Medical image classification method based on frequency domain guidance and channel-space dual-path transformer

Abstract

The application belongs to the technical field of deep learning and medical image intelligent analysis, and particularly relates to an MRI medical image classification method based on frequency domain guidance and channel-space dual path Transformer. The method firstly pre-processes the acquired MRI image, and performs frequency domain decomposition through wavelet transform to generate multi-scale frequency domain enhanced features; then a multi-level spatial semantic feature is extracted by using a convolutional neural network backbone, and a cross-level channel-space fusion module is introduced to interactively enhance the low-level and middle-level features, and a channel-space dual path Transformer module is constructed to process the high-level features, thereby generating global semantic features by modeling spatial long-range dependence and channel correlation in parallel; finally, the multi-level features are fused to output slice-level classification results, and patient-level prediction is realized based on the majority voting strategy. The application effectively fuses local perception and global modeling capability, improves the accuracy, robustness and computational efficiency of MRI image classification, and is suitable for clinical auxiliary diagnosis scenes.

Description

Technical Field

[0001] This invention relates to the fields of deep learning and intelligent medical image analysis technology, specifically to MRI (Magnetic Resonance Imaging) medical image classification technology, and particularly to an MRI image classification method and system based on frequency domain guidance and channel-space dual-path Transformer. This invention is applicable to scenarios such as clinical disease auxiliary diagnosis and quantitative analysis of medical images, and can achieve high-precision intelligent recognition and classification of pathological features in MRI images. Background Technology

[0002] MRI imaging, with its advantages of being radiation-free and having high soft tissue resolution, has become one of the core imaging bases for clinical disease diagnosis (such as brain diseases and tumors). With the deep penetration of artificial intelligence technology into the medical field, deep learning-based MRI image classification methods are gradually becoming a key technology for assisted diagnosis. Its core goal is to automatically extract pathological features from images through algorithms to achieve accurate differentiation between disease and healthy states.

[0003] However, existing MRI image classification techniques still have many shortcomings: First, while traditional convolutional neural networks (CNNs) have strong capabilities for extracting local structural features, their ability to capture long-range spatial dependencies in images is limited, making it difficult to fully represent cross-regional pathological association information. Second, while pure Transformer models can model global dependencies, they have high computational complexity and are insufficient in perceiving local fine-grained features, making them susceptible to interference from redundant information. Third, existing methods often employ single-dimensional feature modeling, such as focusing only on spatial or channel features, ignoring the synergistic effect of channel correlation and spatial dependency, resulting in incomplete feature representation. Fourth, there is insufficient mining of frequency domain information in MRI images, failing to fully utilize the multi-scale pathological features contained in different frequency subbands, and the cross-level feature fusion strategy is not perfect, affecting classification accuracy. Fifth, some methods only focus on slice-level prediction results, without considering the correlation between multiple slices of the same subject, which can easily lead to insufficient reliability of patient-level classification results due to misjudgment of a single slice.

[0004] These issues make it difficult for existing MRI image classification models to balance classification accuracy, robustness, and computational efficiency in complex clinical scenarios, thus limiting their large-scale application in clinical auxiliary diagnosis. Therefore, there is an urgent need for an MRI image classification technology that can synergistically utilize local features and global dependencies, integrate multi-dimensional and multi-scale information, and balance efficiency and accuracy.

[0005] Purpose of the invention

[0006] To address the aforementioned shortcomings of existing MRI image classification techniques, the core objective of this invention is to provide an MRI image classification method and system based on frequency domain guidance and a channel-space dual-path Transformer. Specific objectives include:

[0007] 1) Explore the multi-scale features of MRI images in the frequency domain, and enrich the dimensions of feature representation through frequency domain decomposition and enhancement strategies to improve the ability to capture subtle pathological structures.

[0008] 2) By combining the local perception advantage of convolutional neural networks with the global dependency modeling capability of Transformers, the limitations of single models in feature extraction are overcome.

[0009] 3) Construct a cross-level, channel-space dual-path feature fusion mechanism to strengthen the correlation of features at different levels and in different dimensions, and achieve more comprehensive feature representation;

[0010] 4) Optimize the model structure design to reduce computational complexity while ensuring classification performance, thereby achieving a balance between efficiency and accuracy;

[0011] 5) Establish a multi-level prediction fusion strategy at the slice level and patient level to improve the robustness and clinical applicability of classification results and provide reliable support for clinical auxiliary diagnosis.

[0012] Technical solution

[0013] To achieve the above objectives, this invention provides an MRI image classification method based on frequency domain guidance and channel-space dual-path Transformer, comprising the following steps:

[0014] Step 1: Obtain 3D MRI image data from multiple subjects, perform spatial standardization and slice extraction on the raw MRI data, and construct a 2D MRI slice dataset for model input;

[0015] Step 2: Preprocess the single-channel MRI image by introducing a wavelet-guided multi-scale representation module to perform frequency domain decomposition on the input image and generate multi-scale frequency domain enhancement features;

[0016] Step 3: Construct the backbone module of the convolutional neural network, perform layer-by-layer convolution processing on the frequency domain enhancement features, and extract multi-level spatial semantic features;

[0017] Step 4: Introduce a cross-level channel-spatial feature fusion module into the convolutional backbone to perform channel interaction modeling and spatial attention fusion on low-level and mid-level features to generate cross-level enhanced features;

[0018] Step 5: Construct a channel-space dual-path Transformer module to perform parallel modeling of spatial and channel paths for high-level features, capture long-range spatial dependencies and channel correlations, and generate global semantically enhanced features;

[0019] Step 6: Perform multi-feature fusion on the enhanced features at different levels, construct a classification head, and output slice-level classification prediction results;

[0020] Step 7: Use the cross-entropy loss function to train the model end-to-end and output the optimized MRI image classification model;

[0021] Step 8: The classification prediction results of multiple two-dimensional MRI slices of the same subject are fused, and the final classification prediction result at the subject level is obtained by using a majority voting strategy.

[0022] Furthermore, in step 1, three-dimensional MRI image data of multiple subjects are acquired, and the raw MRI data undergoes spatial standardization, slice extraction, and intensity normalization to construct a two-dimensional MRI slice dataset for model input. The specific steps are as follows:

[0023] Step 1-1: Collect three-dimensional MRI image data from multiple subjects, including a patient group and a healthy control group.

[0024] Steps 1-2: Spatial registration of 3D MRI images is performed using a standard brain template to unify the MRI data of different subjects into the same spatial coordinate system.

[0025] Steps 1-3: Extract multiple consecutive two-dimensional slices along the axial direction from each spatially registered 3D MRI image.

[0026] Steps 1-4: Perform size unification and intensity normalization processing on the two-dimensional MRI slices to form a single-channel two-dimensional MRI input image.

[0027] Furthermore, in step 2, the single-channel MRI image is preprocessed by introducing a wavelet-guided multi-scale representation module to perform frequency domain decomposition on the input image and generate multi-scale frequency domain enhancement features. The specific steps are as follows:

[0028] Step 2-1: Acquire the input single-channel MRI image First, the image is normalized in intensity and its spatial dimensions are adjusted to obtain the preprocessed input image:

[0029]

[0030] in, Represents the original single-channel MRI image; This indicates an intensity normalization operation; This indicates a space size adjustment operation; , These represent the height and width of the target image, respectively. This represents the preprocessed image after normalization and resizing.

[0031] Step 2-2: Apply two-dimensional discrete wavelet transform to the input image The decomposition yields four frequency sub-bands: , , , ,in, This represents the low-frequency subband, used to characterize the global structure and main energy information of an image; , , These represent high-frequency subbands in the horizontal, vertical, and diagonal directions, respectively, used to characterize the edge and texture details of an image.

[0032] Steps 2-3: Perform scale alignment on each frequency sub-band to unify their spatial resolution.

[0033]

[0034] in, , , , This represents the frequency sub-band after scale alignment;

[0035] Steps 2-4: Apply convolution mapping to each frequency sub-band to obtain the corresponding sub-band features:

[0036]

[0037] in, Represents the scale-aligned first... One frequency sub-band; This represents the convolutional features of the corresponding frequency subband; express Convolution operation.

[0038] Steps 2-5: Concatenate the four sub-band features along the channel dimension, and then... Convolutional fusion yields frequency domain enhanced features:

[0039]

[0040] in, This indicates a splicing operation at the channel level; express Convolution operation; This indicates a batch normalization operation; Represents a non-linear activation function; This represents the frequency domain enhancement features after fusion.

[0041] Furthermore, in step 3, a convolutional neural network backbone module is constructed to perform layer-by-layer convolution processing on the frequency domain enhancement features, extracting multi-level spatial semantic features. The specific steps are as follows:

[0042] Step 3-1: Input the frequency domain enhanced features into the backbone of the convolutional neural network and use the ResNet structure for feature extraction;

[0043] Step 3-2: Through multiple residual stages, feature representations at different levels are obtained:

[0044]

[0045] in, It represents low-level features, has high spatial resolution, and is used to characterize local texture and edge information; This represents mid-level features, used to characterize mid-scale spatial structure information; , It represents high-level features, used to characterize high-level semantic information.

[0046] Furthermore, in step 4, a cross-level channel-spatial feature fusion module is introduced into the convolutional backbone to perform channel interaction modeling and spatial attention fusion on low-level and mid-level features, generating cross-level enhanced features. The specific steps are as follows:

[0047] Step 4-1: Low-level features With mid-layer features Apply global average pooling to each channel to obtain the corresponding channel description vectors:

[0048]

[0049] in, This indicates a global average pooling operation; , These represent the channel description vectors for low-level and mid-level features, respectively.

[0050] Step 4-2: Generate the corresponding channel weights through fully connected mapping:

[0051]

[0052] in, , Represents the learnable fully connected weight matrix; This represents the Sigmoid activation function; , This represents the channel weight of the corresponding feature.

[0053] Step 4-3: Use a cross-modulation strategy to perform channel interaction and obtain channel enhancement features:

[0054]

[0055] in, This indicates a channel-by-channel multiplication operation; , This represents the low- and mid-layer features after cross-modulation.

[0056] Step 4-4: To further highlight the consistent salient regions of the two layers of features in the spatial dimension, spatial compression is performed on the enhanced features of each channel, and a joint spatial attention map is constructed:

[0057]

[0058] in: , These represent average pooling and max pooling along the channel dimension, respectively; Indicates feature concatenation operation; The kernel size is Two-dimensional convolution; This is a spatial attention weight map.

[0059] Steps 4-5: Apply the spatial attention map to both low-level and mid-level features simultaneously to obtain the final spatial-channel joint enhancement result:

[0060]

[0061] middle: , These are the output features; the output feature size is the same as the original. , Maintain consistency for subsequent high-level feature extraction and classification.

[0062] Furthermore, in step 5, a channel-space dual-path Transformer module is constructed to perform parallel modeling of spatial and channel paths for high-level features, capturing long-range spatial dependencies and channel correlations to generate global semantically enhanced features. The specific steps are as follows:

[0063] Step 5-1: Process the high-level features obtained in Step 3 and Perform flattening and dimensional rearrangement respectively to construct the corresponding token sequence:

[0064]

[0065] in, These are the high-level semantic features output by the residual stage of the convolutional neural network. Indicates by the first A sequence of tokens obtained from high-level feature mappings; These represent the spatial size and number of channels of the feature map, respectively.

[0066] Step 5-2: For each token sequence Construct spatial paths and model global dependencies between different spatial locations using a multi-head self-attention mechanism:

[0067]

[0068]

[0069] in, It is a learnable linear mapping matrix; the multi-head self-attention mechanism computes the correlation between spatial locations in parallel through multiple attention heads; This represents the spatial enhancement features of the spatial path output.

[0070] Step 5-3: For each token sequence Construct the channel paths, perform global spatial aggregation on them, and model the dependencies between channels:

[0071]

[0072]

[0073] in, This represents global average pooling along the token (space) dimension; The channel enhancement features represent the output of the channel path; the Transformer encoder performs self-attention and feedforward operations in the channel dimension.

[0074] Step 5-4: Broadcast the channel path features to the token dimension, and use residual fusion to fuse the spatial path and channel path features:

[0075]

[0076]

[0077] in, This indicates a dimension expansion operation; These are learnable fusion weight coefficients; This represents the token sequence after fusion via a channel-space dual-path Transformer.

[0078] Step 5-5: Reverse map the fused token sequence into a two-dimensional feature map to obtain the output features of the dual-path Transformer:

[0079]

[0080] in, These represent the high-level features enhanced by the dual-path Transformer.

[0081] Furthermore, in step 6, the enhanced features at different levels are fused to construct a classification head and output slice-level classification prediction results. The specific steps are as follows:

[0082] Step 6-1: Enhance the features at different levels obtained in Step 4 and Step 5 Apply adaptive global average pooling to obtain the corresponding global feature vectors:

[0083]

[0084] in, This represents a global average pooling operation performed on the spatial dimension; Indicates the first Global semantic representation of layer-enhanced features.

[0085] Step 6-2: Concatenate the global feature vectors of each level along the channel dimension to construct a multi-layer feature fusion representation:

[0086]

[0087] in, This indicates a feature concatenation operation along the channel dimension; This represents a global feature vector that integrates low-level detailed information with high-level semantic information.

[0088] Step 6-3: Input the fused feature vector into the classification head, and obtain the final classification prediction result through fully connected mapping, normalization, nonlinear activation and regularization operations.

[0089] in, Indicates a fully connected layer; This indicates a batch normalization operation; Represents a non-linear activation function; This represents a random deactivation regularization operation; Represents the classification prediction probability distribution of slice-level MRI images

[0090] Furthermore, in step 7, the model is trained end-to-end using the cross-entropy loss function to output an optimized MRI image classification model. The specific steps are as follows:

[0091] The cross-entropy loss function is used to supervise the training of the predicted results and the true labels:

[0092] in, Indicates the number of training samples; Indicates the first The true category label corresponding to each sample; The model represents the first The probability value of a sample being predicted as the true class; This represents the overall loss function value of the model during the training phase.

[0093] The model parameters are updated through backpropagation and gradient descent until the model converges, resulting in the final MRI image classification model.

[0094] Furthermore, in step 8, the classification prediction results of multiple two-dimensional MRI slices from the same subject are fused, and a majority voting strategy is used to obtain the final classification prediction result at the subject level. The specific steps are as follows:

[0095] Step 8-1: For each two-dimensional MRI slice, output the corresponding slice-level predicted category based on the classification head.

[0096] Step 8-2: Statistically analyze the prediction results of multiple slices of the same subject.

[0097] Step 8-3: Using a majority voting strategy, the category with the highest frequency is taken as the final prediction result for the subject.

[0098] The patient-level prediction results are expressed as follows:

[0099]

[0100] in, Indicates the first Predicted category of 2D MRI slices This indicates the total number of slices for that subject.

[0101] Beneficial effects:

[0102] This invention, through its innovative technical solution, achieves the following significant advantages compared to existing technologies:

[0103] 1) The MRI images are decomposed in the frequency domain by wavelet-guided multi-scale representation module to extract the global structural information of the low-frequency subband and the detailed pathological features of the high-frequency subband. Combined with the cross-level feature fusion strategy, the full utilization of multi-scale and cross-level features is realized, which greatly improves the richness and effectiveness of feature representation.

[0104] 2) The channel-space dual-path Transformer module models long-range spatial dependencies and channel correlations in parallel. This retains the ability of CNN to perceive local fine-grained features and leverages the advantages of Transformer in capturing global correlation information. It solves the shortcomings of single-dimensional modeling and makes feature representation have both local accuracy and global integrity.

[0105] 3) Through cross-modulated channel interaction, spatial attention weighting, multi-feature fusion and patient-level majority voting strategies, the impact of single feature misjudgment and single slice bias is effectively reduced, significantly improving the accuracy and robustness of MRI image classification, especially suitable for classification tasks in complex pathological scenarios.

[0106] 4) The model adopts an end-to-end training method, which is convenient to operate. The patient-level prediction results are highly consistent with the clinical diagnostic logic, which can provide doctors with intuitive and reliable auxiliary diagnostic basis. It is applicable to a variety of MRI-related disease diagnosis scenarios (such as brain diseases, tumor screening, etc.) and has broad clinical application value. Attached Figure Description

[0107] Figure 1 This is a schematic diagram of the steps of the present invention.

[0108] Figure 2 This is the overall model framework diagram of the present invention.

[0109] Figure 3 This is a flowchart of the preprocessing and frequency domain decomposition enhancement feature generation process of the present invention.

[0110] Figure 4 This is a flowchart of the cross-level channel-spatial feature fusion module of the present invention.

[0111] Figure 5 This is a flowchart of the channel-space dual-path Transformer module processing of the present invention. Detailed Implementation

[0112] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings. An MRI image classification method based on frequency domain guidance and channel-space dual-path Transformer includes the following steps:

[0113] The overall process is as follows Figure 2 As shown,

[0114] Step 1: The flowchart for this step is as follows Figure 3 As shown, three-dimensional MRI image data of multiple subjects were acquired, and the original MRI data were spatially standardized and sliced ​​to construct a two-dimensional MRI slice dataset for model input.

[0115] Step 1-1: Collect three-dimensional MRI image data from multiple subjects, including a patient group and a healthy control group.

[0116] Steps 1-2: Spatial registration of 3D MRI images is performed using a standard brain template to unify the MRI data of different subjects into the same spatial coordinate system.

[0117] Steps 1-3: Extract multiple consecutive two-dimensional slices along the axial direction from each spatially registered 3D MRI image.

[0118] Steps 1-4: Perform size unification and intensity normalization processing on the two-dimensional MRI slices to form a single-channel two-dimensional MRI input image.

[0119] Step 2: The flowchart for this step is as follows. Figure 3 As shown, single-channel MRI images are preprocessed by introducing a wavelet-guided multi-scale representation module to perform frequency domain decomposition on the input image and generate multi-scale frequency domain enhancement features.

[0120] Step 2-1: Acquire the input single-channel MRI image First, the image is normalized in intensity and its spatial dimensions are adjusted to obtain the preprocessed input image:

[0121]

[0122] in, Represents the original single-channel MRI image; This indicates an intensity normalization operation; This indicates a space size adjustment operation; , These represent the height and width of the target image, respectively. This represents the preprocessed image after normalization and resizing.

[0123] Step 2-2: Apply two-dimensional discrete wavelet transform to the input image The decomposition yields four frequency sub-bands: , , , ,in, This represents the low-frequency subband, used to characterize the global structure and main energy information of an image; , , These represent high-frequency subbands in the horizontal, vertical, and diagonal directions, respectively, used to characterize the edge and texture details of an image.

[0124] Steps 2-3: Perform scale alignment on each frequency sub-band to unify their spatial resolution.

[0125]

[0126] in, , , , This represents the frequency sub-band after scale alignment;

[0127] Steps 2-4: Apply convolution mapping to each frequency sub-band to obtain the corresponding sub-band features:

[0128]

[0129] in, Represents the scale-aligned first... One frequency sub-band; This represents the convolutional features of the corresponding frequency subband; express Convolution operation.

[0130] Steps 2-5: Concatenate the four sub-band features along the channel dimension, and then... Convolutional fusion yields frequency domain enhanced features:

[0131]

[0132] in, This indicates a splicing operation at the channel level; express Convolution operation; This indicates a batch normalization operation; Represents a non-linear activation function; This represents the frequency domain enhancement features after fusion.

[0133] Step 3: Construct the backbone module of the convolutional neural network, perform layer-by-layer convolution processing on the frequency domain enhancement features, and extract multi-level spatial semantic features;

[0134] Step 3-1: Input the frequency domain enhanced features into the backbone of the convolutional neural network and use the ResNet structure for feature extraction;

[0135] Step 3-2: Through multiple residual stages, feature representations at different levels are obtained:

[0136]

[0137] in, It represents low-level features, has high spatial resolution, and is used to characterize local texture and edge information; This represents mid-level features, used to characterize mid-scale spatial structure information; , It represents high-level features, used to characterize high-level semantic information.

[0138] Step 4: The flowchart for this step is as follows. Figure 4 As shown, a cross-level channel-spatial feature fusion module is introduced into the convolutional backbone to perform channel interaction modeling and spatial attention fusion on low-level and mid-level features, generating cross-level enhanced features.

[0139] Step 4-1: Low-level features With mid-layer features Apply global average pooling to each channel to obtain the corresponding channel description vectors:

[0140]

[0141] in, This indicates a global average pooling operation; , These represent the channel description vectors for low-level and mid-level features, respectively.

[0142] Step 4-2: Generate the corresponding channel weights through fully connected mapping:

[0143]

[0144] in, , Represents the learnable fully connected weight matrix; This represents the Sigmoid activation function; , This represents the channel weight of the corresponding feature.

[0145] Step 4-3: Use a cross-modulation strategy to perform channel interaction and obtain channel enhancement features:

[0146]

[0147] in, This indicates a channel-by-channel multiplication operation; , This represents the low- and mid-layer features after cross-modulation.

[0148] Step 4-4: To further highlight the consistent salient regions of the two layers of features in the spatial dimension, spatial compression is performed on the enhanced features of each channel, and a joint spatial attention map is constructed:

[0149]

[0150] in: , These represent average pooling and max pooling along the channel dimension, respectively; Indicates feature concatenation operation; The kernel size is Two-dimensional convolution; This is a spatial attention weight map.

[0151] Steps 4-5: Apply the spatial attention map to both low-level and mid-level features simultaneously to obtain the final spatial-channel joint enhancement result:

[0152]

[0153] in: , These are the output features; the output feature size is the same as the original. , Maintain consistency for subsequent high-level feature extraction and classification.

[0154] Step 5: The flowchart for this step is as follows. Figure 5 As shown, a channel-space dual-path Transformer module is constructed to perform parallel modeling of spatial and channel paths for high-level features, capture long-range spatial dependencies and channel correlations, and generate global semantically enhanced features.

[0155] Step 5-1: Process the high-level features obtained in Step 3 and Perform flattening and dimensional rearrangement respectively to construct the corresponding token sequence:

[0156]

[0157] in, These are the high-level semantic features output by the residual stage of the convolutional neural network. Indicates by the first A sequence of tokens obtained from high-level feature mappings; These represent the spatial size and number of channels of the feature map, respectively.

[0158] Step 5-2: For each token sequence Construct spatial paths and model global dependencies between different spatial locations using a multi-head self-attention mechanism:

[0159]

[0160]

[0161] in, It is a learnable linear mapping matrix; the multi-head self-attention mechanism computes the correlation between spatial locations in parallel through multiple attention heads; This represents the spatial enhancement features of the spatial path output.

[0162] Step 5-3: For each token sequence Construct the channel paths, perform global spatial aggregation on them, and model the dependencies between channels:

[0163]

[0164]

[0165] in, This represents global average pooling along the token (space) dimension; The channel enhancement features represent the output of the channel path; the Transformer encoder performs self-attention and feedforward operations in the channel dimension.

[0166] Step 5-4: Broadcast the channel path features to the token dimension, and use residual fusion to fuse the spatial path and channel path features:

[0167]

[0168]

[0169] in, This indicates a dimension expansion operation; These are learnable fusion weight coefficients; This represents the token sequence after fusion via a channel-space dual-path Transformer.

[0170] Step 5-5: Reverse map the fused token sequence into a two-dimensional feature map to obtain the output features of the dual-path Transformer:

[0171]

[0172] in, These represent the high-level features enhanced by the dual-path Transformer.

[0173] Step 6: Perform multi-feature fusion on the enhanced features at different levels, construct a classification head, and output slice-level classification prediction results;

[0174] Step 6-1: Enhance the features at different levels obtained in Step 4 and Step 5 Apply adaptive global average pooling to obtain the corresponding global feature vectors:

[0175]

[0176] in, This represents a global average pooling operation performed on the spatial dimension; Indicates the first Global semantic representation of layer-enhanced features.

[0177] Step 6-2: Concatenate the global feature vectors of each level along the channel dimension to construct a multi-layer feature fusion representation:

[0178]

[0179] in, This indicates a feature concatenation operation along the channel dimension; This represents a global feature vector that integrates low-level detailed information with high-level semantic information.

[0180] Step 6-3: Input the fused feature vector into the classification head, and obtain the final classification prediction result through fully connected mapping, normalization, nonlinear activation and regularization operations.

[0181] in, Indicates a fully connected layer; This indicates a batch normalization operation; Represents a non-linear activation function; This represents a random deactivation regularization operation; This represents the probability distribution for classification prediction of slice-level MRI images.

[0182] Step 7: Use the cross-entropy loss function to train the model end-to-end and output the optimized MRI image classification model;

[0183] The cross-entropy loss function is used to supervise the training of the predicted results and the true labels:

[0184]

[0185] in, Indicates the number of training samples; Indicates the first The true category label corresponding to each sample; The model represents the first The probability value of a sample being predicted as the true class; This represents the overall loss function value of the model during the training phase.

[0186] The model parameters are updated through backpropagation and gradient descent until the model converges, resulting in the final MRI image classification model.

[0187] Step 8: The classification prediction results of multiple two-dimensional MRI slices of the same subject are fused, and the final classification prediction result at the subject level is obtained by using a majority voting strategy.

[0188] Step 8-1: For each two-dimensional MRI slice, output the corresponding slice-level predicted category based on the classification head.

[0189] Step 8-2: Statistically analyze the prediction results of multiple slices of the same subject.

[0190] Step 8-3: Using a majority voting strategy, the category with the highest frequency is taken as the final prediction result for the subject.

[0191] The patient-level prediction results are expressed as follows:

[0192]

[0193] in, Indicates the first Predicted category of 2D MRI slices This indicates the total number of slices for that subject.

Claims

1. An MRI image classification method based on frequency domain guidance and channel-space dual-path Transformer, characterized in that, The method includes the following steps: Step 1: Obtain 3D MRI image data from multiple subjects, perform spatial standardization and slice extraction on the raw MRI data, and construct a 2D MRI slice dataset for model input; Step 2: Preprocess the single-channel MRI image by performing frequency domain decomposition on the input image using a wavelet-guided multi-scale representation module and generating multi-scale frequency domain enhancement features. ; Step 3: Construct the backbone module of the convolutional neural network and enhance the frequency domain features generated in Step 2. The main module is input for layer-by-layer convolution processing to extract multi-level spatial semantic features, including low-level features. Mid-layer characteristics and high-level characteristics , ; Step 4: Introduce a cross-level channel-spatial feature fusion module into the convolutional backbone to process the low-level features obtained in Step 3. With mid-layer features Channel interaction modeling and spatial attention fusion are performed to generate cross-level enhanced features. , ; Step 5: Construct a channel-space dual-path Transformer module to process the high-level features obtained in Step 3. , Parallel modeling of spatial paths and channel paths is performed to capture long-range spatial dependencies and channel correlations, generating global semantically enhanced features. , ; Step 6: Apply the cross-level enhancement features obtained in Step 4 , Compared with the global semantic enhancement features obtained in step 5 , Perform multi-feature fusion to construct a classification head and output slice-level classification prediction results. ; Step 7: Use the cross-entropy loss function to train the model end-to-end and output the optimized MRI image classification model; Step 8: Fuse the classification prediction results of multiple 2D MRI slices of the same subject, and use a majority voting strategy to obtain the final classification prediction result at the subject level. .

2. The MRI image classification method based on frequency domain guidance and channel-space dual-path Transformer according to claim 1, characterized in that, Step 1 includes the following sub-steps: Step 1-1: Collect three-dimensional MRI image data from multiple subjects, including a patient group and a healthy control group; Steps 1-2: Spatial registration of 3D MRI images is performed using a standard brain template to unify the MRI data of different subjects into the same spatial coordinate system; Steps 1-3: Extract multiple consecutive two-dimensional slices along the axial direction from each spatially registered 3D MRI image.

3. The MRI image classification method based on frequency domain guidance and channel-space dual-path Transformer according to claim 1, characterized in that, Step 2 includes the following sub-steps: Step 2-1: Acquire the input single-channel MRI image First, the image is normalized in intensity and its spatial dimensions are adjusted to obtain the preprocessed input image: ； in, Represents the original single-channel MRI image; This indicates an intensity normalization operation; This indicates a space size adjustment operation; , These represent the height and width of the target image, respectively. This represents the preprocessed image after normalization and resizing. Step 2-2: Apply two-dimensional discrete wavelet transform to the input image The decomposition yields four frequency sub-bands: , , , ,in, This represents the low-frequency subband, used to characterize the global structure and main energy information of an image; , , These represent high-frequency subbands in the horizontal, vertical, and diagonal directions, respectively, used to characterize the edge and texture details of an image; Steps 2-3: Perform scale alignment on each frequency sub-band to unify their spatial resolution. The aligned subbands are obtained: ； Steps 2-4: Apply convolution mapping to each frequency sub-band to obtain the corresponding sub-band features: ； in, Represents the scale-aligned first... One frequency sub-band; This represents the convolutional features of the corresponding frequency subband; express Convolution operation; Steps 2-5: Concatenate the four sub-band features along the channel dimension, and then... Convolutional fusion yields frequency domain enhanced features: ； in, This indicates a concatenation operation along the channel dimension; express Convolution operation; This indicates a batch normalization operation; Represents a non-linear activation function; This indicates the frequency domain enhancement features after fusion.

4. The MRI image classification method based on frequency domain guidance and channel-space dual-path Transformer according to claim 3, characterized in that, Step 3 includes the following sub-steps: Step 3-1: Enhance frequency domain features Input the convolutional neural network backbone and use a ResNet structure for feature extraction; Step 3-2: Pass through multiple residual stages sequentially to obtain feature representations at different levels: ； in, It represents low-level features, has high spatial resolution, and is used to characterize local texture and edge information; This represents mid-level features, used to characterize mid-scale spatial structure information; , It represents high-level features, used to characterize high-level semantic information.

5. The MRI image classification method based on frequency domain guidance and channel-space dual-path Transformer according to claim 4, characterized in that, Step 4 includes the following sub-steps: Step 4-1: Low-level features With mid-layer features Apply global average pooling to each channel to obtain the corresponding channel description vectors: ； in, This indicates a global average pooling operation; , These represent the channel description vectors for low-level and mid-level features, respectively. Step 4-2: Generate the corresponding channel weights through fully connected mapping: ； in, , Represents the learnable fully connected weight matrix; This represents the Sigmoid activation function; , This represents the channel weights corresponding to the features; Step 4-3: Use a cross-modulation strategy to perform channel interaction and obtain channel enhancement features: ； in, This indicates a channel-by-channel multiplication operation; , This represents the low- and mid-layer features after cross-modulation; Step 4-4: To further highlight the consistent salient regions of the two layers of features in the spatial dimension, spatial compression is performed on the enhanced features of each channel, and a joint spatial attention map is constructed: ； in, , These represent average pooling and max pooling along the channel dimension, respectively; Indicates feature concatenation operation; The kernel size is Two-dimensional convolution; This is a spatial attention weight map. , Features and Spatial height and width; Steps 4-5: Apply the spatial attention map to both low-level and mid-level features simultaneously to obtain the final spatial-channel joint enhancement result: ； in, , These are the output features, and their sizes are the same as the original. , Maintain consistency.

6. The MRI image classification method based on frequency domain guidance and channel-space dual-path Transformer according to claim 5, characterized in that, Step 5 includes the following sub-steps: Step 5-1: Process the high-level features obtained in Step 3 and Perform flattening and dimensional rearrangement respectively to construct the corresponding token sequence: ； in, , These are the high-level semantic features output by the residual stage of the convolutional neural network. Indicates by the first A sequence of tokens obtained from high-level feature mappings; , , These represent the spatial dimensions and number of channels of the feature map, respectively. Step 5-2: For each token sequence Construct spatial paths and model global dependencies between different spatial locations using a multi-head self-attention mechanism: ； in, , , It is a learnable linear mapping matrix; the multi-head self-attention mechanism computes the correlation between spatial locations in parallel through multiple attention heads; The spatial augmentation features represent the output of the spatial path. The number of tokens; Step 5-3: For each token sequence Construct channel paths, globally aggregate their spatial dimensions, and model inter-channel dependencies using a channel Transformer encoder: ； in, This represents global average pooling along the token (space) dimension; This represents a Transformer encoder that performs self-attention and feedforward operations along the channel dimension. This represents the channel enhancement feature output by the channel path; It is a non-linear activation function; Step 5-4: Broadcast the channel path features to the token dimension, and use residual fusion to fuse the spatial path and channel path features: ； in, This indicates a dimension expansion operation, which copies the channel features to each token location; , These are learnable fusion weight coefficients; This represents the token sequence after fusion via a channel-space dual-path Transformer. Step 5-5: Reverse map the fused token sequence into a two-dimensional feature map to obtain the output features of the dual-path Transformer: ； in, , These represent the high-level features enhanced by the dual-path Transformer.

7. The MRI image classification method based on frequency domain guidance and channel-space dual-path Transformer according to claim 6, characterized in that, Step 6 includes the following sub-steps: Step 6-1: Enhance the cross-level features obtained in Step 4 , Compared with the global semantic enhancement features obtained in step 5 , Apply adaptive global average pooling to obtain the corresponding global feature vectors: ； in, This represents a global average pooling operation performed on the spatial dimension; Indicates the first Global semantic representation of layer-enhanced features; Step 6-2: Concatenate the global feature vectors of each level along the channel dimension to construct a multi-layer feature fusion representation: ； in, This indicates a feature concatenation operation along the channel dimension; This represents a global feature vector that integrates low-level detailed information with high-level semantic information. Step 6-3: Input the fused feature vector into the classification head, and obtain the final classification prediction result through fully connected mapping, batch normalization, non-linear activation, random deactivation regularization, and Softmax operation. ； in, Indicates a fully connected layer; This indicates a batch normalization operation; Represents a nonlinear activation function; This represents a random deactivation regularization operation; This represents the probability distribution for classification prediction of slice-level MRI images.

8. The MRI image classification method based on frequency domain guidance and channel-space dual-path Transformer according to claim 7, characterized in that, Step 7 includes the following sub-steps: The cross-entropy loss function is used to supervise the training of the predicted results and the true labels: ； in, Indicates the number of training samples; Indicates the first The true category label corresponding to each sample; The model represents the first The probability value of a sample being predicted as the true class; This represents the overall loss function value of the model during the training phase; The model parameters are updated through backpropagation and gradient descent until the model converges, resulting in the final MRI image classification model.

9. The MRI image classification method based on frequency domain guidance and channel-space dual-path Transformer according to claim 1, characterized in that, Step 8 includes the following sub-steps: Step 8-1: For each 2D MRI slice, output the corresponding slice-level predicted category based on the classification head; Step 8-2: Statistically analyze the prediction results of multiple slices from the same subject; Step 8-3: Using a majority voting strategy, the category with the highest frequency is taken as the final prediction result for the subject; The patient-level prediction results are expressed as follows: ； in, Indicates the first Predicted category of 2D MRI slices This indicates the total number of slices for that subject.

Images