A method for early diagnosis of alzheimer's disease

Abstract

The application relates to the technical field of medical big data processing and artificial intelligence auxiliary diagnosis, in particular to an early Alzheimer's disease diagnosis method, which comprises the following steps: preprocessing and registering structural magnetic resonance imaging (sMRI) and resting-state functional magnetic resonance imaging (rs-fMRI) image data of a to-be-diagnosed object, constructing a mixed feature pyramid to extract multi-scale anatomical features, adopting a space-time manifold embedding module to extract dynamic functional features, strengthening pathological correlation features through a cross-dimension double attention mechanism, and finally realizing diagnosis classification through multi-modal feature adaptive fusion. The application can accurately capture the deep correlation between brain structure microlesions and functional network abnormalities, and significantly improve the diagnosis accuracy of Alzheimer's disease and early mild cognitive impairment.

Description

Technical Field

[0001] This invention relates to the fields of medical big data processing and artificial intelligence-assisted diagnosis. Background Technology

[0002] Alzheimer's disease is a neurodegenerative disease with insidious onset and progressive development. Intervention in the early stages of the progression from mild cognitive impairment (MCI) to Alzheimer's is crucial to slowing disease progression. Currently, computer-aided diagnosis based on medical imaging (such as structural magnetic resonance imaging (sMRI) and functional magnetic resonance imaging (fMRI) has become a research hotspot for the early diagnosis of Alzheimer's disease.

[0003] However, existing auxiliary diagnostic methods still have many technical problems:

[0004] Loss of features of small lesions: Most existing 3D-CNN models directly perform convolution operations on whole brain images, ignoring the differences in brain atrophy at different anatomical scales (such as local texture of the hippocampus and macroscopic atrophy of the whole brain), resulting in the loss of features of small lesions during downsampling, which affects the sensitivity of early diagnosis.

[0005] Insufficient capture of spatiotemporal coupling: When processing fMRI data, existing methods usually process time series and spatial structure separately (e.g., CNN first and then RNN), which fails to effectively capture the deep coupling between spatial activation and time dependence of brain regions and makes it difficult to reflect the dynamic changes in brain function.

[0006] Implicit feature identification is difficult: Traditional attention mechanisms focus on salient regions, but early pathological changes in Alzheimer's disease often manifest as weak changes in the connection strength of specific brain functional networks rather than obvious structural lesions. Existing algorithms struggle to capture these implicit long-distance dependent features, resulting in low accuracy in early diagnosis. Summary of the Invention

[0007] To overcome the problems of loss of features of small lesions, insufficient capture of spatiotemporal coupling relationships, and difficulty in identifying implicit long-distance dependence features in existing diagnostic methods, this invention provides an early diagnostic method for Alzheimer's disease, comprising the following steps:

[0008] S1. Acquire structural magnetic resonance imaging (sMRI) and resting-state functional magnetic resonance imaging (rs-fMRI) image data of the subject to be diagnosed;

[0009] S2. Input the sMRI image data into a multi-scale panoramic-slice hybrid feature pyramid network to extract features, and then perform feature fusion to generate multi-scale anatomical feature vectors.

[0010] S3. Extract dynamic functional feature vectors from rs-fMRI image data using a spatiotemporal manifold embedding module;

[0011] S4. Perform weighted optimization on multi-scale anatomical feature vectors and dynamic functional feature vectors;

[0012] S5. Concatenate and stitch the optimized feature vectors along the feature dimension, input them into a fully connected layer with DropBlock regularization to achieve fusion, and output the probability distribution of the subject to be diagnosed as the normal control group (NC), early mild cognitive impairment (EMCI), late mild cognitive impairment (LMCI) or Alzheimer's disease (AD) to complete the early diagnosis of Alzheimer's disease.

[0013] Preferably, in step S1, the image data undergoes preprocessing including skull removal, temporal layer correction, and head motion correction, and then spatial registration of rs-fMRI and sMRI is completed using a rigid body transformation algorithm based on mutual information.

[0014] Preferably, in step S2, the extracted features are extracted through three parallel branches: 3D dilated convolution is used to extract macroscopic atrophy features of the whole brain, slices are taken along the sagittal, coronal and axial planes, high-frequency texture features are extracted using 2D residual networks, and different brain regions are segmented based on brain atlases and local geometric features are extracted.

[0015] Preferably, in step S2, feature fusion is performed by hierarchically fusing the three features through a Feature Pyramid Network (FPN), and then the adaptive slice-voxel reprojection algorithm is used for feature fusion, as shown in the formula:

[0016] ;

[0017] in, It is a voxel-level feature. For slice-level feature sets, For ROI map features, For tensor addition based on residual connections, For the Hadamard product of the channel dimension, , These represent the learnable weight matrix and bias vector in the ROI space attention mechanism, respectively. It is the Sigmoid activation function. This is a 3D feature fusion mapping function. For a specific perspective (slices) The 2D feature extraction operation performed The initial 3D whole-brain structural feature tensor extracted from the backbone network; The slice back projection operator is defined by the following formula:

[0018] ;

[0019] in, For interpolation operations, For dimensional expansion operations, It serves as a mask for brain parenchyma.

[0020] Preferably, the extraction of dynamic functional feature vectors specifically involves treating rs-fMRI as a series of 3D tensors, using a sliding window operation of 3D convolution kernels on the time axis to extract local spatiotemporal features and calculate the dynamic functional connectivity matrix between brain regions to generate dynamic functional feature vectors.

[0021] Let the input data sequence be X(t), and the local spatiotemporal feature descriptor be:

[0022] ;

[0023] in, The first-order difference gradient in the time dimension. For smooth convolution in three-dimensional space, The regularization coefficient is . For the space Laplace operator, Gradient operator for spatial dimension.

[0024] Preferably, in the sliding window operation, the sliding window length is 20-50 TR, and the step size is 1-2 TR.

[0025] Preferably, in step S4, the multi-scale anatomical feature vector and the dynamic functional feature vector are input into the cross-dimensional dual attention module, and the optimized multi-scale anatomical feature vector and the dynamic functional feature vector are obtained through dual attention weighting.

[0026] The cross-dimensional dual attention module includes a channel-temporal joint attention unit and a spatial-connectivity cross-covariance attention unit. Dual attention weighting first uses the channel-temporal joint attention unit to weight the channel importance of different feature vectors and the key frames of the time series, and then uses the spatial-connectivity cross-covariance attention unit to calculate the cross-covariance between multi-scale anatomical feature vectors and dynamic functional feature vectors.

[0027] Preferably, the energy function of the spatial-connected cross-covariance attention unit is:

[0028] ;

[0029] in, The query vector mapped from structural features. The key vector mapped from the functional connectivity matrix. This is the pathological prior bias matrix. The dimension of the key vector. For functional feature vectors; The reduced-dimensional reflective function of the manifold is given by the formula:

[0030] ;

[0031] in, For the first A set of voxels from each brain region This is a global average pooling operation. voxels Features For position encoding.

[0032] Preferably, in step S5, a multi-center adversarial constraint loss function is introduced to train the fully connected layer, as shown in the formula:

[0033] ;

[0034] in, For cross-entropy loss, , These are the weighting coefficients. Loss at the center The loss is due to orthogonality constraints; the formula for the orthogonality constraint term is:

[0035] ;

[0036] in, For the first Multi-scale anatomical feature vectors For the first A dynamic functional feature vector, It is the Frobenius norm.

[0037] The beneficial effects of this invention are as follows:

[0038] This invention employs a panoramic-slice hybrid feature pyramid network, combined with an adaptive slice-voxel reprojection algorithm, to simultaneously capture macroscopic atrophy of the whole brain, high-frequency texture of slices, and morphological features of local brain regions, avoiding the loss of features from minute lesions and greatly improving the sensitivity to early, minor lesions. The invention utilizes a spatiotemporal manifold embedding module and a spatiotemporal gradient flow operator to effectively capture the deep coupling relationship between spatial activation and temporal dependence of brain regions, enhancing the representation ability of dynamic functional features. The proposed cross-covariance mechanism can clearly uncover the causal relationship between structural atrophy and functional degradation, effectively solving the problem that multimodal data is merely a simple "stitching" without "interaction." Furthermore, through spatiotemporal manifold embedding, the system has stronger anti-interference capabilities against head movement noise commonly found in fMRI data, reducing the misdiagnosis rate in clinical settings. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the overall process of an embodiment of the present invention. Figure I ;

[0040] Figure 2 This is a schematic diagram of the overall process of an embodiment of the present invention. Figure II . Detailed Implementation

[0041] Embodiments of the present invention provide a method for early diagnosis of Alzheimer's disease, such as... Figure 1 , 2 As shown, it includes the following steps:

[0042] S1. Acquire structural magnetic resonance imaging (sMRI) and resting-state functional magnetic resonance imaging (rs-fMRI) image data of the subject to be diagnosed (data format is DICOM or NIfTI 3D / 4D matrix). Perform preprocessing on the image data, including skull removal, time-layer correction, and head motion correction. Then, use a rigid body transformation algorithm based on mutual information to register the rs-fMRI to the anatomical space of the sMRI to achieve spatial alignment of structure and function.

[0043] S2. Establish a multi-scale panoramic-slice hybrid feature pyramid network and input the registered sMRI image data. Extract features through three parallel branches: use 3D dilated convolution to extract macroscopic atrophy features of the whole brain; use 2D residual network to extract high-frequency texture features after slicing along the sagittal, coronal, and axial planes; and segment specific brain regions such as the hippocampus and entorhinal cortex based on brain atlas and extract local geometric features. Perform hierarchical fusion of the three features through the feature pyramid network (FPN) and combine the adaptive slice-voxel reprojection algorithm to complete feature fusion and generate multi-scale anatomical feature vectors.

[0044] The formula for feature fusion using the adaptive slice-voxel reprojection algorithm is as follows:

[0045] ;

[0046] in, It is a voxel-level feature. It is a slice-level feature set (corresponding to the sagittal, coronal, and axial planes respectively). For ROI map features, For tensor addition based on residual connections, The Hadamard Product is a channel-dimensional feature used to gate and adjust spatial features by leveraging prior knowledge of the ROI. , These are the learnable weight matrix and bias vector in the ROI space attention mechanism, respectively, used to perform linear transformations on prior knowledge; The sigmoid activation function is used to map numerical values ​​to... Intervals are used to generate spatial gating weights; For 3D feature fusion mapping function (usually by Composed of convolutional layers and nonlinear activation functions, it is used to deeply integrate global volumetric features and multi-view slice features. To indicate a specific perspective (Sagittal, coronal, or axial) sections The 2D feature extraction operation performed The initial 3D whole-brain structural feature tensor extracted from the backbone network; The slice back projection operator is defined by the following formula:

[0047] ;

[0048] in, For interpolation operations, For dimensional expansion operations, As a brain parenchyma mask, this operator stretches 2D slice features along the normal direction and utilizes the brain parenchyma mask. Background noise is eliminated, thereby filling the slice texture information back into the 3D spatial coordinate system and compensating for the loss of texture details caused by pure 3D convolution.

[0049] In this step, features are extracted by three parallel branches, and combined with the adaptive slice-voxel reprojection algorithm, lossless fusion of macroscopic and microscopic features is achieved, avoiding the loss of features of small lesions (such as hippocampal micro-atrophy) during downsampling.

[0050] S3. Establish a spatiotemporal manifold embedding module. The preprocessed and registered rs-fMRI image data is used to extract dynamic functional feature vectors through the spatiotemporal manifold embedding module. The extraction of dynamic functional feature vectors is to treat 4D rs-fMRI as a series of 3D tensors, use a sliding window operation of 3D convolution kernels on the time axis to extract local spatiotemporal features and calculate the dynamic functional connectivity matrix (dFC) between brain regions to generate dynamic functional feature vectors.

[0051] In the sliding window operation, the window length is 20-50 TR, and the step size is 1-2 TR. A window that is too short will lead to unstable correlation calculations, while a window that is too long will smooth out short-term dynamic changes and fail to capture transient features. A smaller step size can generate denser time series, which helps to capture continuous dynamic fluctuations in brain function. In this embodiment, the window length and step size are adjusted according to specific experiments.

[0052] The spatiotemporal manifold embedding module uses the spatiotemporal gradient flow operator to extract dynamic features. Let the input rs-fMRI data sequence be X(t), and the local spatiotemporal feature descriptor be:

[0053] ;

[0054] in, The first-order difference gradient in the time dimension is used to capture instantaneous changes in neural activity; For smooth convolution in three-dimensional space; The regularization coefficient is used. This is the spatial Laplacian operator, used to regularize spatial smoothness; The gradient operator is a spatial dimension; this formula highlights anomalous activation points (i.e., early neuronal anomalous firing regions) that fluctuate dramatically over time and are highly localized in space by subtracting the spatial divergence from the magnitude of the temporal gradient.

[0055] In this step, the spatiotemporal manifold embedding module directly extracts local spatiotemporal coupling features through a 3D convolution kernel time axis sliding window operation, avoiding the separation of spatial and temporal information processing, effectively capturing the deep association between brain region activation and time dependence, and solving the problem of insufficient capture of spatiotemporal coupling relationship.

[0056] S4. Input the multi-scale anatomical feature vector and dynamic functional feature vector into the cross-dimensional dual attention module, and obtain the optimized multi-scale anatomical feature vector and dynamic functional feature vector through dual attention weighting.

[0057] The cross-dimensional dual attention module includes a channel-temporal joint attention unit and a spatial-connectivity cross-covariance attention unit. Dual attention weighting first uses the channel-temporal joint attention unit to weight the channel importance of different feature vectors and keyframes of the time series to suppress noise signals. Then, the spatial-connectivity cross-covariance attention unit calculates the cross-covariance between multi-scale anatomical feature vectors and dynamic functional feature vectors, strengthening associated regions with structural shrinkage and significantly reduced functional connectivity.

[0058] The energy function of the spatial-connected cross-covariance attention unit is:

[0059] ;

[0060] in, The query vector mapped from structural features represents the degree of anatomical brain atrophy. The key vector, mapped from the functional connectivity matrix, represents the information transmission efficiency between brain regions; This is the pathological prior bias matrix, used to force the introduction of the default network (DMN) damage pattern prior common to AD in this attention mechanism; The dimension of the key vector. For functional feature vectors; Let be the manifold reduction reflective function, used to map a high-dimensional 3D structural feature tensor to a topological node space with the same functional connectivity matrix, as shown in the formula:

[0061] ;

[0062] in, For the first A set of voxels from each brain region This is a global average pooling operation. voxels Features For position encoding.

[0063] In this step, by calculating the cross-covariance between the structural feature map and the functional connectivity matrix, and combining it with the pathological prior bias of the AD default network (DMN), the latent association region is strengthened, and early non-explicit long-distance dependency anomalies of functional networks are accurately identified, thus solving the problem of latent long-distance dependency feature identification.

[0064] S5. The optimized multi-scale anatomical feature vector and dynamic functional feature vector are concatenated and spliced ​​in the feature dimension. The spliced ​​feature input includes a DropBlock regularized fully connected layer. The learnable parameters of the fully connected layer realize the adaptive weighted fusion of multimodal features. Finally, the probability distribution of the subject to be diagnosed as belonging to the normal control group (NC), early mild cognitive impairment (EMCI), late mild cognitive impairment (LMCI) or Alzheimer's disease (AD) is output. This probability distribution is the target result of early diagnosis of Alzheimer's disease.

[0065] To address the heterogeneity issue of multi-slot data, a multi-center adversarial loss function is introduced to train the fully connected layer, as shown in the formula:

[0066] ;

[0067] in, For cross-entropy loss, , These are the weighting coefficients. Loss at the center The loss is due to orthogonality constraints; the formula for the orthogonality constraint term is:

[0068] ;

[0069] in, For the first Multi-scale anatomical feature vectors For the first A dynamic functional feature vector, The Frobenius norm is used; this constraint requires that structural and functional features remain orthogonal in the feature space, thereby forcing the network to mine complementary pathological information in sMRI and fMRI, rather than repeatedly learning obvious features.

[0070] This invention has been described through embodiments. Those skilled in the art will understand that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of this invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, this invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of this invention.

Claims

1. A method for early diagnosis of Alzheimer's disease, characterized in that, Includes the following steps: S1. Acquire structural magnetic resonance imaging (sMRI) and resting-state functional magnetic resonance imaging (rs-fMRI) image data of the subject to be diagnosed; S2. Input the sMRI image data into a multi-scale panoramic-slice hybrid feature pyramid network to extract features, and then perform feature fusion to generate multi-scale anatomical feature vectors. S3. Extract dynamic functional feature vectors from rs-fMRI image data using a spatiotemporal manifold embedding module; S4. Perform weighted optimization on multi-scale anatomical feature vectors and dynamic functional feature vectors; S5. Concatenate and stitch the optimized feature vectors along the feature dimension, input them into a fully connected layer with DropBlock regularization to achieve fusion, and output the probability distribution of the subject to be diagnosed as the normal control group (NC), early mild cognitive impairment (EMCI), late mild cognitive impairment (LMCI) or Alzheimer's disease (AD) to complete the early diagnosis of Alzheimer's disease.

2. The method for early diagnosis of Alzheimer's disease according to claim 1, characterized in that, In step S1, the image data undergoes preprocessing including skull removal, temporal layer correction, and head motion correction. Then, spatial registration of rs-fMRI and sMRI is completed using a rigid body transformation algorithm based on mutual information.

3. The method for early diagnosis of Alzheimer's disease according to claim 1, characterized in that, In step S2, the extracted features are extracted through three parallel branches: 3D dilated convolution is used to extract macroscopic atrophy features of the whole brain, slices are taken along the sagittal, coronal and axial planes, high-frequency texture features are extracted using 2D residual networks, and different brain regions are segmented based on brain atlases and local geometric morphological features are extracted.

4. The method for early diagnosis of Alzheimer's disease according to claim 3, characterized in that, In step S2, the feature fusion performs hierarchical fusion of the three features through a Feature Pyramid Network (FPN), and then uses an adaptive slice-voxel reprojection algorithm for feature fusion, as shown in the formula: ； in, It is a voxel-level feature. For slice-level feature sets, For ROI map features, For tensor addition based on residual connections, For the Hadamard product of the channel dimension, , These represent the learnable weight matrix and bias vector in the ROI space attention mechanism, respectively. It is the Sigmoid activation function. This is a 3D feature fusion mapping function. For a specific perspective (slices) The 2D feature extraction operation performed The initial 3D whole-brain structural feature tensor extracted from the backbone network; The slice back projection operator is defined by the following formula: ； in, For interpolation operations, For dimensional expansion operations, It serves as a mask for brain parenchyma.

5. The method for early diagnosis of Alzheimer's disease according to claim 1, characterized in that, The extraction of dynamic functional feature vectors involves treating rs-fMRI as a series of 3D tensors, using a sliding window operation of 3D convolution kernels on the time axis to extract local spatiotemporal features and calculate the dynamic functional connectivity matrix between brain regions to generate dynamic functional feature vectors. Let the input data sequence be X(t), and the local spatiotemporal feature descriptor be: ； in, The first-order difference gradient in the time dimension. For smooth convolution in three-dimensional space, The regularization coefficient is . For the space Laplace operator, Gradient operator for spatial dimension.

6. The method for early diagnosis of Alzheimer's disease according to claim 5, characterized in that, In the sliding window operation, the sliding window length is 20-50TR, and the step size is 1-2TR.

7. The method for early diagnosis of Alzheimer's disease according to claim 1, characterized in that, In step S4, the multi-scale anatomical feature vector and the dynamic functional feature vector are input into the cross-dimensional dual attention module, and the optimized multi-scale anatomical feature vector and the dynamic functional feature vector are obtained through dual attention weighting. The cross-dimensional dual attention module includes a channel-temporal joint attention unit and a spatial-connectivity cross-covariance attention unit. Dual attention weighting first uses the channel-temporal joint attention unit to weight the channel importance of different feature vectors and the key frames of the time series, and then uses the spatial-connectivity cross-covariance attention unit to calculate the cross-covariance between multi-scale anatomical feature vectors and dynamic functional feature vectors.

8. The method for early diagnosis of Alzheimer's disease according to claim 7, characterized in that, The energy function of the spatial-connected cross-covariance attention unit is: ； in, The query vector mapped from structural features. The key vector mapped from the functional connectivity matrix. This is the pathological prior bias matrix. The dimension of the key vector. For functional feature vectors; The reduced-dimensional reflective function of the manifold is given by the formula: ； in, For the first A set of voxels from each brain region This is a global average pooling operation. voxels Features For position encoding.

9. The method for early diagnosis of Alzheimer's disease according to claim 1, characterized in that, In step S5, a multi-center adversarial constraint loss function is introduced to train the fully connected layer, as shown in the formula: ； in, For cross-entropy loss, , These are the weighting coefficients. Loss at the center The loss is due to orthogonality constraints; the formula for the orthogonality constraint term is: ； in, For the first Multi-scale anatomical feature vectors For the first A dynamic functional feature vector, It is the Frobenius norm.

Images