaptive expert selection for different input samples.

[0073] Then, the weighted expert features are fused with the global features through residual connections:

[0074] ;

[0075] This formula indicates that firstly, based on routing weight... Characteristics of each expert We perform a weighted summation to obtain the fused expert features, and then connect them with the global features via residual connections. By combining these features, the final feature representation can be obtained. . Finally, it is fed into the global classifier for prediction.

[0076] In addition, during the classifier calibration process, a loss consisting of three parts—classification level, prediction level, and feature level—was designed to ensure the stability of the training process.

[0077] At the prediction level, a knowledge distillation loss based on KL divergence is introduced. The purpose of this loss is to perform discriminative knowledge transfer from experts with distribution awareness to global experts, ensuring that the global prediction is consistent with the experts' predictions on common categories.

[0078] For experts The loss from predictive knowledge distillation is calculated as follows:

[0079] ;

[0080] Where T is the temperature coefficient for adjustment, and the coefficient is... Used for scaling distillation losses. It is a softmax function. The formula for calculating the loss is based on the final fused features. Input global classifier Obtain the prediction results and base them on category subsets. Extract the predicted probability for the corresponding category; and simultaneously... Expert features obtained from subsets Input the corresponding expert classifier The predicted probabilities are obtained. Then, the logits are smoothed using the temperature coefficient T, and the probability distribution is obtained using the softmax function. Finally, the difference between the two is measured by KL divergence, thereby aligning the prediction distribution of the expert model with the prediction distribution of the global model, achieving prediction-level knowledge distillation.

[0081] At the feature level, consistency constraints are set to preserve global features as much as possible to prevent large deviations in feature learning during the adaptive expert routing process:

[0082] ;

[0083] The final global loss is calculated as follows:

[0084] ;

[0085] and These are the hyperparameters that control the loss at the prediction level and the loss at the feature level, respectively.

[0086] Example 2

[0087] This embodiment 2 provides a non-transitory computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, they implement the medical image recognition method based on a distributed-aware hybrid expert model as described above. The method includes:

[0088] Acquire the medical image data to be identified;

[0089] A pre-trained distribution-aware hybrid expert model is used to process the acquired medical image data to be identified, resulting in medical image recognition. The learning objective of the distribution-aware hybrid expert model is to learn a mapping that yields a more discriminative representation in long-tail classification tasks. The distribution-aware hybrid expert model includes a distribution-aware expert knowledge learning module and an expert knowledge-guided classifier calibration module. In the distribution-aware expert knowledge learning module, multiple parallel expert encoders are used to learn distribution-specific knowledge and to learn an initial global representation for the entire dataset. In the expert knowledge-guided classifier calibration module, all expert encoders are frozen, and the global classifier head is optimized through adaptive routing and fusion of distribution-aware expert features, thereby calibrating the classifier to reduce head class dominance and improve tail class recognition ability.

[0090] Example 3

[0091] This embodiment 3 provides a computer device, including a memory and a processor, wherein the processor and the memory communicate with each other, and the memory stores program instructions that can be executed by the processor. The processor calls the program instructions to execute the medical image recognition method based on a distributed sensing hybrid expert model as described above, the method including:

[0092] Acquire the medical image data to be identified;

[0093] A pre-trained distribution-aware hybrid expert model is used to process the acquired medical image data to be identified, resulting in medical image recognition. The learning objective of the distribution-aware hybrid expert model is to learn a mapping that yields a more discriminative representation in long-tail classification tasks. The distribution-aware hybrid expert model includes a distribution-aware expert knowledge learning module and an expert knowledge-guided classifier calibration module. In the distribution-aware expert knowledge learning module, multiple parallel expert encoders are used to learn distribution-specific knowledge and to learn an initial global representation for the entire dataset. In the expert knowledge-guided classifier calibration module, all expert encoders are frozen, and the global classifier head is optimized through adaptive routing and fusion of distribution-aware expert features, thereby calibrating the classifier to reduce head class dominance and improve tail class recognition ability.

[0094] Example 4

[0095] This embodiment 4 provides an electronic device, including: a processor, a memory, and a computer program; wherein, the processor is connected to the memory, and the computer program is stored in the memory. When the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions for implementing the medical image recognition method based on a distributed sensing hybrid expert model as described above. The method includes:

[0096] Acquire the medical image data to be identified;

[0097] A pre-trained distribution-aware hybrid expert model is used to process the acquired medical image data to be identified, resulting in medical image recognition. The learning objective of the distribution-aware hybrid expert model is to learn a mapping that yields a more discriminative representation in long-tail classification tasks. The distribution-aware hybrid expert model includes a distribution-aware expert knowledge learning module and an expert knowledge-guided classifier calibration module. In the distribution-aware expert knowledge learning module, multiple parallel expert encoders are used to learn distribution-specific knowledge and to learn an initial global representation for the entire dataset. In the expert knowledge-guided classifier calibration module, all expert encoders are frozen, and the global classifier head is optimized through adaptive routing and fusion of distribution-aware expert features, thereby calibrating the classifier to reduce head class dominance and improve tail class recognition ability.

[0098] In summary, the long-tail distribution learning method for medical image recognition based on a distribution-aware hybrid expert model proposed in this invention addresses the problem of extremely long-tail distributions and high class similarity in medical image disease recognition. This method designs a two-stage training framework that decouples feature learning from classifier learning, enabling the learning of more discriminative feature representations in long-tail medical image classification. This invention first acquires distribution knowledge of specific types, then dynamically filters data through a classifier calibration mechanism, thereby achieving unbiased diagnosis. This helps alleviate the dominance of head classes during the learning process and improves the recognition ability of tail classes. The solution to the problem of extremely long-tail distributions and high class similarity in medical image recognition is a method unprecedented in similar tasks.

[0099] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0100] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0101] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0102] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment, whereby a series of operational steps are performed to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0103] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that, based on the technical solutions disclosed in the present invention, various modifications or variations that can be made by those skilled in the art without creative effort should be included within the scope of protection of the present invention.

Claims

1. A medical image recognition method based on a distributed sensing hybrid expert model, characterized in that, include: Acquire the medical image data to be identified; A pre-trained distribution-aware hybrid expert model is used to process the acquired medical image data to be identified, resulting in medical image recognition. The learning objective of the distribution-aware hybrid expert model is to learn a mapping that yields a more discriminative representation in long-tail classification tasks. The distribution-aware hybrid expert model includes a distribution-aware expert knowledge learning module and an expert knowledge-guided classifier calibration module. In the distribution-aware expert knowledge learning module, multiple parallel expert encoders are used to learn distribution-specific knowledge and to learn an initial global representation for the entire dataset. In the expert knowledge-guided classifier calibration module, all expert encoders are frozen, and the global classifier head is optimized through adaptive routing and fusion of distribution-aware expert features, thereby calibrating the classifier to reduce head class dominance and improve tail class recognition ability.

2. The medical image recognition method based on a distributed sensing hybrid expert model according to claim 1, characterized in that, In the distribution-aware expert knowledge learning module, the original dataset is first divided into disjoint subsets based on the number of samples in each category. Each subset contains categories with similar sample numbers. For all subsets, features are extracted using encoders with the same structure and no shared singular parameters. Each encoder has an independent classifier, enabling the encoder to learn the representation of a specific distribution.

3. The medical image recognition method based on a distributed sensing hybrid expert model according to claim 2, characterized in that, To enhance the learning of expert knowledge for specific distributions and the learning of global representations, a progressive reweighting training strategy is adopted for all encoders, enabling all encoders to first learn stable and generalized representations, and then progressively learn low-frequency categories.

4. The medical image recognition method based on a distributed sensing hybrid expert model according to claim 1, characterized in that, In the expert knowledge-guided classifier calibration module, the learned expert knowledge is connected to the long-tail decision boundary, thereby effectively solving the problem of unclear classification of the posterior categories. In this module, all expert encoders are frozen, while only the head classifier is optimized and calibrated, thus enabling stable knowledge transfer from experts with distribution perception capabilities to the global classifier.

5. The medical image recognition method based on a distributed sensing hybrid expert model according to claim 4, characterized in that, In the expert knowledge-guided classifier calibration module, for all image patch-level features extracted by the encoder, in order to avoid the image-level features being tightly coupled with features of a specific distribution, they are averaged to obtain image-level features. Then, these image-level features are projected into a unified representation space through a learnable linear mapping. In order to effectively integrate knowledge of a specific distribution, the global image representation is regarded as a benchmark, and a routing function is designed to measure the relevance of each expert to the benchmark. Specifically, global features are used to generate routing weights for experts of all specific distributions; routing scores are obtained through linear projection, thereby generating adaptive fusion weights for each sample; Then, the weighted expert features and global features are fused through residual connections, and finally fed into the global classifier for prediction.

6. The medical image recognition method based on a distributed sensing hybrid expert model according to claim 5, characterized in that, During the classifier calibration process, a loss consisting of three parts—classification level, prediction level, and feature level—was designed to ensure the stability of the training process. At the prediction level, a knowledge distillation loss based on KL divergence is introduced. The purpose of this loss is to perform discriminative knowledge transfer from experts with distribution awareness to global experts, ensuring that the global prediction is consistent with the experts' predictions on common categories. At the feature level, consistency constraints are set to preserve global features as much as possible to prevent large deviations in feature learning during the adaptive expert routing process.

7. A medical image recognition system based on a distributed sensing hybrid expert model, characterized in that, include: The acquisition module is used to acquire the medical image data to be identified; The processing module is used to process the acquired medical image data to be identified using a pre-trained distribution-aware hybrid expert model to obtain medical image recognition results. The learning objective of the distribution-aware hybrid expert model is to learn a mapping that yields a more discriminative representation in long-tail classification tasks. The distribution-aware hybrid expert model includes a distribution-aware expert knowledge learning module and an expert knowledge-guided classifier calibration module. In the distribution-aware expert knowledge learning module, multiple parallel expert encoders are used to learn distribution-specific knowledge and to learn an initial global representation for the entire dataset. In the expert knowledge-guided classifier calibration module, all expert encoders are frozen, and the global classification head is optimized through adaptive routing and fusion of distribution-aware expert features, thereby calibrating the classifier to reduce head class dominance and improve tail class recognition ability.

8. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium is used to store computer instructions, which, when executed by a processor, implement the medical image recognition method based on a distributed-aware hybrid expert model as described in any one of claims 1-6.

9. A computer device, characterized in that, The method includes a memory and a processor, the processor and the memory communicating with each other, the memory storing program instructions that can be executed by the processor, and the processor calling the program instructions to execute the medical image recognition method based on a distributed sensing hybrid expert model as described in any one of claims 1-6.

10. An electronic device, characterized in that, include: The device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions for implementing the medical image recognition method based on a distributed sensing hybrid expert model as described in any one of claims 1-6.

Images