Distributed medical image classification method
Through the distributed medical image classification method, the image parameter feature extractor and coordinator are used to align the feature distribution, which solves the problem of multi-center data differences, improves the stability and training efficiency of the medical image classification model, and adapts to complex clinical needs.
Patent Information
- Application Number
- CN202411345388.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-26
- Publication Date
- 2025-10-10
- Estimated Expiration
- 2044-09-26
AI Technical Summary
Existing medical image classification methods find it difficult to effectively handle data distribution differences between multiple centers while protecting patient privacy, and traditional resampling and eigenvalue normalization strategies have the risk of image detail loss and overfitting.
A distributed medical image classification method is adopted to automatically classify medical images by training an image parameter feature extractor, aligning feature distributions using an image coordinator, and performing classification through a feature map extractor and a classification model, avoiding the direct training of a huge parameter model.
It improves the performance stability and consistency of classification models in multi-center environments, increases training speed and resource utilization efficiency, and adapts to medical image analysis tasks that meet complex clinical needs.
Smart Images
Figure CN119323687B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of artificial intelligence and medical image analysis, and particularly relates to a distributed medical image classification method. BACKGROUND
[0002] Advanced medical image classification methods can extract key information from medical images, address traditionally difficult-to-quantify problems for specific clinical issues, and have wide application prospects and important clinical value. In recent years, the medical artificial intelligence field has increasingly emphasized the patient privacy security issues and data distribution difference issues existing in the multi-center application process. For the former, distributed models have emerged, which allow joint learning without sharing original data, thereby protecting patient privacy and reducing the risk of data leakage. For example, federated learning (Federated Learning) as a new distributed machine learning method can enable multiple institutions to train models on local data and then send the trained model parameters to a central server for aggregation. This method not only retains the locality of data, but also realizes the global optimization of the model. By introducing distributed models, medical institutions are expected to overcome the privacy security issues and data island effect in traditional centralized learning methods, realize larger-scale data integration and analysis under the premise of ensuring data security, and thus promote cross-institutional collaborative research and accelerate innovation and progress in the field of medical artificial intelligence.
[0003] However, distributed models cannot fully address the issue of data distribution disparity across multiple centers. Research has shown that the performance of medical image classification models can be affected by imaging parameters such as hardware configuration and software algorithms, increasing the risk of model overfitting. To address the impact of slice thickness, pixel size, noise, reconstruction algorithms, detector accuracy, and other imaging parameters on medical images, researchers are currently implementing the following strategies to improve the stability of medical image classification model output: 1) Resampling: Scanned images are resampled before feature extraction to mitigate differences in slice thickness and pixel size between samples. However, conventional upsampling algorithms have limited capabilities; excessive upsampling can produce noticeable artifacts, while downsampling can result in loss of image detail. Some researchers have combined image resampling with a Butterworth filter for preprocessing to correct for the effects of pixel size changes. However, the Butterworth filter also blurs the image, resulting in loss of image detail essential for radiomics. 2) Eigenvalue Normalization: The ComBat method and the Z-score method are commonly used eigenvalue correction methods. The ComBat method can realign the feature distribution within the test set, improving the validation accuracy of radiomics models. The Z-score method corrects the feature value by calculating the mean and standard deviation of each feature in each data set. 3) Stable feature screening: Improve the generalization performance of the model by screening out a stable feature set under different imaging parameters. However, the method of evaluating features one by one is relatively inefficient and does not easily provide a direct reference for the subsequent development of new features and models. This problem is particularly evident when developing deep learning networks for direct prediction. Moreover, the screening process actually narrows the feature space that has potential correlation with the prediction target to a certain extent, which has a negative impact on the training of high-precision models.
[0004] Therefore, in order to solve the problem that existing medical image classification methods are difficult to handle data distribution differences while protecting patient privacy, it is urgent to propose a distributed medical image classification method based on the imaging parameter characteristics of medical images and the difficulties of data processing. Summary of the Invention
[0005] To solve the above technical problems, the present invention proposes a distributed medical image classification method, which provides a basis for building a classification model with stable performance in multiple centers to solve the problems existing in the above-mentioned existing technologies.
[0006] To achieve the above object, the present invention provides a distributed medical image classification method, comprising the following steps:
[0007] The medical images of each branch center server are input into the trained image parameter feature extractor for processing to obtain image parameter features;
[0008] classifying the medical image into a corresponding image parameter category based on the image parameter features, and then converting the medical image into a coordinated image based on a trained image coordinator corresponding to the image parameter category;
[0009] Inputting the coordinated image into a trained feature map extractor for conversion to obtain a feature map;
[0010] The feature map is input into the trained classification model for processing, and the classification result of the medical image is obtained and aggregated.
[0011] Optionally, before inputting the medical image of each sub-center server into the trained image parameter feature extractor for processing to obtain the image parameter feature, the method further includes:
[0012] Based on an interpolation algorithm, the pixel sizes of the medical image in the x-axis, y-axis, and z-axis directions are adjusted to fixed values, and the pixel intensity of the medical image is adjusted to a fixed range.
[0013] Optionally, the method for obtaining a trained image parameter feature extractor includes: constructing an image parameter feature extractor based on a three-dimensional convolutional neural network and a multi-layer perceptron, collecting medical images and corresponding imaging parameters from multiple sub-centers to form an image parameter feature extractor training sample set; training the image parameter feature extractor based on the image parameter feature extractor training sample set; and adjusting the parameters of the image parameter feature extractor based on the infoNCE loss function until the training end conditions are met, thereby obtaining a trained image parameter feature extractor.
[0014] Optionally, the process of classifying the medical image into corresponding image parameter categories based on the image parameter characteristics includes: in each sub-center server, calculating the correlation between the image parameter characteristics of the medical image and the central image parameter characteristics of each image parameter category, sorting them, and classifying the medical image into the image parameter category with the highest correlation.
[0015] Optionally, the process of acquiring the image parameter category includes: performing K-means clustering on the image parameter features, and taking each cluster of the clustering result as the image parameter category.
[0016] Optionally, the process of acquiring the central image parameter feature includes: for each image parameter category, calculating the sum of the correlations between each image feature parameter in each category and all other image parameter features, sorting them, and selecting the image parameter feature with the highest sum of correlations as the central image parameter feature of the corresponding image parameter category.
[0017] Optionally, the process of obtaining the trained image coordinator includes: constructing an image coordinator based on a three-dimensional convolutional neural network, and constructing an image coordination discriminator based on an image parameter feature extractor and a softmax layer; constructing an image coordinator training sample set based on medical images of image parameter categories corresponding to the image coordinator; obtaining image coordination target samples, and training the image coordinator and the image coordination discriminator based on the image coordinator training sample set and the image coordination target samples respectively until the number of training rounds is reached, aggregating the parameters of the final image coordinator and image coordination discriminator through a federated averaging algorithm, updating the parameters of the global image coordinator and the global image coordination discriminator, and obtaining the trained image coordinator.
[0018] Optionally, the process of obtaining image coordination target samples includes: selecting an image parameter category with the most medical images as an initial image coordination target sample; preprocessing the initial image coordination target sample, and inputting it into a trained image coordination target sample generator for processing to obtain an image coordination target sample.
[0019] Optionally, the process of obtaining the trained image coordination target sample generator includes: calculating the cosine similarity between the image parameter features of the image coordination target sample and the central image parameter features of each image parameter category, and taking the negative of the sum of the cosine similarities as the loss function value until the loss function value meets the training end condition, thereby obtaining the trained image coordination target sample generator.
[0020] Optionally, the trained feature map extractor is a three-dimensional convolutional neural network, the end of which includes a convolutional layer with L1 regularization for generating multiple two-dimensional feature maps.
[0021] Optionally, the process of obtaining the trained classification model includes: constructing a classification model based on a convolutional neural network, a global average pooling layer, a multi-layer perceptron and a softmax layer; constructing a classification model training sample set based on the actual classification category corresponding to the medical image, training the classification model based on the classification model training sample set, and adjusting the parameters of the classification model based on the cross-entropy loss function value; aggregating the parameters of the received classification model through a federated averaging algorithm, updating the parameters of the global classification model, and obtaining the trained classification model.
[0022] Compared with the prior art, the present invention has the following advantages and technical effects:
[0023] (1) The present invention can automatically classify medical images into corresponding image parameter categories by training an image parameter feature extractor. This process eliminates the reliance on manual input of image parameters and enables the system to independently identify and process various imaging parameters.
[0024] (2) The present application converts medical images into coordinated images by training an image coordinator, aligns the feature distribution of each image, reduces the distribution difference problem caused by different image sources, and improves the performance stability and consistency of the classification model in a multi-center environment. It is suitable for large-scale, multi-center medical image analysis tasks.
[0025] (3) The present application uses a feature map extractor to convert the coordinated image into a feature map, and uses a classification model to classify the feature map. This method avoids directly training a classification model with a huge parameter size, greatly improves the training speed and resource utilization efficiency of the distributed medical image classification model. In addition, the separation design of the feature map extractor and the classification model makes the system more flexible in processing different medical image classification tasks and adapts to various complex clinical needs. BRIEF DESCRIPTION OF DRAWINGS
[0026] The accompanying drawings, which form a part of this application, are intended to provide further understanding of the application and are incorporated herein in their entirety. The schematic embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:
[0027] Figure 1 The figure is a schematic diagram of the framework of the distributed medical image classification method of the embodiment of the present application;
[0028] Figure 2 The figure is a schematic diagram of the training method flow of the image parameter feature extractor of the embodiment of the present application;
[0029] Figure 3 The figure is a schematic diagram of the training method flow of the feature map extractor of the embodiment of the present application. DETAILED DESCRIPTION
[0030] It should be noted that the embodiments and features in the embodiments of the present application can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.
[0031] It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in an order different from that shown herein.
[0032] As Figure 1 shown, the present embodiment provides a distributed medical image classification method, which comprises the following steps:
[0033] The medical images of each sub-center server are input into the trained image parameter feature extractor for processing to obtain image parameter features;
[0034] classifying the medical image into a corresponding image parameter category based on the image parameter features, and then converting the medical image into a coordinated image based on a trained image coordinator and a trained image coordination discriminator corresponding to the image parameter category;
[0035] Inputting the coordinated image into a trained feature map extractor for conversion to obtain a feature map;
[0036] The feature map is input into the trained classification model for processing, and the classification result of the medical image is obtained and aggregated.
[0037] As a specific embodiment, the medical image is three-dimensional medical image data obtained by scanning and reconstructing a certain medical imaging device such as CT, MRI, or PET.
[0038] The medical images of each branch center server are input into the trained image parameter feature extractor for processing. Before obtaining the image parameter features, the medical images are also preprocessed. The preprocessing method is: using an interpolation algorithm to adjust the pixel size of the medical image in the x-axis, y-axis, and z-axis directions to a fixed value, preferably, the fixed value is 1mm; adjusting the pixel intensity of the medical image to a fixed range, preferably, the fixed range is [-1,1].
[0039] The method for extracting image parameter features from medical images through an image parameter feature extractor is as follows: in each sub-center server, the medical image is input into the image parameter feature extractor, and after being processed by the image parameter feature extractor, a set of vectors is output as image parameter features.
[0040] The image parameter feature extractor is a combination of a three-dimensional convolutional neural network and a multi-layer perceptron. Preferably, the three-dimensional convolutional neural network uses 3D ResNet-34, the dimension of the output feature vector is 1024, the hidden layer dimension of the multi-layer perceptron is 2048, and the output dimension is 512.
[0041] like Figure 2 As shown, the training method of the image parameter feature extractor is:
[0042] Step A10: Collect 100 medical images from each of the multiple branch centers, record imaging parameters corresponding to the medical images, including pixel size, slice thickness, reconstruction algorithm, and device model, and obtain a training sample set for an image parameter feature extractor;
[0043] Step A20, preprocessing the medical images of the image parameter feature extractor training sample set by using the “preprocessing the medical images” of the distributed medical image classification method;
[0044] Step A30: Randomly select 16 training samples from each sub-center server, input the medical images into an image parameter feature extractor, and obtain image parameter features of the medical images;
[0045] Step A40, transmitting the image parameter characteristics of the medical image and the corresponding imaging parameters to the central server;
[0046] Step A50: In the central server, the current performance of the image parameter feature extractor is evaluated using the infoNCE loss function for the image parameter features of the two different medical images. The calculation formula is:
[0047]
[0048] Among them, z i and z j Image parameter features representing medical images with the same imaging parameters, i.e., positive sample pairs; z k Represents z i All negative samples and z j ; τ is a positive temperature parameter; N is z i The number of negative samples plus 1. sim(A,B) is the cosine similarity function, and the calculation formula is:
[0049]
[0050] Where A and B are two non-zero vectors; A·B represents the dot product of A and B; ||A|| and ||B|| are the Euclidean lengths of A and B, respectively. If the loss function value does not meet the training termination criteria, the parameters of the parameter feature extractor are adjusted, the adjusted parameters are transmitted to each sub-center server, and the process jumps to step A30 for iterative training until the loss function value meets the training termination criteria, thereby obtaining a trained parameter feature extractor.
[0051] Medical images are classified into corresponding image parameter categories based on image parameter features, and then converted into coordinated images based on a trained image coordinator and a trained image coordination discriminator corresponding to the image parameter categories. The method for classifying medical images into corresponding image parameter categories based on image parameter features comprises calculating, in each sub-center server, the correlation between the image parameter features of the medical image and the central image parameter features of each image parameter category. The correlation is measured using the Euclidean distance, and the medical image is classified into the image parameter category with the highest correlation.
[0052] The method for obtaining the image parameter category is as follows: the image parameter features used for training the image parameter feature extractor in each sub-center are transmitted to the central server; in the central server, the image parameter features are subjected to K-means clustering; and each clustering cluster of the clustering result is used as the image parameter category.
[0053] The method for selecting the number of clusters for K-means clustering is as follows: using the elbow rule, calculate the total variation under different numbers of clusters, and select the point that significantly reduces the total variation as the candidate number of clusters; using the silhouette coefficient to evaluate the clustering effect under the candidate number of clusters, and select the number of clusters with the highest silhouette coefficient as the final number of clusters.
[0054] The method for obtaining the central image parameter feature is as follows: in the central server, for each image parameter category, the sum of the correlations between each image parameter feature in the category and all other image parameter features in the category is calculated, and the image parameter feature with the highest sum of correlations is selected as the central image parameter feature of the image parameter category.
[0055] The method for converting a medical image into a coordinated image based on the trained image coordinator and the trained image coordination discriminator corresponding to the image parameter category is as follows: selecting an image coordinator corresponding to the image parameter category of the medical image, inputting the medical image into the image coordinator, and outputting an image as a coordinated image after processing by the image coordinator.
[0056] There are multiple image coordinators, and each image parameter category corresponds to an image coordinator.
[0057] The image coordinator is a three-dimensional convolutional neural network, the end of which includes a convolution layer for generating a coordinated image. Preferably, the three-dimensional convolutional neural network uses 3D ResNet-34.
[0058] The training method of the image coordinator is:
[0059] Step B10, determining the image parameter category corresponding to the image coordinator to be trained;
[0060] Step B20: Collect 50 medical images belonging to the image parameter category in each sub-center server to obtain an image coordinator training sample set; for sub-center servers that fail to collect a sufficient number of medical images, exclude them from the image coordinator training process.
[0061] Step B30: In each sub-center server, the medical images of the image coordinator training sample set are pre-processed by the “pre-processing of medical images” method of the distributed medical image classification method;
[0062] Step B40: randomly select 16 training samples from each sub-center server, input the medical images into the image coordinator, and obtain a coordinated image of the medical images;
[0063] Step B50: In each sub-center server, the coordination image is input into the image coordination discriminator to obtain a discrimination result;
[0064] Step B60: In the central server, the image coordination target sample is input into the image coordination discriminator to obtain a discrimination result;
[0065] In step B70, the discriminant loss function value of the image coordination discriminator is calculated in each sub-center server and the central server. The training goal of the image coordination discriminator is to accurately distinguish which coordinated images are generated from the image coordination training samples and which are the image coordination target samples. The parameters of the image coordination discriminator are adjusted based on the discriminant loss function value. The formula of the discriminant loss function is:
[0066]
[0067] Among them, x i represents the image coordination target sample, m represents the number of image coordination target samples, z j represents the image coordinator training sample, n represents the number of image coordinator training samples, D(·) is the image coordination discriminator, and G(·) is the image coordinator.
[0068] In step B80, the image coordinator's generation loss function is calculated in each sub-center server and the central server. The image coordinator's goal is to deceive the image coordination discriminator into believing that the coordinated image generated from the image coordinator's training sample is the image coordination target sample. Based on the generation loss function value, the image coordinator's parameters are adjusted. The formula for the generation loss function is:
[0069]
[0070] Step B90, performing multiple rounds of local training in each sub-center server and the central server. After each round of training, steps B40 to E80 are repeated until a predetermined number of local training rounds is reached.
[0071] Step B100, transmitting the parameters of the image coordinator and the image coordination discriminator of each sub-center server to the central server;
[0072] Step B110 : In the central server, the received parameters of the image coordinator and the image coordination discriminator are aggregated by a federated averaging algorithm to update the parameters of the global image coordinator and the global image coordination discriminator.
[0073] Step B120, distribute the updated parameters of the global image coordinator and the global image coordination discriminator to each sub-center server as the parameters of the image coordinator and the image coordination discriminator.
[0074] Step B130, if the sum of the discrimination loss function values or the sum of the generation loss function values of each sub-center does not meet the training end condition, and the number of training rounds does not reach the predetermined number, jump to step B40 for iterative training.
[0075] The structure of the image coordination discriminator is a combination of an image parameter feature extractor and a softmax layer, and the parameters thereof are initialized using the parameters of the image parameter feature extractor, and are used to determine whether the input image is an image coordination target sample.
[0076] The method for obtaining the image coordination target sample is: selecting an image parameter category with the most medical images, and taking the medical images of the category as image coordination target initial samples; transmitting the image coordination target initial samples to the central server; in the central server, preprocessing the image coordination target initial samples by the "preprocessing medical images" of the distributed medical image classification method; in the central server, inputting the image coordination target initial samples into the image coordination target sample generator to obtain the image coordination target sample.
[0077] The structure of the image coordination target sample generator is the same as that of the image coordinator, which is a three-dimensional convolutional neural network, and the end thereof includes a convolutional layer for generating the image coordination target sample.
[0078] The training method of the image coordination target sample generator is:
[0079] Step C10, transmit the center image parameter features of each image parameter category to the central server;
[0080] Step C20, in the central server, input the preprocessed image coordination target initial samples into the image coordination target sample generator to obtain the image coordination target sample;
[0081] Step C30, in the central server, input the image coordination target sample into the image parameter feature extractor to obtain the image parameter features of the image coordination target sample;
[0082] In step C40, the central server calculates the cosine similarity between the image parameter features of the image coordination target sample and the central image parameter features of each image parameter category, and uses the negative of the sum of the cosine similarities as the loss function value. If the loss function value does not meet the training end condition, the parameters of the image coordination target sample generator are adjusted, and the process jumps to step C20 for iterative training until the loss function value meets the training end condition, thereby obtaining a trained image coordination target sample generator.
[0083] The method for converting the coordination image into a feature map using a feature map extractor is as follows: In each sub-center server, the coordination image is input into the feature map extractor, which processes the image and outputs the feature map. The feature map extractor is a three-dimensional convolutional neural network with a convolutional layer with L1 regularization at the end, which is used to generate multiple two-dimensional feature maps.
[0084] like Figure 3 As shown, the training method of the feature map extractor is:
[0085] Step D10: Combine the feature map extractor with the global average pooling layer, the multi-layer perceptron, and the softmax layer as the feature extraction initial classification network;
[0086] Step D20: In the central server, the image coordination target sample is input into the feature extraction initial classification network to obtain the predicted classification category;
[0087] Step D30: In the central server, the predicted classification category is compared with the actual classification category, and a cross entropy loss function is calculated;
[0088] Step D40: In the central server, if the cross entropy loss function value does not meet the training end condition, the parameters of the feature extraction initial classification network are adjusted and the process jumps to step D20 for iterative training until the cross entropy loss function value meets the training end condition, and a trained feature map extractor is obtained and transmitted to each sub-center server.
[0089] The feature map is input into a trained classification model for processing to obtain a classification result for the medical image. In each sub-center server, the feature map is input into the classification model, and after processing by the classification model, a predicted classification category is output. The classification model is a combination of a convolutional neural network with a global average pooling layer, a multi-layer perceptron, and a softmax layer to generate the predicted classification category. Preferably, the convolutional neural network uses Densenet-121.
[0090] The training method of the classification model is:
[0091] Step E10: Collect 50 medical images from multiple sub-centers, record the actual classification categories corresponding to the medical images, and obtain a classification model training sample set;
[0092] Step E20: In each sub-center server, the medical images in the classification model training sample set are processed through the first to fifth steps of the distributed medical image classification method to obtain corresponding feature maps;
[0093] Step E30: In each sub-center server, the feature map is input into the classification model to obtain a predicted classification category;
[0094] Step E40: In each sub-center server, the predicted classification category is compared with the actual classification category, a cross entropy loss function is calculated, and the parameters of the classification model are adjusted according to the cross entropy loss function value;
[0095] Step E50: Perform multiple rounds of local training in each sub-center server. After each round of training, repeat steps E30 to E40 until a predetermined number of local training rounds is reached.
[0096] Step E60, transmitting the parameters of the classification model of each sub-center server to the central server;
[0097] Step E70: In the central server, the received classification model parameters are aggregated by using a federated averaging algorithm to update the parameters of the global classification model.
[0098] Step E80: Distribute the updated parameters of the global classification model to each sub-center server as parameters of the classification model.
[0099] In step E90, if the sum of the cross entropy loss function values of the sub-centers does not meet the training end condition and the number of training rounds does not reach the predetermined number, jump to step E30 for iterative training.
[0100] Technical effect:
[0101] (1) The present invention can automatically classify medical images into corresponding image parameter categories by training an image parameter feature extractor. This process eliminates the reliance on manual input of image parameters, enabling the system to independently identify and process various imaging parameters. Furthermore, even if new image parameter types are not included in the training data, the present invention can effectively handle them, thereby significantly improving the generalization and promotion capabilities of the classification model and adapting it to more practical application scenarios.
[0102] (2) The application converts medical images into coordinated images by training an image coordinator, aligns the feature distribution of each image, reduces the distribution difference problem caused by different image sources, and improves the performance stability and consistency of the classification model in a multi-center environment, suitable for large-scale, multi-center medical image analysis tasks.
[0103] (3) The application uses a feature map extractor to convert the coordinated images into feature maps, and uses a classification model to classify the feature maps. This method avoids directly training a classification model with a huge number of parameters, greatly improves the training speed and resource utilization efficiency of the distributed medical image classification model. In addition, the separation design of the feature map extractor and the classification model makes the system more flexible in handling different medical image classification tasks and adapting to various complex clinical needs.
[0104] The above is only the preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.
Claims
1. A distributed medical image classification method, characterized in that: The following steps are involved: The medical images of each branch center server are input into the trained image parameter feature extractor for processing to obtain image parameter features; classifying the medical image into a corresponding image parameter category based on the image parameter features, and then converting the medical image into a coordinated image based on a trained image coordinator corresponding to the image parameter category; Inputting the coordinated image into a trained feature map extractor for conversion to obtain a feature map; Inputting the feature map into a trained classification model for processing, obtaining a classification result of the medical image and performing aggregation processing; The process of classifying the medical image into a corresponding image parameter category based on the image parameter features includes: calculating, in each sub-center server, a correlation between the image parameter features of the medical image and the central image parameter features of each image parameter category, sorting the correlations, and classifying the medical image into an image parameter category with the highest correlation; The process of obtaining the image parameter category includes: performing K-means clustering on the image parameter features, and taking each cluster of the clustering result as the image parameter category; The process of acquiring the trained image coordinator includes: constructing an image coordinator based on a three-dimensional convolutional neural network, and constructing an image coordination discriminator based on an image parameter feature extractor and a softmax layer; The training method of the image coordinator is: Step B10, determining the image parameter category corresponding to the image coordinator to be trained; Step B20: In each sub-center server, medical images belonging to the image parameter category are collected to obtain an image coordinator training sample set; Step B30: pre-processing the medical images of the image coordinator training sample set in each sub-center server; Step B40: randomly select a number of training samples in each sub-center server, input the medical images into the image coordinator, and obtain a coordinated image of the medical images; Step B50: In each sub-center server, the coordination image is input into the image coordination discriminator to obtain a discrimination result; Step B60: In the central server, the image coordination target sample is input into the image coordination discriminator to obtain a discrimination result; Step B70: Calculate the discriminant loss function value of the image coordination discriminator in each sub-center server and the central server; the training goal of the image coordination discriminator is to discriminate between the coordinated images and the image coordination target samples in the image coordinator training samples; and adjust the parameters of the image coordination discriminator according to the discriminant loss function value. Step B80: Calculate the generation loss function value of the image coordinator in each sub-center server and the central server; the goal of the image coordinator is to deceive the image coordination discriminator into believing that the coordinated image generated from the image coordinator training sample is the image coordination target sample; adjust the parameters of the image coordinator according to the generation loss function value; Step B90, performing multiple rounds of local training in each sub-center server and the central server until a predetermined number of local training rounds is reached; Step B100, transmitting the parameters of the image coordinator and the image coordination discriminator of each sub-center server to the central server; Step B110: In the central server, the received parameters of the image coordinator and the image coordination discriminator are aggregated by a federated averaging algorithm to update the parameters of the global image coordinator and the global image coordination discriminator. Step B120 : Distribute the updated parameters of the global image coordinator and the global image coordination discriminator to each sub-center server as the parameters of the image coordinator and the image coordination discriminator.
2. The distributed medical image classification method according to claim 1, characterized in that: Before inputting the medical image of each sub-center server into the trained image parameter feature extractor for processing to obtain the image parameter feature, the method further includes: Based on an interpolation algorithm, the pixel sizes of the medical image in the x-axis, y-axis, and z-axis directions are adjusted to fixed values, and the pixel intensity of the medical image is adjusted to a fixed range.
3. The distributed medical image classification method according to claim 1, characterized in that: The method for obtaining a trained image parameter feature extractor includes: constructing an image parameter feature extractor based on a three-dimensional convolutional neural network and a multi-layer perceptron, collecting medical images and corresponding imaging parameters from multiple sub-center servers to form an image parameter feature extractor training sample set; training the image parameter feature extractor based on the image parameter feature extractor training sample set; and adjusting the parameters of the image parameter feature extractor based on the infoNCE loss function until the training end conditions are met, thereby obtaining a trained image parameter feature extractor.
4. The distributed medical image classification method according to claim 1, characterized in that The process of obtaining the central image parameter feature includes: for each image parameter category, calculating the sum of the correlations between each image feature parameter in each category and all other image parameter features, sorting them, and selecting the image parameter feature with the highest sum of correlations as the central image parameter feature of the corresponding image parameter category.
5. The distributed medical image classification method according to claim 1, characterized in that: The process of obtaining image coordination target samples includes: selecting image parameter categories with the most medical images as image coordination target initial samples; after preprocessing the image coordination target initial samples, inputting them into a trained image coordination target sample generator for processing to obtain image coordination target samples.
6. The distributed medical image classification method according to claim 5, characterized in that: The process of obtaining the trained image coordination target sample generator includes: calculating the cosine similarity between the image parameter features of the image coordination target sample and the central image parameter features of each image parameter category, and using the negative of the sum of the cosine similarities as a loss function value until the loss function value meets the training end condition, thereby obtaining the trained image coordination target sample generator.
7. The distributed medical image classification method according to claim 1, characterized in that: The process of obtaining the trained classification model includes: constructing a classification model based on a convolutional neural network, a global average pooling layer, a multi-layer perceptron and a softmax layer; constructing a classification model training sample set based on the actual classification categories corresponding to the medical image, training the classification model based on the classification model training sample set, and adjusting the parameters of the classification model based on the cross-entropy loss function value; aggregating the parameters of the received classification model through a federated averaging algorithm, updating the parameters of the global classification model, and obtaining the trained classification model.