A medical image analysis method, system, electronic device, and readable storage medium

Abstract

The present invention discloses a medical image analysis method, system, electronic device, and readable storage medium, belonging to the technical field of deep learning. The method includes: S100: obtaining lung CT images with true labels for pneumonia prediction; S200: constructing a medical image analysis network model, the model including an image encoder, a feature perturbation module, a feature correction module, a cyclic correction module, and a dual adaptive network; S300: classifying the input lung CT images through the medical image analysis network model. The present invention can improve the robustness and generalization ability of the model in image classification tasks, and can show stronger adaptability and higher diagnostic accuracy when facing data distribution shift and diverse attack means.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and more particularly to a medical image analysis method, system, electronic device, and readable storage medium. Background Art

[0002] Disease diagnosis and anomaly detection are key aspects of medical image analysis (Medical image analysis technology refers to the use of computer algorithms and artificial intelligence technology to process and analyze medical images (such as X-rays, CTs, MRIs, etc.) to assist doctors in disease diagnosis, anomaly detection, tissue classification, treatment planning, etc.). Accurate image analysis can help doctors detect lesions earlier, develop personalized treatment plans, and monitor the progression of diseases, improving the diagnostic accuracy and efficiency. The application of deep learning models has broad application prospects in the field of medical image analysis. However, the performance of the model highly depends on the data distribution. When facing complex clinical environments, distribution shift can cause the model to perform poorly in the new environment, increasing the misdiagnosis rate and missed diagnosis rate. At the same time, due to privacy protection requirements and the suddenness of medical health events, the high sensitivity and lack of medical data pose great challenges to model training. Therefore, how to improve the model performance using limited data while protecting privacy has become an urgent problem to be solved.

[0003] The Test-time Adaptation (TTA) mechanism has thus been proposed. It is set that the source data cannot be accessed, allowing the model to self-optimize according to the data in the actual application scenario after deployment, enabling the model to quickly adapt to the new environment, thereby improving its generalization ability and robustness. Traditional transfer learning or domain adaptation methods usually require a large amount of labeled data, which may not be very practical. In contrast, TTA can not only perform fine-tuning using a small number of unlabeled samples without violating privacy regulations but also effectively handle data distribution differences brought about by different times, devices, shooting conditions, and imaging protocols. This makes the TTA mechanism particularly suitable for disease diagnosis and anomaly detection tasks in medical image analysis.

[0004] Traditionally, medical image analysis relies on a single network to obtain image features. However, a single path may not be able to fully capture the multi-scale information of the image, especially when faced with complex structures such as lung CT scans. To comprehensively and multi-dimensionally capture information, some studies have introduced a dual structure. For example, Li et al. (Li, Y., Chen, X., & Hu, J. (2018). A multi-task learning framework for high-resolution and low-resolution image feature extraction. Journal of Medical Imaging, 5(3), 034501.) proposed a multi-task learning framework that combines two convolutional neural networks (CNNs) with different depths for feature extraction of high-resolution and low-resolution images respectively. This method enhances the model's ability to capture information at different scales, but also brings the problem of increased model complexity. However, these methods usually require more computing resources, and how to effectively utilize and integrate dual models or even multi-models remains a challenge. The Contrastive Test-Time Adaptation proposed by Chen et al. fine-tunes the model during the test phase through contrastive learning, which not only retains the original knowledge of the model but also enhances its adaptability to new domains, providing a new idea and direction for the application of multi-models in TTA. (Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In CVPR, 2022.)

[0005] To address the domain difference problem, the TTA method mainly focuses on adjusting the Batch Normalization (BN) layer, improving output prediction, or domain space alignment. Methods for adjusting the BN layer, such as the AdaBN (Adaptive Batch Normalization) method that dynamically updates the statistics and affine parameters of the BN layer according to the statistics of the target domain for adaptation, are widely adopted, and its performance is equivalent to updating the entire BN layer. However, it is prone to the inability to accurately estimate the test data distribution due to insufficient sample size in the target domain, resulting in the failure of BN layer update. In contrast, AdaIN (Adaptive Instance Normalization) is more suitable for single or few-shot learning environments. Because it applies normalization processing to each instance individually, allowing for personalized adjustment for each input. At the same time, AdaIN has similarities with the MixStyle method, and both involve operations on the feature map style (i.e., channel-level mean and standard deviation) to generate diverse samples. The direct feature-level domain synthesis method effectively generates different potential target domains by randomly perturbing the feature channel statistics. However, if the perturbation intensity is too large, it may cause highly activated channels to be overly interfered, resulting in a decline in model performance. In addition, overly drastic changes may also cause the model to lose its original knowledge. Therefore, a more refined control perturbation mechanism needs to be developed.

[0006] In summary, how to provide a medical image analysis method, system, electronic device, and readable storage medium is an urgent problem for those skilled in the art. Summary of the Invention

[0007] In view of this, the present invention provides a medical image analysis method, system, electronic device, and readable storage medium, aiming to improve the robustness and generalization ability of the model in image classification tasks, thereby improving the accuracy and reliability of disease diagnosis and anomaly detection, and is particularly suitable for cross-institutional data sharing and rapid response to new diseases.

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] A medical image analysis method includes:

[0010] S100: Obtain lung CT images with true labels for pneumonia prediction;

[0011] S200: Construct a medical image analysis network model, the model includes an image encoder, a feature perturbation module, a feature correction module, a cyclic correction module, and a dual adaptive network;

[0012] S300: Classify the input lung CT images through the medical image analysis network model;

[0013] Among them, the image encoder is used to extract features from the input lung CT image to obtain a feature map; the feature perturbation module is used to apply local or global perturbations to the feature map in the shallow layer of the medical image analysis network model to obtain perturbed features; the feature correction module is used to correct the perturbed features to obtain corrected features; the cyclic correction module is used to iteratively apply a correction network to perform secondary correction on the corrected features to obtain a statistically represented cyclic correction; the dual adaptive network includes two sub-networks constructed based on the same ResNet architecture but with independently adjusted parameters, which are respectively used to process the feature map, generate their respective image classification results and output, and finally integrate them.

[0014] Further, the image encoder is used to extract features from the input lung CT image to obtain a feature map with a dimension of ; where ; among them, represents the channel dimension of the feature map, and respectively represent the height and width of the feature map.

[0015] Further, the feature perturbation module includes: a local perturbation module, a global perturbation module, and a global statistic update mechanism; the processing steps of the feature perturbation module include: calculating the feature channel statistics based on the feature map; randomly introducing local and global perturbations according to the conditional probability in a specific network layer; for local perturbations, first create a zero matrix and a one matrix with the same dimension as the mean of the feature map, and then randomly generate noise coefficients from a normal distribution with a mean of 0 and a standard deviation of 0.75. These noise coefficients are used to control the injection process of Gaussian noise, and the feature channel statistics (mean and standard deviation) of the original feature map are replaced with the perturbed channel statistics. Then, through Adaptive Instance Normalization, the perturbation parameters (i.e., the perturbed mean and standard deviation) are applied to the affine transformation of the features, thereby realizing the adjustment of the feature map; if global perturbations are applied, the dataset statistics of the current layer are determined, and the perturbed mean and standard deviation are calculated using the global perturbation factor. Similar to local perturbations, a perturbed feature map is then generated.

[0016] Further, the feature correction module creates a corresponding sequence model from several perturbation layers, including: three convolutional layers Conv2d, two custom normalization layers LayerNorm2d, and two non-linear activation functions ReLU; among them, the custom normalization layer: normalizes the feature map of each channel and calculates the mean and standard deviation of each channel.

[0017] Further, the cyclic correction module includes:

[0018] Determine whether to make corrections according to the random probabilities of local perturbations and global perturbations, and use the corresponding perturbation parameters for processing when the conditions are met;

[0019] Adjust the mean and standard deviation of the correction features according to the perturbation parameters;

[0020] Perform standardization and rescaling translation operations to generate a new feature distribution;

[0021] Transfer the processed features to the domain adapter of the correction network, and use the split tensor method to avoid gradient calculation to ensure that the gradient flow remains unchanged;

[0022] Calculate and update the mean and standard deviation of the feature map for subsequent loss calculation.

[0023] Furthermore, the dual adaptive network includes two encoders with the same structure but different values, and their respective decoders, which are used to map the features to the final classification results;

[0024] Among them, the two encoders with the same structure but different values share the same network configuration, but their initial values, learning rates, or optimizers are randomly different;

[0025] Input the lung CT images into the two dual adaptive encoders respectively to extract multi-level feature representations, and then output the integrated prediction output through the two identical decoders.

[0026] Furthermore, the medical image analysis network is also trained using the cross-entropy loss function, similarity loss function, and alignment cycle loss function until the loss function converges, so as to obtain the optimal medical image analysis model for classifying and outputting the input lung images.

[0027] A medical image analysis system, comprising:

[0028] Data acquisition module: Acquire lung CT images with real labels for pneumonia prediction;

[0029] Model construction module: Construct a medical image analysis network model, which includes an image encoder, a feature perturbation module, a feature correction module, a cyclic correction module, and a dual adaptive network;

[0030] Recognition and detection module: Classify the input lung CT images through the medical image analysis network model;

[0031] Among them, the image encoder is used to extract features from the input lung CT image to obtain a feature map; the feature perturbation module is used to apply local or global perturbations to the feature map in the shallow layer of the medical image analysis network model to obtain perturbed features; the feature correction module is used to correct the perturbed features to obtain corrected features; the cyclic correction module is used to iteratively apply a correction network to perform secondary correction on the corrected features to obtain a statistically represented result after cyclic correction; the dual adaptive network includes two sub-networks constructed based on the same ResNet architecture but with independently adjusted parameters, which are respectively used to process the feature map, generate their respective image classification results and output them, and finally integrate them.

[0032] An electronic device includes a memory and a processor. The memory is used to store a program that supports the processor to execute a medical image analysis method, and the processor is configured to execute the program stored in the memory.

[0033] A computer-readable storage medium stores a computer program. When the computer program is run by a processor, it executes the steps of a medical image analysis method.

[0034] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a medical image analysis method, system, electronic device, and readable storage medium. The beneficial effects are as follows:

[0035] 1. The present invention uses isomorphic but heterogenous feature extractors (such as initialization values, learning rates, optimizers, etc.) to form a two-branch structure to capture diverse feature representations. Each branch generates two paths: one retains the original features, and the other is the enhanced features after perturbation and correction. By integrating the prediction results of the two decoders through the idea of ensemble learning, the expression ability and robustness of the model are significantly enhanced. Especially when facing complex data distribution changes, it can better capture diverse features, thereby improving the stability and accuracy of the final decision.

[0036] 2. The present invention randomly introduces local and global perturbations in the shallow layer of the network, and adjusts the perturbed features through a designed correction network, enabling the model to learn more robust feature representations and improving the perception ability of abnormal regions. And a cyclic correction mechanism is adopted to further optimize the feature representation, making the model show stronger adaptability and higher diagnostic accuracy when dealing with diverse attack means. This method enhances the adaptability of the model to different environmental conditions and avoids overfitting to specific data patterns. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained according to the provided drawings.

[0038] Figure 1 It is a schematic flow chart of the method of the present invention;

[0039] Figure 2 It is a schematic structural diagram of the medical image analysis network model of the present invention. Specific embodiments

[0040] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0041] The purpose of the present invention is to provide a medical image analysis method, system, electronic device, and readable storage medium, including: 1) using feature extractors that are isomorphic but have different initial settings (for example, ResNet, DenseNet, etc. based on the convolutional neural network CNN) to obtain the feature representation of the input image; 2) randomly introducing local perturbations and global perturbations according to a preset probability in the shallow layers of the network (such as the 1st, 2nd, and 3rd layers); 3) each feature extractor generates two branches: one retains the original features, and the other perturbs the features; 4) adjusting the perturbed features through a correction network, and further optimizing the feature representation by adopting a cyclic correction mechanism; 5) using the idea of ensemble learning to fuse the prediction results of the two decoders; 6) using a variety of loss functions such as cross-entropy loss, similarity loss, and alignment cycle loss to guide model optimization; 7) in the adaptive stage during testing, when encountering new target domain samples, the system performs fine-tuning by only updating some network parameters (such as the correction network parameters in the feature extractor) through operations similar to those in the training stage. The present invention can improve the robustness and generalization ability of the model in the image classification task, and can show stronger adaptability and higher diagnostic accuracy when facing data distribution shifts and diverse attack means.

[0042] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the drawings and specific embodiments.

[0043] See Figure 1-2As shown, an embodiment of the present invention discloses a medical image analysis method, including:

[0044] S100: Obtain a lung CT image with a true label for pneumonia prediction;

[0045] S200: Construct a medical image analysis network model, which includes an image encoder, a feature perturbation module, a feature correction module, a cyclic correction module, and a dual adaptive network;

[0046] S300: Classify the input lung CT image through the medical image analysis network model;

[0047] Among them, the image encoder is used to extract features from the input lung CT image to obtain a feature map; the feature perturbation module is used to apply local perturbations or global perturbations to the feature map in the shallow layer of the medical image analysis network model to obtain perturbed features; the feature correction module is used to correct the perturbed features to obtain corrected features; the cyclic correction module is used to iteratively apply a correction network to perform secondary correction on the corrected features to obtain a statistically represented cyclic correction; the dual adaptive network includes two sub-networks constructed based on the same ResNet architecture but with independently adjusted parameters, which are respectively used to process the feature map, generate their respective image classification results and output, and finally integrate the output.

[0048] In a specific embodiment, first obtain a true label for a lung CT image for detecting pneumonia; use a general image encoder to perform feature extraction respectively; obtain a feature map with a dimension of ; where represents the channel dimension of the feature map, and represent the height and width of the feature map respectively.

[0049] Then, construct a feature perturbation module, including: a local perturbation module, a global perturbation module, and a global statistic update mechanism. Randomly apply local perturbations or global perturbations to the feature map at a specific network layer according to a preset probability to simulate image changes under different conditions. Local perturbations simulate minor changes in the lesion area of the image, and global perturbations simulate the effects of factors such as overall image quality or lighting. The specific steps are as follows:

[0050] Step 2.1: Given a feature map , calculate the feature channel statistics, that is, the average value and the standard deviation along each channel dimension;

[0051] Step 2.2. At the 1st, 2nd, and 3rd layers of the network, with the channel dimensions being [64, 128, 256], randomly introduce local perturbations and global perturbations according to the conditional probability. If , then apply local perturbations; if , then apply global perturbations; only when , neither of which satisfies the conditional probability of perturbation, the feature is not subjected to any perturbation processing, and the perturbed feature ;

[0052] Step 2.3. The local perturbation creates two tensors with the same dimension as the mean of the feature map, namely, a zero matrix and a one matrix of [1, C, 1, 1]. Randomly generate noise coefficients and from a normal distribution with a mean of 0 and a standard deviation of 0.75. controls the difference between the feature map and its mean, while controls the contribution of the mean to the feature map.

[0053] Step 2.4. Use two perturbation factors and to control the Gaussian noise injection process of and :

[0054]

[0055]

[0056] In Equation (1), represents the mean of the perturbation; represents the variance of the perturbation.

[0057] Step 2.5. Replace the feature channel statistics of the original feature map with the perturbed channel statistics . Using the idea of adaptive instance normalization (AdaIN), apply the perturbed statistics to the original feature map to obtain the perturbed feature map , which not only retains the structure of the original feature map but also introduces the changes brought by the perturbation. The calculation formula is as follows:

[0058]

[0059] In Equation (2), represents the perturbed feature map.

[0060] Step 2.6: If global perturbations are applied, it is necessary to determine the dataset statistics datum_center and datum_var of the current layer. Similar to the local perturbation module, the perturbation mean and standard deviation are calculated using the perturbation factor to replace the original statistics, and the perturbed feature map is obtained using Equation (2).

[0061] In a specific embodiment, a feature correction module is constructed to correct the bias caused by perturbations, ensuring the consistency and stability of feature representations. The purpose is to correct different target domain features to align with the source domain, including creating corresponding sequence models according to the output channel dimensions [128, 256, 512] of the perturbation layer. Specifically, for each perturbation layer, the created sequence model includes the following components: three convolutional layers Conv2d, two custom normalization layers LayerNorm2d, and two non-linear activation functions ReLU; among them, the custom normalization layer: normalizes the feature map of each channel, calculates the mean of each channel and standard deviation , and the specific process is as follows:

[0062] Step 3.1: Standardize the feature map according to the normalization formula, and use learnable parameters: weights (initialized to all 1s) and biases (initialized to all 0s) to scale and translate the standardized feature map. The learnable parameters allow the model to adaptively adjust the distribution of the feature map, thereby retaining important feature information.

[0063] Step 3.2: The feature correction module uses the perturbed features and their statistics to obtain the preliminary corrected features and their correction factors , .

[0064]

[0065]

[0066] In Equation (3), represents the mean of , represents the standard deviation of , is a constant to prevent division by zero error, usually taking the value .

[0067] Step 3.3: Based on the correction factor, the feature correction module uses Equation (4) to obtain the new mean and standard deviation, replacing the original statistics:

[0068]

[0069]

[0070] In Equation (4), represents the mean of the calibration features; represents the standard deviation of the calibration features.

[0071] Step 3.4. The feature calibration module uses Equation (5) to normalize it to zero mean and unit variance.

[0072]

[0073] The feature calibration module uses Equation (6) to rescale and translate the normalized feature map using the updated mean and standard deviation to achieve the final calibration:

[0074] (6)

[0075] In a specific embodiment, a cyclic correction module is constructed to iteratively apply the perturbation mechanism and the correction network to the features that have been calibrated once for secondary correction to further optimize the feature representation and improve the model's learning ability for complex patterns. The specific process is as follows:

[0076] Step 4.1. Determine whether to perform cyclic correction on the feature map according to the random probabilities of local perturbation and global perturbation. When , that is, the probability condition of local perturbation is satisfied, the local perturbation parameters and are used for cyclic correction; if but , that is, only the probability condition of global perturbation is satisfied, the global perturbation parameters and are used for cyclic correction; if neither condition is satisfied, no correction is performed. If and are both less than 0.5, the condition of will be triggered first. This means that in this case, the local perturbation parameters ( and ) will be preferentially used for cyclic correction.

[0077] Step 4.2. Process the mean and standard deviation of the calibration features according to the perturbation parameters.

[0078]

[0079]

[0080] Step 4.3. Normalize, rescale, and translate: Normalize to zero mean and unit variance, and then rescale and translate using the updated mean and standard deviation to generate a new distribution .

[0081]

[0082] Step 4.4: Pass to the domain adapter of the corresponding correction network, and avoid gradient calculation through the detach() method of the tensor to ensure that the gradient flow of the original feature map is not changed, and generate a secondarily processed feature map 。

[0083] The detach() method returns a new tensor that has the same value as the original tensor but is no longer associated with other operations in the computational graph, so it will not be used to calculate gradients.

[0084] Step 4.5: Calculate the mean and standard deviation of the feature map .

[0085]

[0086]

[0087] Step 4.6: Update the final mean and standard deviation, adjust based on to obtain the final mean ; adjust based on to obtain the final standard deviation for subsequent loss calculations.

[0088]

[0089]

[0090] In Equation (10), represents the feature mean after cyclic correction; represents the feature standard deviation after cyclic correction.

[0091] Step 5: Construct a dual adaptive network, including: a dual adaptive encoder (ResNet18), and a decoder including a simple linear layer;

[0092] Step 5.1: Construct two parallel and The image encoder, where two models share the same network configuration ResNet18, but have randomly different initialization values, learning rates, or optimizers, thus introducing diverse and robust feature representations.

[0093] Step 5.2: Input the image X into the two encoders respectively to extract multi-level feature representations, and process the features and to obtain and , and , and then output the integrated prediction output through two identical decoders ,

[0094]

[0095]

[0096] In Equation (11), represents and the mean of the predicted labels regarding the original features output by the encoders, represents and the mean of the predicted labels regarding the corrected features output by the encoders.

[0097] Step 6: A medical image analysis network is composed of an image encoder, a feature perturbation module, a feature correction module, a cyclic correction module, and a dual adaptive ResNet model, and the cross-entropy loss function of the medical image analysis network is constructed using Equations (12), (13), (14), and (15) respectively , to minimize the difference between the predicted output and the actual label; the similarity loss function ; to measure the similarity and difference of samples in the feature space and reduce the feature distribution difference between the source domain and the target domain; the alignment cyclic loss and are used to constrain the adapter, aiming to enable the adapter to correct the perturbed features back to the source domain space, improve the adaptability and stability of the model in a changing environment, and at the same time ensure that the learned features are both discriminative and maintain internal consistency:

[0098]

[0099]

[0100] In Equation (13), represents the original features in the first encoder , represents the second encoder The enhanced features after perturbation correction in represent the second encoder The original features in represent the first encoder The enhanced features after perturbation correction in

[0101] (14)

[0102] In formula (14), C represents the number of network layers of perturbation, and N represents the number of encoders.

[0103]

[0104] In formula (15), are the features after two corrections.

[0105] Step 7: Train the medical image analysis network through the Adam optimizer, and calculate the total loss function using formula (16) to update the network parameters until the loss function converges, so as to obtain the optimal medical image analysis model for classifying the input lung images and outputting the corresponding disease predictions.

[0106]

[0107] In formula (16), represents the weight of represents and the weights of

[0108] Step 8: In the adaptive stage during testing, the network is still in the training mode. Obtain a target domain image without a label , and continue training according to the above steps. Fine-tune some network parameters, and only allow the correction network and its related parameters to participate in the gradient descent update, while keeping the remaining network parameters unchanged. At the same time, continuously update the global statistical data.

[0109]

[0110] In formula (17), represents the learning rate of the network in the adaptive stage during testing.

[0111] On the other hand, an embodiment of the present invention also discloses an electronic device, including a memory and a processor. The memory is used to store a program that supports the processor to execute the above method, and the processor is configured to execute the program stored in the memory.

[0112] On the other hand, an embodiment of the present invention also discloses a computer-readable storage medium, on which a computer program is stored. When the computer program is run by a processor, the steps of the above method are executed.

[0113] For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple. For related parts, refer to the description in the method part.

[0114] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A medical image analysis method, characterized in that Including: S100: Obtain lung CT images with real labels for pneumonia prediction; S200: Construct a medical image analysis network model, which includes an image encoder, a feature perturbation module, a feature correction module, a cyclic correction module, and a dual adaptive network; S300: Classify the input lung CT images through the medical image analysis network model; Among them, the image encoder is used to extract features from the input lung CT images to obtain feature maps; the feature perturbation module is used to apply local perturbations or global perturbations to the feature maps in the shallow layer of the medical image analysis network model to obtain perturbed features; the feature correction module is used to correct the perturbed features to obtain corrected features; the cyclic correction module is used to iteratively apply the correction network to perform secondary correction on the corrected features to obtain a statistically represented cyclic correction; the dual adaptive network includes two sub-networks constructed based on the same ResNet architecture but with independently adjusted parameters, which are respectively used to process the feature maps, generate their respective image classification results and output them, and finally integrate them; The feature perturbation module includes: a local perturbation module, a global perturbation module, and a global statistic update mechanism; the processing steps of the feature perturbation module include: calculating feature channel statistics based on the feature maps; randomly introducing local perturbations and global perturbations according to conditional probabilities at specific network layers; for local perturbations, first create a zero matrix and a one matrix with the same dimension as the mean of the feature maps, and then randomly generate noise coefficients from a normal distribution with a mean of 0 and a standard deviation of 0.75; the noise coefficients are used to control the injection process of Gaussian noise, and the feature channel statistics of the original feature maps are replaced with the perturbed channel statistics; through adaptive instance normalization, the perturbation parameters are applied to the affine transformation of the features, thereby realizing the adjustment of the feature maps; if global perturbations are applied, determine the dataset statistics of the current layer, calculate the perturbed mean and standard deviation using the global perturbation factor, and similar to local perturbations, further generate perturbed feature maps.

2. The medical image analysis method according to claim 1, wherein, The image encoder is used to extract features from the input lung CT image to obtain a feature map with a dimension of ; where C represents the channel dimension of the feature map, and H and W represent the height and width of the feature map respectively. ​ 3. The medical image analysis method according to claim 1, wherein, The feature correction module creates a corresponding sequence model from several perturbation layers, including: three convolutional layers Conv2d, two custom normalization layers LayerNorm2d, and two non-linear activation functions ReLU; among them, the custom normalization layer: normalizes the feature maps of each channel and calculates the mean and standard deviation of each channel.

4. A medical image analysis method according to claim 1, wherein, The cyclic correction module includes: Determine whether to perform correction according to the random probabilities of local perturbations and global perturbations, and when the conditions are met, process them with the corresponding perturbation parameters respectively; Adjust the mean and standard deviation of the corrected features according to the perturbation parameters; Perform standardization and rescaling translation operations to generate a new feature distribution; Transfer the processed features to the domain adapter of the correction network, and use the separated tensor method to avoid gradient calculation to ensure that the gradient flow remains unchanged; Calculate and update the mean and standard deviation of the feature maps for subsequent loss calculation.

5. A medical image analysis method according to claim 1, characterized in that, The dual adaptive network includes two encoders with the same structure but different values, and their respective decoders, which are used to map features to the final classification results; Among them, the two encoders with the same structure but different values share the same network configuration, but the initial values, learning rates or optimizers are randomly different; The lung CT images are respectively input into the two dual adaptive encoders to extract multi-level feature representations, and then the integrated prediction outputs are output through the two identical decoders.

6. The medical image analysis method according to claim 1, wherein The medical image analysis network is also trained using the cross-entropy loss function, similarity loss function and alignment cycle loss function until the loss function converges, so as to obtain the optimal medical image analysis model for classifying and outputting the input lung images.

7. A medical image analysis system using the medical image analysis method according to any one of claims 1 to 6, characterized in that, It includes: Data acquisition module: Acquire lung CT images with real labels for pneumonia prediction; Model construction module: Construct a medical image analysis network model, which includes an image encoder, a feature perturbation module, a feature correction module, a cycle correction module and a dual adaptive network; Recognition and detection module: Classify the input lung CT images through the medical image analysis network model; Among them, the image encoder is used to extract features from the input lung CT images to obtain feature maps; the feature perturbation module is used to apply local or global perturbations to the feature maps in the shallow layer of the medical image analysis network model to obtain perturbed features; the feature correction module is used to correct the perturbed features to obtain corrected features; the cycle correction module is used to iteratively apply the correction network to perform secondary correction on the corrected features to obtain the statistical representation after cycle correction; the dual adaptive network includes two sub-networks constructed based on the same ResNet architecture but with independently adjusted parameters, which are respectively used to process the feature maps, generate their respective image classification results and output them, and finally integrate them.

8. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports the processor to execute any one of the medical image analysis methods in claims 1 to 6, and the processor is configured to execute the program stored in the memory.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is run by the processor, it executes the steps of any one of the medical image analysis methods in claims 1 to 6.

Images