A medical image classification system based on feature interaction and adaptive feature fusion

Abstract

The application provides a kind of medical image classification system based on feature interaction and adaptive feature fusion in the technical field of computer vision and artificial intelligence, for: obtaining historical medical image dataset, and the medical image dataset is preprocessed and divided, to obtain training set, verification set and test set;Build the classification network model including first feature extraction branch, second feature extraction branch and feature fusion module;Through the training set, verification set and test set, the classification network model is trained, verified and tested respectively, and the classification network model that passes the test is deployed;Get the medical image to be classified, input the medical image to be classified into the deployed classification network model, to obtain the corresponding classification result.The application has the advantages that: the accuracy and robustness of medical image classification are greatly improved.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and artificial intelligence, and in particular to a medical image classification system based on feature interaction and adaptive feature fusion. Background Technology

[0002] With the rapid development of deep learning technology, image classification methods based on convolutional neural networks (CNNs) have been widely used in the field of medical image analysis. These methods aim to automatically extract discriminative features from medical images such as X-rays and computed tomography (CT) scans, and output corresponding category labels to provide technical support for auxiliary analysis.

[0003] However, existing deep learning-based medical image classification methods still face many technical bottlenecks when dealing with the unique challenges of medical images, which restricts further improvement in classification performance.

[0004] First, medical images commonly suffer from strong imaging noise interference. To extract high-level semantic features, mainstream CNN models typically rely on successive downsampling operations (such as pooling or strided convolution) to expand the receptive field. However, this layer-by-layer downsampling is a double-edged sword. While achieving feature abstraction and dimensionality compression, it inevitably leads to the loss of high-frequency details in the image and may introduce spectral aliasing during the downsampling process. The loss of high-frequency details blurs key structural and textural information in the image, while spectral aliasing distorts feature representations, which is particularly detrimental to medical image classification tasks that require precise feature discrimination.

[0005] Secondly, target regions (such as lesions) in medical images often exhibit significant scale variations. Standard convolution operations, limited by their fixed, local receptive fields, struggle to effectively capture features at different scales simultaneously. Specifically, small receptive fields excel at extracting local details but are ill-suited for modeling broad contextual relationships; conversely, increasing the receptive field may overlook subtle structures. This inherent contradiction between prioritizing "local details" and "global semantics" limits the feature representation capabilities of single-architecture models when dealing with multi-scale targets.

[0006] Finally, in the feature fusion stage, existing methods typically employ simple element-wise addition or channel concatenation operations to integrate features from different depths or branches of the network. This "one-size-fits-all" fusion strategy treats all feature channels equally, ignoring the differences in the importance of the information contained in different feature channels. Deep feature maps are rich in high-level semantic information but have low spatial resolution and blurred details; shallow feature maps, while having higher spatial resolution and retaining more details, have lower semantic levels and are susceptible to noise interference. Undifferentiated fusion of them fails to achieve complementary advantages and may instead introduce redundant or noisy information, affecting the accuracy of the final classification decision.

[0007] In summary, existing technologies suffer from the following main drawbacks: 1) Traditional downsampling operations result in the loss of high-frequency details and aliasing noise when extracting high-level semantics; 2) Single model architectures (such as pure CNNs or pure Transformers) struggle to balance local detail perception with global context modeling; 3) Traditional feature fusion strategies cannot adaptively select and enhance key features while suppressing irrelevant information. These technical shortcomings collectively limit further improvements in the performance of medical image classification systems.

[0008] Therefore, how to provide a medical image classification system based on feature interaction and adaptive feature fusion to improve the accuracy and robustness of medical image classification has become an urgent technical problem to be solved. Summary of the Invention

[0009] The technical problem to be solved by this invention is to provide a medical image classification system based on feature interaction and adaptive feature fusion, so as to improve the accuracy and robustness of medical image classification.

[0010] This invention is implemented as follows: a medical image classification system based on feature interaction and adaptive feature fusion, comprising the following modules: The medical image dataset partitioning module is used to acquire historical medical image datasets and preprocess and partition the medical image datasets to obtain training sets, validation sets, and test sets. The classification network model building module is used to build a classification network model that includes a first feature extraction branch, a second feature extraction branch, and a feature fusion module. The first feature extraction branch is used to perform hierarchical feature extraction and downsampling on the input medical image through multiple concatenated residual blocks, and to perform adaptive denoising processing based on discrete wavelet transform on the image features extracted at each layer to obtain the first branch features; The second feature extraction branch is used to extract second branch features from the input medical image by combining the CNN sub-branch with the Transformer sub-branch of the cross-feature channel attention mechanism; The feature fusion module is used to adaptively weight and fuse the first branch features and the second branch features through a multi-scale channel attention and gating mechanism to obtain fused features, process the fused features to extract deep features for classification, and output classification results based on the deep features. The classification network model training module is used to train, validate, and test the classification network model using the training set, validation set, and test set, respectively, and to deploy the classification network model that passes the test. The medical image classification module is used to acquire medical images to be classified, input the medical images to be classified into a deployed classification network model, and obtain the corresponding classification results.

[0011] Furthermore, the medical image dataset partitioning module is specifically used for: A historical medical image dataset containing a large number of medical images is obtained, and each medical image in the dataset is preprocessed in a standardized manner. The preprocessed historical medical image dataset is then divided into a training set, a validation set, and a test set according to a preset ratio.

[0012] Furthermore, in the classification network model construction module, the adaptive denoising process based on discrete wavelet transform is performed on the image features extracted from each layer to obtain the first branch features, specifically as follows: The image features after downsampling are subjected to discrete wavelet transform to decompose them into low-frequency approximate components and multiple high-frequency detail components. An adaptive gating system is designed for each of the high-frequency detail components to dynamically adjust the weight coefficients of each high-frequency detail component according to the feature content of the image features, thereby achieving noise suppression and key feature enhancement. The weighted low-frequency approximation components and high-frequency detail components are subjected to inverse discrete wavelet transform to reconstruct a denoised feature map. The denoised feature map is residually fused with the original features before downsampling to obtain the first branch feature.

[0013] Furthermore, in the classification network model construction module, the extraction of second-branch features from the input medical image by combining the CNN sub-branch with the Transformer sub-branch using a cross-feature channel attention mechanism specifically involves: The medical image is input into the CNN sub-branch and Transformer sub-branch of the second feature extraction branch to obtain the CNN feature map and Transformer feature map, respectively. Channel compression is performed on the CNN feature map and Transformer feature map to obtain CNN channel feature vectors and Transformer channel feature vectors; The CNN channel feature vectors and Transformer channel feature vectors are excited and compressed to obtain enhanced CNN attention feature vectors and Transformer attention feature vectors; Construct a cross-feature channel correlation matrix based on the CNN attention feature vector and the Transformer attention feature vector; Based on the cross-feature channel correlation matrix, the CNN feature map is projected onto a subspace related to the Transformer feature map, and the Transformer feature map is projected onto a subspace related to the CNN feature map, so as to interact the CNN feature map and the Transformer feature map to obtain the second branch feature.

[0014] Furthermore, before outputting the second branch features, spatial-level fusion is performed on the interacted CNN feature map and Transformer feature map, specifically as follows: The interactive CNN feature map is subjected to large kernel convolution to enhance spatial dependencies and concatenated with the interactive Transformer feature map; and the interactive Transformer feature map is subjected to small kernel convolution to extract local details and weighted summed with the interactive CNN feature map.

[0015] Furthermore, in the classification network model construction module, the adaptive weighted fusion of the first branch features and the second branch features through multi-scale channel attention and gating mechanisms to obtain the fused features specifically involves: Add the first branch feature to the second branch feature to obtain the summed feature; The additive features are input into the multi-scale channel attention module, and multi-scale channel attention weights for the additive features are generated by combining local channel attention and global channel attention. The multi-scale channel attention weights are used as gating weights to weight and concatenate the first branch features and the second branch features to obtain the fused features.

[0016] Furthermore, in the classification network model construction module, the classification network model adopts an encoder-decoder architecture, where the first feature extraction branch and the second feature extraction branch constitute the encoder and are connected to the decoder through skip connections.

[0017] Furthermore, in the classification network model construction module, the process of processing the fused features to extract deep features for classification specifically involves: The fused features are sequentially upsampled and decoded by at least one decoder to gradually restore the image resolution and output deep features for classification.

[0018] Furthermore, the classification network model training module is specifically used for: The classification network model is trained using the training set, validated and hyperparameters tuned using the validation set, tested using the test set, and then deployed using the tested classification network model.

[0019] The advantages of this invention are: 1. First, adaptive denoising based on discrete wavelet transform is used to selectively suppress noise and enhance key details during feature extraction downsampling, thus preserving cleaner and more informative features. Second, a dual-branch interactive network integrating the local perception capabilities of CNN and the global modeling capabilities of Transformer is constructed to achieve complementary and deep fusion of local details and global contextual information, enabling the model to understand image content more comprehensively. Finally, an adaptive feature fusion module based on multi-scale channel attention and gating mechanism is designed. This module can dynamically evaluate and weight the importance of feature channels from different branches, thereby intelligently selecting and strengthening the most discriminative features and suppressing redundant or interfering information. These three technologies work in a progressive manner to ensure that the model can extract more robust and discriminative deep features from complex and varied medical images, ultimately achieving more accurate and reliable classification.

[0020] 2. By employing a dual-branch heterogeneous feature extraction architecture, the completeness and complementarity of feature representation are achieved, enhancing the model's feature representation capabilities: An innovative heterogeneous feature extraction network consisting of a first feature extraction branch and a second feature extraction branch is designed. The first branch, based on a residual network and incorporating an adaptive denoising mechanism in the frequency domain, excels at extracting robust, noise-free hierarchical structural features. The second branch creatively integrates CNN and Transformer sub-branches through a cross-feature channel attention mechanism, enabling simultaneous capture of local detail features and global contextual dependencies. This dual-branch design allows the model to understand medical images from different dimensions and in different ways, ensuring comprehensive and multi-faceted analysis of input information. This lays a rich feature foundation for subsequent high-precision classification and effectively overcomes the limitation of feature extraction perspectives in single-architecture models.

[0021] 3. By employing adaptive denoising based on discrete wavelet transform, the robustness of the model to noise in medical images is significantly enhanced, while preserving key diagnostic information. Addressing common noise and artifact interference in medical images, an innovative adaptive denoising mechanism is introduced in the first feature extraction branch. This mechanism does not perform simple filtering but instead applies discrete wavelet transform to the downsampled feature map, decomposing it into components of different frequencies. An adaptive gating system is specifically designed for high-frequency detail components to dynamically distinguish noise from useful edges and texture information. By suppressing noise-corresponding components and enhancing key feature-corresponding components before reconstruction, accurate and adaptive denoising at the feature level is achieved. This method effectively purifies features, reduces misjudgments caused by noise, and avoids feature blurring or loss that may occur with traditional denoising methods, ensuring high quality and high fidelity of features used in subsequent processing.

[0022] 4. Deep feature interaction between CNN and Transformer is achieved through a cross-feature channel attention mechanism, giving full play to the synergistic advantages of the two architectures: In the second feature extraction branch, instead of simply parallelizing or concatenating CNN and Transformer, a deep interaction mechanism is proposed. This mechanism guides the two features to be projected into each other's related subspaces by constructing a cross-feature channel correlation matrix between the feature vectors of CNN and Transformer. This process enables the local detail features extracted by CNN to be integrated into the global relevance information that Transformer focuses on, while also enabling the global context features of Transformer to focus on the local important regions emphasized by CNN. This interaction achieves true complementarity and integration of the advantages of the two mainstream architectures, producing a feature enhancement effect of "1+1>2", enabling the extracted second branch features to have both strong local descriptive ability and global semantic understanding ability.

[0023] 5. By introducing an adaptive feature fusion strategy with multi-scale channel attention and gating mechanisms, dynamic optimization of fusion weights is achieved: In the feature fusion module, the traditional fixed weight addition or splicing method is abandoned, and a more refined adaptive weighted fusion strategy is adopted. This strategy first analyzes the addition results of the first and second branch features from multiple scales, such as local and global, through the multi-scale channel attention module to generate attention weights that can accurately reflect the importance of each channel. Then, these weights are used as gating to perform weighted splicing of the features of the two branches. This mechanism enables the network to dynamically and discriminatively evaluate and fuse information from the two branches according to the specific content of the current input image. For example, in a certain region, the denoised structural features are trusted more, while in another region, the contextual features after interaction are relied upon more, thereby realizing the intelligent and optimized information fusion and greatly improving the representation quality of the fused features.

[0024] 6. Construct a complete, end-to-end automated classification system from data preparation to model deployment, which is highly practical and easy to implement: It not only focuses on the innovation of the core network model structure, but also completes the entire process system including data partitioning, model building, training and testing, and final application; it clarifies the standardized process from historical data preprocessing and partitioning to model training, verification, testing and deployment, forming a closed-loop solution. This end-to-end system design ensures the repeatability and engineering convenience of the technical solution. Users do not need to design additional complex data processing or model integration pipelines, and can directly apply it to actual medical image classification tasks, which significantly reduces the technical application threshold. Attached Figure Description

[0025] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0026] Figure 1 This is a schematic diagram of the structure of a medical image classification system based on feature interaction and adaptive feature fusion according to the present invention.

[0027] Figure 2 This is an architecture diagram of the classification network model of this invention.

[0028] Figure 3 This is a schematic diagram of the adaptive denoising based on discrete wavelet transform of the present invention.

[0029] Figure 4 This is a schematic diagram of the dual-flow feature interaction of the present invention.

[0030] Figure 5 This is a schematic diagram of the attention-gated fusion of the present invention. Detailed Implementation

[0031] The overall approach of the technical solution in this application is as follows: First, the historical medical image dataset is standardized and preprocessed, and divided into training, validation, and test sets to provide a data foundation for model learning. Next, a classification network model is constructed. The core design idea of ​​the classification network model is "divide and conquer, intelligent fusion": the first branch (first feature extraction branch) focuses on purifying the downsampled features from a frequency domain perspective through discrete wavelet transform and adaptive gating, enhancing their resistance to noise interference and extracting purer hierarchical structural features; the second branch (second feature extraction branch) is dedicated to fusing the advantages of different modalities. Through a cross-feature channel attention mechanism, CNN, which excels in local perception, and Transformer, which excels in global modeling, engage in deep feature interaction, thereby extracting complementary features that simultaneously contain rich details and global semantics. Subsequently, through a multi-scale channel attention mechanism in the feature fusion module, the importance of heterogeneous features from the two branches is evaluated, and a gating mechanism is used to achieve dynamic and discriminative weighted fusion, generating a more discriminative fused feature. This fused feature is then upsampled and decoded by the decoder, finally outputting the classification result. The constructed model is optimized and evaluated through a standardized training, validation, and testing process. Once the test is passed, it can be deployed to automatically and accurately classify new medical images to be classified.

[0032] Please refer to Figures 1 to 5 As shown, a preferred embodiment of the medical image classification system based on feature interaction and adaptive feature fusion of the present invention includes the following modules: The medical image dataset partitioning module is used to acquire historical medical image datasets and preprocess and partition the medical image datasets to obtain training sets, validation sets, and test sets. In practice, the public CXR datasets GWCMC and HWCMC can be selected and divided into training set, validation set and test set in a ratio of 8:1:1. The classification network model building module is used to build a classification network model that includes a first feature extraction branch, a second feature extraction branch, and a feature fusion module. The first feature extraction branch is used to perform hierarchical feature extraction and downsampling on the input medical image through multiple concatenated residual blocks (in specific implementation, four ResNet blocks can be concatenated), and to perform adaptive denoising processing based on discrete wavelet transform on the image features extracted at each layer to obtain the first branch features; In medical image classification tasks, accurately extracting local detail features from images is crucial. Common convolutional classification models typically require layered convolution and downsampling operations on input features to extract multi-scale local features. However, layer-by-layer downsampling leads to the loss of detail information, resulting in aliasing. This is because high-frequency detail features undergo spectral aliasing during downsampling, being incorrectly folded into lower frequency bands. This causes the subtle structures of lesions (such as ground-glass opacities and abnormal textures in pneumonia) to be blurred or distorted, severely impairing the model's ability to extract and discriminate fine-grained lesion features. Frequency domain denoising methods can effectively separate effective signals from noise interference in the frequency dimension. To address the aliasing and noise accumulation problems caused by downsampling, a wavelet denoising residual convolution (WDRC) module was constructed, aiming to suppress high-frequency noise while preserving key lesion detail features to the greatest extent. Considering that the downsampling operation in convolution can effectively perform feature abstraction and dimensionality compression, which is a key step in deep networks to extract high-level semantics, the downsampling path in convolution is retained, and the wavelet denoising residual module is introduced serially only after this operation. This approach retains the powerful feature representation capabilities of deep networks while effectively eliminating aliasing noise introduced during downsampling, achieving a synergistic improvement in feature abstraction efficiency and signal-to-noise ratio.

[0033] Specifically, the input medical image is H and W represent the height and width of the medical image, respectively, and C represents the number of channels. Represents real numbers. Function The function represents the output after convolution, batch normalization, and activation function layers; the function Wavelet represents the output of the wavelet denoising module; the function ResBlock represents the output of the convolution module, as shown in the following formula: ; ; ; in, Indicates wavelet transform; Indicates inverse wavelet transform; This represents an adaptive gating / filtering mechanism for high-frequency subband signals based on a learnable threshold. and Indicates learnable weights; Indicates pooling operation; The specific operation of wavelet denoising is as follows: For input X, it is first decomposed into a low-frequency approximation component LL and three high-frequency detail components (including LH, HL, and HH) through discrete wavelet transform. Among them, the low-frequency approximation component LL preserves the overall structure and contour information of the image, while the high-frequency detail components LH, HL, and HH are used to capture edge texture details in the horizontal, vertical, and diagonal directions, respectively.

[0034] To address the characteristic of subtle texture variations in lesion regions of medical images, such as pneumonia images, an adaptive gating mechanism was designed for each high-frequency sub-band. By learning the weight coefficients of each sub-band, noise interference is dynamically suppressed and key features related to the lesion are enhanced. Specifically, the adaptive gating automatically adjusts the contribution of each sub-band based on the feature content, removing redundant information while retaining high-frequency details beneficial to the classification task. The four sub-bands after gating and weighting are reconstructed into enhanced feature maps through inverse discrete wavelet transform (InverseDWT, IDWT), and residual fusion is performed with the original features before pooling. This achieves maximum preservation of fine-grained lesion information while reducing spatial resolution, as shown below: ; in, This represents the output signal, i.e., the characteristics of the first branch; This represents the learnable signal threshold, which is adaptively adjusted as the model is trained. The second feature extraction branch is used to extract second branch features from the input medical image by combining the CNN sub-branch with the Transformer sub-branch of the cross-feature channel attention mechanism; Convolutional neural networks (CNNs) excel at capturing rich local details, while Transformers excel at modeling long-range global dependencies; their strengths are significantly complementary. To fully leverage this complementarity, a correlation matrix is ​​constructed based on channel feature descriptors to identify correlations between heterogeneous features. This aims to filter redundant information and refine feature representations, thereby achieving bidirectional feature interaction between the CNN and Transformer branches. Specifically, a Dual-Stream Feature Interaction Network (DFIN) is designed, integrating a Cross-Feature Channel Attention (CFCA) module into the parallel dual-stream architecture of CNN and Transformer. By establishing correlation mappings between channels, local structural information is transferred from CNN to Transformer, and global contextual information flows from Transformer to CNN. This layer-by-layer bidirectional interaction mechanism not only strengthens the fusion and mutual guidance of multi-scale features but also effectively improves the model's ability to extract key discriminative features, ultimately significantly improving the model's feature representation quality and classification performance.

[0035] Specifically, the two feature maps are refined by constructing a correlation matrix based on channel features. The outputs of the CNN and Transformer modules first extract multi-channel feature maps, respectively. and .in, and These represent the number of feature channels for the CNN and Transformer, respectively. Next, the multi-channel feature maps are compressed into channel feature vectors, where each element represents the feature of each channel. To achieve efficient compression, adaptive average pooling is used, mapping the multi-channel feature maps into a one-dimensional vector by calculating the average of the feature maps for each channel. Then, the multi-channel feature maps M and N can be compressed as follows: ; ; in, and Let M and N represent the compressed feature vectors for channels M and N, respectively. Generally, the feature map extracted by the Transformer has more channels than that extracted by the CNN. To facilitate subsequent cross-channel focus on establishing the correlation between high and low channel features, the channel feature vectors of the CNN and Transformer were subjected to activation-then-compression and compression-then-activation operations, respectively. The activation-then-compression operation first uses a linear transformation to... The process involves mapping to a higher-dimensional space, followed by the ReLU activation function, and finally compressing it back to the original-dimensional space through another linear transformation. ; in, Used to remap (compress) high-dimensional features after ReLU activation back to the original channel dimension. ; Used to input feature vectors Mapped to a higher-dimensional space (“Induction” phase); This represents the attention feature vector of the CNN branch obtained after the "excitement-then-compression" operation; This represents the Sigmoid activation function, which maps the linearly transformed values ​​to the (0,1) interval to generate attention weights.

[0036] Similarly, the compression-then-excitation operation first transforms the linear transformation... The space is compressed to a lower dimension, followed by the ReLU activation function, and finally compressed back to the original dimension through another linear transformation: ; in, This represents the attention feature vector of the Transformer branch obtained after the "excitement-then-compression" operation.

[0037] Then, we begin constructing the cross-feature channel correlation matrix. Through the correlation matrix Q, we can identify the correlated channel features between the CNN and Transformer feature maps. We then project M onto the subspace correlated with N, and N onto the subspace correlated with M. ; ; ; in, Used to normalize the channel correlation matrix; Represents a 1D tensor product; T represents the transpose; and Both represent projection features.

[0038] To enable CNN feature maps to contain global information and address the lack of local features in Transformer feature maps, the projected features are directly incorporated. and The features are weighted and added to the original features M and N respectively, thus achieving feature fusion, as shown in the following formula: ; ; in, This represents the CNN feature map after fusion with Transformer features; This represents the Transformer feature map after being fused with CNN features.

[0039] In addition to feature fusion at the channel level, CNN and Transformer features are also fused at the spatial level to better complement each other. Spatial differences between their feature representations are reduced by using convolutions. Specifically, 5×5 convolutions are applied... , and its feature dimensions from Convert to Then combine this feature with... Weighted concatenation is performed, where the 5×5 convolution provides a relatively large receptive field, thereby enhancing the CNN's ability to capture spatial dependencies. Meanwhile, in Perform a 3×3 convolution on top, reducing the channel dimension from Reduce to Using a smaller kernel helps extract finer local details to complement Transformer features. The generated features are then compared with... Weighted additions are performed. Finally, the fused results are concatenated and output. The process is shown in the following equation: ; ; ; in, This represents the final feature of the Transformer branch after spatial fusion processing; This represents the final features of the CNN branch after spatial fusion processing; This represents the final output feature of the Two-Stream Feature Interaction Network (DFIN) module, i.e., the second branch feature; This indicates a channel splicing operation.

[0040] The feature fusion module is used to adaptively weight and fuse the first branch features and the second branch features through a multi-scale channel attention and gating mechanism to obtain fused features. The fused features are then processed to extract deep features for classification (these features can come from texture features, edge information, gray-level distribution or topological structure in the image), and the classification result is output based on the deep features. Deep feature maps are rich in high-level semantic information but have low spatial resolution and suffer from severe loss of detail. Conversely, shallow feature maps have high spatial resolution but weak semantic expressive power and are easily affected by image detail noise. Feature fusion aims to integrate feature representations from different levels or branches through operations such as element-wise addition or channel concatenation. To compensate for the spatial details and correlation information lost during feature extraction, an Attention-Gated Fusion Module (AGFM) is designed. First, a Multi-Scale Channel Attention Module (MS-CAM) is introduced to optimize the features, suppressing responses in non-target regions while preserving contextual information, thereby optimizing the feature representation of the two branches. Subsequently, a gating mechanism is used to adaptively fuse the feature maps, suppressing redundant information and enhancing salient features to achieve efficient fusion of multi-branch features.

[0041] Specifically, given two feature maps Let Y be a feature map with a large receptive field. and The summation is used as input to the multi-scale channel attention module. Its core idea is to achieve local and global channel attention by changing the size of the spatial representation. Then... and Perform gating splicing as shown in the following formula: ; ; ; ; ; in, This indicates that the input feature maps are composed of two input feature maps. and The fusion feature obtained by adding elements one by one; This represents the output of the local channel attention branch, which is related to the input. A series of transformations are performed to capture the local dependencies between channels; B(.) represents the normalization operation; and Both represent pointwise convolutions, with kernel sizes of respectively and , is a "bottleneck" structure that first reduces dimensions and then increases them, used to model the relationship between channels, where r is the reduction ratio; Represents the ReLU activation function; This represents the global average pooling operation, which applies to the input feature map. The spatial dimensions (H, W) are compressed to obtain a channel feature vector, which is used to capture global context information; This represents the output of the global channel attention branch; This represents the final output of the multi-scale channel attention module, i.e., the generated attention weight map, with dimensions of [dimensionality missing]. The value is between 0 and 1, and is used to weight the feature importance of different channels and locations; This represents the final output of the attention-gated fusion module, i.e., the fused feature map.

[0042] To address the loss of high-frequency details and spectral aliasing effects caused by traditional convolutional downsampling, a wavelet transform denoising module (WDRC) was designed. This module is introduced serially after convolutional downsampling and decomposes the features into low-frequency approximate components and three high-frequency detail components through discrete wavelet transform. An adaptive gating mechanism is designed for each high-frequency sub-band to dynamically suppress noise interference and enhance key lesion features, achieving effective suppression of high-frequency noise and fidelity enhancement of feature information. Secondly, to solve the problem that a fixed receptive field cannot simultaneously capture local details and global distribution, a dual-stream feature interaction module based on channel attention (DFIN) is proposed. This module constructs... A parallel dual-stream architecture of CNN and Transformer was constructed, and a cross-feature channel attention mechanism was used to achieve bidirectional flow of local structural information and global contextual information, which significantly improved the model's ability to represent local and global features. In addition, in the feature fusion stage, an attention-gated fusion module (AGFM) was adopted to replace the traditional simple addition or splicing operation. This module adaptively weighted the fusion of multi-branch features through multi-scale channel attention and gating mechanism, suppressing redundant information and achieving efficient fusion of multi-branch features. Through the coordinated design of the three modules, the accuracy and robustness of medical image classification were significantly improved.

[0043] The classification network model training module is used to train, validate, and test the classification network model using the training set, validation set, and test set, respectively, and to deploy the classification network model that passes the test. The medical image classification module is used to acquire medical images to be classified, input the medical images to be classified into a deployed classification network model, and obtain the corresponding classification results.

[0044] The medical image dataset partitioning module is specifically used for: A historical medical image dataset containing a large number of medical images is obtained, and each medical image in the dataset is preprocessed in a standardized manner. The preprocessed historical medical image dataset is then divided into a training set, a validation set, and a test set according to a preset ratio.

[0045] In the classification network model construction module, the adaptive denoising process based on discrete wavelet transform is performed on the image features extracted from each layer to obtain the first branch features, specifically as follows: The image features after downsampling are subjected to discrete wavelet transform to decompose them into low-frequency approximate components and multiple high-frequency detail components. An adaptive gating system is designed for each of the high-frequency detail components to dynamically adjust the weight coefficients of each high-frequency detail component according to the feature content of the image features, thereby achieving noise suppression and key feature enhancement. The weighted low-frequency approximation components and high-frequency detail components are subjected to inverse discrete wavelet transform to reconstruct a denoised feature map. The denoised feature map is residually fused with the original features before downsampling to obtain the first branch feature.

[0046] In the classification network model construction module, the extraction of second-branch features from the input medical image by combining the CNN sub-branch with the cross-feature channel attention mechanism and the Transformer sub-branch specifically involves: The medical image is input into the CNN sub-branch and Transformer sub-branch of the second feature extraction branch to obtain the CNN feature map and Transformer feature map, respectively. Channel compression is performed on the CNN feature map and Transformer feature map to obtain CNN channel feature vectors and Transformer channel feature vectors; The CNN channel feature vectors and Transformer channel feature vectors are excited and compressed to obtain enhanced CNN attention feature vectors and Transformer attention feature vectors; Construct a cross-feature channel correlation matrix based on the CNN attention feature vector and the Transformer attention feature vector; Based on the cross-feature channel correlation matrix, the CNN feature map is projected onto a subspace related to the Transformer feature map, and the Transformer feature map is projected onto a subspace related to the CNN feature map, so as to interact the CNN feature map and the Transformer feature map to obtain the second branch feature.

[0047] Before outputting the second branch features, spatial-level fusion is performed on the interacted CNN feature map and Transformer feature map, specifically as follows: The interactive CNN feature map is subjected to large kernel convolution to enhance spatial dependencies and concatenated with the interactive Transformer feature map; and the interactive Transformer feature map is subjected to small kernel convolution to extract local details and weighted summed with the interactive CNN feature map.

[0048] In the classification network model construction module, the step of adaptively weighting and fusing the first branch features and the second branch features through a multi-scale channel attention and gating mechanism to obtain the fused features specifically involves: Add the first branch feature to the second branch feature to obtain the summed feature; The additive features are input into the multi-scale channel attention module, and multi-scale channel attention weights for the additive features are generated by combining local channel attention and global channel attention. The multi-scale channel attention weights are used as gating weights to weight and concatenate the first branch features and the second branch features to obtain the fused features.

[0049] In the classification network model construction module, the classification network model adopts an encoder-decoder architecture. The first feature extraction branch and the second feature extraction branch constitute the encoder and are connected to the decoder through skip connections.

[0050] In the classification network model construction module, the specific steps of processing the fused features to extract deep features for classification are as follows: The fused features are sequentially upsampled and decoded by at least one decoder to gradually restore the image resolution and output deep features for classification.

[0051] The classification network model training module is specifically used for: The classification network model is trained using the training set, validated and its hyperparameters tuned using the validation set, and tested using the test set. The tested classification network model is then deployed. During training, the loss function, optimizer function, and learnable hyperparameters are adjusted to find the optimal combination of hyperparameters that best performs on the test set, thus establishing the model.

[0052] To verify the effectiveness of this invention, experiments were conducted to compare it with eight existing methods on the GZCMC, HWCMC, and FJCH datasets. To ensure fairness, all models were configured uniformly. The experimental results show that this invention outperforms other models on all three datasets, demonstrating superior localization capabilities for lesion regions and resulting in better classification performance. Furthermore, the CNN-based method generally outperforms the Transformer-based method on all three datasets, indicating that in medical image classification tasks, local details are more important than global semantic information and have a greater impact on the model's final classification ability.

[0053] First, Table 1 shows the performance comparison between our invention and other mainstream classification models on the three-class classification dataset of the Women and Children's Medical Center of City A (GZCMC). This dataset contains three categories: normal, bacterial pneumonia, and viral pneumonia, and is a widely used benchmark dataset for pediatric pneumonia classification tasks. The experimental results show that our invention (Ours) achieved the best classification performance on the GZCMC dataset, with all metrics significantly outperforming the comparison methods. Specifically, our invention achieved an accuracy of 83.64%, precision of 83.42%, recall of 83.76%, F1 score of 83.61%, and specificity of 89.37%. Compared to the second-best performing VGG network, our invention improved accuracy from 81.94% to 83.64%, an improvement of 1.70 percentage points; and improved precision, recall, F1 score, and specificity by 2.36, 1.76, 2.57, and 0.80 percentage points, respectively. Compared to other traditional CNN models, the advantages of this invention are more significant: compared to ResNet (79.56%), the accuracy is improved by 4.08 percentage points; compared to ResNeXt (77.62%), it is improved by 6.02 percentage points; compared to Inception (80.07%), it is improved by 3.57 percentage points; compared to ConvNeXt (74.28%), it is improved by 9.36 percentage points; and compared to Xception (80.92%), it is improved by 2.72 percentage points. It is worth noting that the Vision Transformer series models (ViT, Swin, PVT) perform slightly lower than traditional CNN models on this dataset, with ViT achieving an accuracy of 79.73%, Swin 76.83%, and PVT 78.67%. This may be because the medical image dataset is relatively small, and the Transformer architecture's ability to handle large-scale data is not fully utilized. This invention, through a dual-stream collaborative design of CNN and Transformer, effectively combines the advantages of both, overcoming the limitations of a single architecture. Furthermore, this invention achieves a specificity of 89.37%, an improvement of 0.80%-6.41% compared to other methods, indicating that the model has a stronger ability to correctly identify normal samples and can effectively reduce the false diagnosis rate. In particular, the recall rate reaches 83.76%, demonstrating that the model can more comprehensively identify cases and reduce missed diagnoses. In summary, this invention demonstrates superior classification performance and good generalization ability on the GZCMC three-class classification task, validating the effectiveness of the proposed wavelet denoising, two-stream feature interaction, and attention-gated fusion modules.

[0054] Table 1: Performance of different methods on the GZCMC dataset

[0055] Next, comparative experiments were conducted on the five-class classification dataset of the Women and Children's Medical Center of Hubei Province (HWCMC), and the results are shown in Table 2. This dataset contains five categories: normal, bacterial pneumonia, mixed pneumonia, mycoplasma pneumonia, and viral pneumonia. The classification granularity is finer, and the task difficulty is relatively high. The performance of all models is significantly lower than that of binary and three-class classification tasks. The experimental results show that the present invention still achieves the best classification performance on this dataset, and all indicators are significantly better than the comparison methods. Specifically, the present invention achieves an accuracy of 56.53%, a precision of 56.32%, a recall of 56.41%, an F1 score of 56.25%, and a specificity of 75.15% on the HWCMC dataset. Compared with the second-best performing ResNet network, the present invention improves the accuracy from 54.62% to 56.53%, an improvement of 1.91 percentage points; and improves the precision, recall, F1 score, and specificity by 2.01, 1.68, 1.61, and 1.83 percentage points, respectively. Compared to other models, the advantages of this invention are more significant: compared to VGG (53.24%), the accuracy is improved by 3.29 percentage points; compared to ResNeXt (49.14%), it is improved by 7.39 percentage points; compared to Inception (52.81%), it is improved by 3.72 percentage points; and compared to Xception (51.34%), it is improved by 5.19 percentage points. It is worth noting that the VisionTransformer series models (ViT, Swin, PVT) performed relatively poorly on this five-class classification task, with accuracy rates all around 50%, with ViT at only 49.52% and Swin at 49.43%. This further verifies that in fine-grained medical image classification tasks, the pure Transformer architecture, due to its lack of local inductive bias and high data requirements, struggles to fully leverage its advantages on medium-sized datasets. This invention, through a two-stream collaborative design of CNN and Transformer, effectively overcomes this problem. Furthermore, this invention achieves a specificity of 75.15%, an improvement of 1.83%-5.10% compared to other methods, indicating its superior ability to correctly identify normal samples. Despite the challenging five-class classification task, this invention maintains a relatively leading advantage, fully demonstrating the effectiveness and robustness of the proposed wavelet denoising, two-stream feature interaction, and attention-gated fusion modules in handling complex multi-class medical image classification tasks, showcasing superior performance in more challenging classification scenarios.

[0056] Table 2: Performance of different methods on the HWCMC dataset

[0057] Secondly, comparative experiments were conducted on the binary classification dataset of Children's Hospital of C Province (FJCH), and the results are shown in Table 3. This dataset contains two categories: normal and pneumonia, and is an important benchmark for verifying the binary classification ability of the model in clinical applications. The experimental results show that our invention (Ours) achieved the best classification performance on this dataset, with all indicators significantly better than the comparative methods. Specifically, our invention achieved an accuracy of 82.75%, precision of 82.61%, recall of 82.54%, F1 score of 82.32%, and specificity of 81.73% on the FJCH dataset. Compared with the second-best performing VGG network, our invention improved the accuracy from 79.40% to 82.75%, an improvement of 3.35 percentage points; and improved precision, recall, F1 score, and specificity by 3.34, 3.15, 2.53, and 5.37 percentage points, respectively. Compared to other models, the advantages of this invention are more significant: compared to ResNet (77.27%), the accuracy is improved by 5.48 percentage points; compared to ResNeXt (76.68%), it is improved by 6.07 percentage points; compared to Inception (78.87%), it is improved by 3.88 percentage points; compared to ConvNeXt (76.47%), it is improved by 6.28 percentage points; and compared to Xception (75.75%), it is improved by 7.00 percentage points. It is worth noting that the VisionTransformer series models (ViT, Swin, PVT) perform slightly lower than traditional CNN models on this dataset, with ViT achieving an accuracy of 78.07%, Swin 77.54%, and PVT only 74.75%. This further verifies that in medical image binary classification tasks, the pure Transformer architecture, due to its lack of local inductive bias and high data requirements, struggles to fully leverage its advantages on medium-sized datasets. This invention, through a dual-stream collaborative design of CNN and Transformer, effectively combines the advantages of both. Furthermore, this invention achieves a specificity of 81.73%, an improvement of 5.37%-9.41% compared to other methods. This indicates that the model has a stronger ability to correctly identify normal samples and can effectively reduce the false diagnosis rate. In particular, the recall rate reaches 82.54%, demonstrating that the model can more comprehensively identify categories. In summary, this invention exhibits superior classification performance and good generalization ability on the FJCH binary classification task, validating the effectiveness and practicality of the proposed wavelet denoising, two-stream feature interaction, and attention-gated fusion modules in handling real-world clinical applications.

[0058] Table 3: Performance of different methods on the FJCH dataset

[0059] Ablation Experiment Results and Analysis Ablation experiments were conducted on the GZCMC dataset to verify the effectiveness of the current design by adjusting the wavelet denoising residual convolution module (WDRC), the two-stream feature interaction network (DFIM), and the attention-gated fusion module (AGFM). The experimental results are shown in Table 4. Without introducing any improved modules, the base model achieved an accuracy of 81.09%, a precision of 80.84%, a recall of 80.96%, an F1 score of 80.79%, and a specificity of 83.25%, providing a performance benchmark for subsequent improvements. When the wavelet denoising residual convolution module was introduced alone, all metrics showed significant improvement, with an accuracy of 81.84%, an improvement of 0.75 percentage points compared to the baseline model, and a specificity improvement to 84.19%. This indicates that the wavelet denoising residual convolution module can effectively suppress aliasing noise during downsampling, separate effective signals from interference in the frequency domain, thereby extracting purer feature information and improving the model's classification performance. By introducing a two-stream feature interaction network (AGFM) while retaining the baseline, the model performance continued to improve, achieving an accuracy of 82.34%, an F1 score of 82.16%, and a specificity of 85.31%. Compared to using only the baseline, the accuracy improved by 1.25%. This demonstrates that the AGFM effectively achieves bidirectional flow of local structural information and global contextual information by establishing channel correlation mapping between CNN and Transformer branches, enhancing the model's ability to extract key discriminative features. To verify the independent contribution of the attention-gated fusion module, a configuration containing only this module was tested. The results show that this configuration achieved an accuracy of 83.06%, an improvement of 1.97% compared to the baseline, and a specificity of 86.08%. This indicates that the AGFM module, through multi-scale channel attention and gating mechanisms, can adaptively filter redundant information and enhance salient features, effectively integrating multi-scale feature information and significantly improving the model's representational ability.

[0060] When all three modules are enabled, the model achieves optimal performance, with an accuracy of 83.64%, precision of 83.42%, recall of 83.76%, F1 score of 83.61%, and a peak specificity of 86.37%. Compared to the baseline model, accuracy is improved by 2.55%, and specificity by 3.12%. This result fully demonstrates that the wavelet denoising residual convolution module, the two-stream feature interaction network, and the attention-gated fusion module have significant complementarity; the synergistic effect of each module can maximize the discriminative ability; and the invention has significant effectiveness on the GZCMC dataset.

[0061] Table 4: Performance of different methods on the GZCMC dataset

[0062] In summary, the advantages of this invention are: 1. First, adaptive denoising based on discrete wavelet transform is used to selectively suppress noise and enhance key details during feature extraction downsampling, thus preserving cleaner and more informative features. Second, a dual-branch interactive network integrating the local perception capabilities of CNN and the global modeling capabilities of Transformer is constructed to achieve complementary and deep fusion of local details and global contextual information, enabling the model to understand image content more comprehensively. Finally, an adaptive feature fusion module based on multi-scale channel attention and gating mechanism is designed. This module can dynamically evaluate and weight the importance of feature channels from different branches, thereby intelligently selecting and strengthening the most discriminative features and suppressing redundant or interfering information. These three technologies work in a progressive manner to ensure that the model can extract more robust and discriminative deep features from complex and varied medical images, ultimately achieving more accurate and reliable classification.

[0063] 2. By employing a dual-branch heterogeneous feature extraction architecture, the completeness and complementarity of feature representation are achieved, enhancing the model's feature representation capabilities: An innovative heterogeneous feature extraction network consisting of a first feature extraction branch and a second feature extraction branch is designed. The first branch, based on a residual network and incorporating an adaptive denoising mechanism in the frequency domain, excels at extracting robust, noise-free hierarchical structural features. The second branch creatively integrates CNN and Transformer sub-branches through a cross-feature channel attention mechanism, enabling simultaneous capture of local detail features and global contextual dependencies. This dual-branch design allows the model to understand medical images from different dimensions and in different ways, ensuring comprehensive and multi-faceted analysis of input information. This lays a rich feature foundation for subsequent high-precision classification and effectively overcomes the limitation of feature extraction perspectives in single-architecture models.

[0064] 3. By employing adaptive denoising based on discrete wavelet transform, the robustness of the model to noise in medical images is significantly enhanced, while preserving key diagnostic information. Addressing common noise and artifact interference in medical images, an innovative adaptive denoising mechanism is introduced in the first feature extraction branch. This mechanism does not perform simple filtering but instead applies discrete wavelet transform to the downsampled feature map, decomposing it into components of different frequencies. An adaptive gating system is specifically designed for high-frequency detail components to dynamically distinguish noise from useful edges and texture information. By suppressing noise-corresponding components and enhancing key feature-corresponding components before reconstruction, accurate and adaptive denoising at the feature level is achieved. This method effectively purifies features, reduces misjudgments caused by noise, and avoids feature blurring or loss that may occur with traditional denoising methods, ensuring high quality and high fidelity of features used in subsequent processing.

[0065] 4. Deep feature interaction between CNN and Transformer is achieved through a cross-feature channel attention mechanism, giving full play to the synergistic advantages of the two architectures: In the second feature extraction branch, instead of simply parallelizing or concatenating CNN and Transformer, a deep interaction mechanism is proposed. This mechanism guides the two features to be projected into each other's related subspaces by constructing a cross-feature channel correlation matrix between the feature vectors of CNN and Transformer. This process enables the local detail features extracted by CNN to be integrated into the global relevance information that Transformer focuses on, while also enabling the global context features of Transformer to focus on the local important regions emphasized by CNN. This interaction achieves true complementarity and integration of the advantages of the two mainstream architectures, producing a feature enhancement effect of "1+1>2", enabling the extracted second branch features to have both strong local descriptive ability and global semantic understanding ability.

[0066] 5. By introducing an adaptive feature fusion strategy with multi-scale channel attention and gating mechanisms, dynamic optimization of fusion weights is achieved: In the feature fusion module, the traditional fixed weight addition or splicing method is abandoned, and a more refined adaptive weighted fusion strategy is adopted. This strategy first analyzes the addition results of the first and second branch features from multiple scales, such as local and global, through the multi-scale channel attention module to generate attention weights that can accurately reflect the importance of each channel. Then, these weights are used as gating to perform weighted splicing of the features of the two branches. This mechanism enables the network to dynamically and discriminatively evaluate and fuse information from the two branches according to the specific content of the current input image. For example, in a certain region, the denoised structural features are trusted more, while in another region, the contextual features after interaction are relied upon more, thereby realizing the intelligent and optimized information fusion and greatly improving the representation quality of the fused features.

[0067] 6. Construct a complete, end-to-end automated classification system from data preparation to model deployment, which is highly practical and easy to implement: It not only focuses on the innovation of the core network model structure, but also completes the entire process system including data partitioning, model building, training and testing, and final application; it clarifies the standardized process from historical data preprocessing and partitioning to model training, verification, testing and deployment, forming a closed-loop solution. This end-to-end system design ensures the repeatability and engineering convenience of the technical solution. Users do not need to design additional complex data processing or model integration pipelines, and can directly apply it to actual medical image classification tasks, which significantly reduces the technical application threshold.

[0068] While specific embodiments of the present invention have been described above, those skilled in the art should understand that the specific embodiments described are merely illustrative and not intended to limit the scope of the present invention. Equivalent modifications and variations made by those skilled in the art in accordance with the spirit of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A medical image classification system based on feature interaction and adaptive feature fusion, characterized in that: Includes the following modules: The medical image dataset partitioning module is used to acquire historical medical image datasets and preprocess and partition the medical image datasets to obtain training sets, validation sets, and test sets. The classification network model building module is used to build a classification network model that includes a first feature extraction branch, a second feature extraction branch, and a feature fusion module. The first feature extraction branch is used to perform hierarchical feature extraction and downsampling on the input medical image through multiple concatenated residual blocks, and to perform adaptive denoising processing based on discrete wavelet transform on the image features extracted at each layer to obtain the first branch features; The second feature extraction branch is used to extract second branch features from the input medical image by combining the CNN sub-branch with the Transformer sub-branch of the cross-feature channel attention mechanism; The feature fusion module is used to adaptively weight and fuse the first branch features and the second branch features through a multi-scale channel attention and gating mechanism to obtain fused features, process the fused features to extract deep features for classification, and output classification results based on the deep features. The classification network model training module is used to train, validate, and test the classification network model using the training set, validation set, and test set, respectively, and to deploy the classification network model that passes the test. The medical image classification module is used to acquire medical images to be classified, input the medical images to be classified into a deployed classification network model, and obtain the corresponding classification results.

2. The medical image classification system based on feature interaction and adaptive feature fusion as described in claim 1, characterized in that: The medical image dataset partitioning module is specifically used for: A historical medical image dataset containing a large number of medical images is obtained, and each medical image in the dataset is preprocessed in a standardized manner. The preprocessed historical medical image dataset is then divided into a training set, a validation set, and a test set according to a preset ratio.

3. The medical image classification system based on feature interaction and adaptive feature fusion as described in claim 1, characterized in that: In the classification network model construction module, the adaptive denoising process based on discrete wavelet transform is performed on the image features extracted from each layer to obtain the first branch features, specifically as follows: The image features after downsampling are subjected to discrete wavelet transform to decompose them into low-frequency approximate components and multiple high-frequency detail components. An adaptive gating system is designed for each of the high-frequency detail components to dynamically adjust the weight coefficients of each high-frequency detail component according to the feature content of the image features, thereby achieving noise suppression and key feature enhancement. The weighted low-frequency approximation components and high-frequency detail components are subjected to inverse discrete wavelet transform to reconstruct a denoised feature map. The denoised feature map is residually fused with the original features before downsampling to obtain the first branch feature.

4. The medical image classification system based on feature interaction and adaptive feature fusion as described in claim 1, characterized in that: In the classification network model construction module, the extraction of second-branch features from the input medical image by combining the CNN sub-branch with the cross-feature channel attention mechanism and the Transformer sub-branch specifically involves: The medical image is input into the CNN sub-branch and Transformer sub-branch of the second feature extraction branch to obtain the CNN feature map and Transformer feature map, respectively. Channel compression is performed on the CNN feature map and Transformer feature map to obtain CNN channel feature vectors and Transformer channel feature vectors; The CNN channel feature vectors and Transformer channel feature vectors are excited and compressed to obtain enhanced CNN attention feature vectors and Transformer attention feature vectors; Construct a cross-feature channel correlation matrix based on the CNN attention feature vector and the Transformer attention feature vector; Based on the cross-feature channel correlation matrix, the CNN feature map is projected onto a subspace related to the Transformer feature map, and the Transformer feature map is projected onto a subspace related to the CNN feature map, so as to interact the CNN feature map and the Transformer feature map to obtain the second branch feature.

5. A medical image classification system based on feature interaction and adaptive feature fusion as described in claim 4, characterized in that: Before outputting the second branch features, spatial-level fusion is performed on the interacted CNN feature map and Transformer feature map, specifically as follows: The interactive CNN feature map is subjected to large kernel convolution to enhance spatial dependencies and concatenated with the interactive Transformer feature map; and the interactive Transformer feature map is subjected to small kernel convolution to extract local details and weighted summed with the interactive CNN feature map.

6. The medical image classification system based on feature interaction and adaptive feature fusion as described in claim 1, characterized in that: In the classification network model construction module, the step of adaptively weighting and fusing the first branch features and the second branch features through a multi-scale channel attention and gating mechanism to obtain the fused features specifically involves: Add the first branch feature to the second branch feature to obtain the summed feature; The additive features are input into the multi-scale channel attention module, and multi-scale channel attention weights for the additive features are generated by combining local channel attention and global channel attention. The multi-scale channel attention weights are used as gating weights to weight and concatenate the first branch features and the second branch features to obtain the fused features.

7. A medical image classification system based on feature interaction and adaptive feature fusion as described in claim 1, characterized in that: In the classification network model construction module, the classification network model adopts an encoder-decoder architecture. The first feature extraction branch and the second feature extraction branch constitute the encoder and are connected to the decoder through skip connections.

8. A medical image classification system based on feature interaction and adaptive feature fusion as described in claim 1, characterized in that: In the classification network model construction module, the specific steps of processing the fused features to extract deep features for classification are as follows: The fused features are sequentially upsampled and decoded by at least one decoder to gradually restore the image resolution and output deep features for classification.

9. A medical image classification system based on feature interaction and adaptive feature fusion as described in claim 1, characterized in that: The classification network model training module is specifically used for: The classification network model is trained using the training set, validated and hyperparameters tuned using the validation set, tested using the test set, and then deployed using the tested classification network model.

Images