Method and system for analyzing ad characteristic information based on brain medical image

Abstract

The application discloses an AD characteristic information analysis method based on brain medical images, which comprises the following steps: data preprocessing is performed on brain MRI images to be analyzed to obtain standard images of brain regions; spatial characteristic analysis of the whole brain and key brain regions; whole brain image characteristic analysis; cognitive characteristic analysis of the whole brain and key brain regions; and aggregation analysis results are obtained to form corresponding explanation data. The application further discloses an AD characteristic information analysis system based on brain medical images. The method for analyzing cognitive index characteristics such as brain age, logical memory score, visual memory score and long-time delay memory score from brain MRI images by using deep learning technology can provide analysis results of brain MRI images and other important medical indexes for doctors, and can provide effective arguments and explanations for subsequent scientific research analysis and interpretation.

Description

Technical Field

[0001] This invention relates to a medical image processing technology, and more particularly to a method and system for analyzing AD feature information based on brain medical images. Background Technology

[0002] Alzheimer's disease (AD) is an irreversible and often difficult-to-detect neurodegenerative disease. AD primarily affects people over 70 years of age, and its incidence increases with age. The main clinical symptoms include memory decline, cognitive impairment, and behavioral abnormalities, severely impacting the daily lives of the elderly and increasing the care costs for families. With the increasing aging of society, tens of thousands of elderly people worldwide will face the risk of developing AD.

[0003] Mild cognitive impairment (MCI) is an early stage of Alzheimer's disease (AD) and a stage where cognitive function undergoes significant changes. Drug intervention during the MCI stage can improve symptoms and alleviate the condition. Currently, diagnosis of MCI and AD primarily relies on physicians' comprehensive analysis and judgment of clinical data. While these methods have achieved good diagnostic results, they are time-consuming, labor-intensive, and somewhat subjective. Since brain atrophy is common in both MCI and AD patients, magnetic resonance imaging (MRI) can help physicians track changes in brain structure and irreversible brain atrophy, making it an effective method for detecting AD. However, brain structural atrophy due to normal aging can also interfere with diagnosis. Cognitive indicators of the whole brain and key brain regions, such as brain age in the hippocampus, logical memory scores, long-term delayed memory scores, and visual memory scores, can effectively reflect a person's cognitive abilities, significantly improve the accuracy of AD detection, and serve as a basis for clinical interpretation of imaging diagnoses. Typically, cognitive indicators in patients with MCI and AD differ from those in normal individuals. However, MRI cannot directly reflect cognitive indicators in key brain regions, posing a challenge to the analysis of images combined with these indicators.

[0004] In recent years, deep learning technology has developed rapidly and been widely applied to various medical image analyses. It typically uses convolutional neural networks trained on brain medical images and clinical evaluation data with doctor-annotated information for prediction. However, these methods all have limitations. For example, Chinese patents CN113538333B and CN112353381A use deep learning convolutional networks to analyze brain MRI images and obtain classification results for the entire image, but they do not consider incorporating clinical cognitive indicators during the prediction process and do not provide clinical interpretation of the image analysis results. Similarly, Chinese patents CN114983341A and CN111488914A, while incorporating clinical information into the analysis of MRI images, do not provide reasonable analysis of key brain regions and similarly lack clinical interpretation of the results, failing to provide convincing auxiliary explanations for doctors' clinical diagnoses and scientific research analyses. Summary of the Invention

[0005] The purpose of this invention is to address the problems existing in the prior art and provide an AD feature information analysis method based on brain medical imaging. This method can analyze the spatial, imaging, and cognitive indicators of the whole brain and key brain regions, providing effective data support and explanation of the analysis process for subsequent clinical research analysis. It can effectively solve the problem that existing technologies lack explanation of the reasoning process and effectively assist doctors in conducting subsequent research.

[0006] The objective of this invention is achieved through the following technical solution:

[0007] On the one hand, an AD feature information analysis method based on brain medical imaging includes the following steps:

[0008] S1. Perform data preprocessing on the brain MRI images to be analyzed to obtain standard images of each brain region;

[0009] S2. Perform spatial feature analysis of the whole brain and key brain regions;

[0010] S3. Perform whole-brain imaging feature analysis;

[0011] S4. Conduct cognitive feature analysis of the whole brain and key brain regions;

[0012] S5. Aggregate the above analysis results to form corresponding explanatory data.

[0013] In step S1, the data preprocessing method includes the following steps:

[0014] S11. Obtain the raw MRI images to be analyzed;

[0015] S12. Perform bias field correction on the MRI image;

[0016] S13. Register the corrected MRI images with the MNI152 standard brain template and remove non-brain regions;

[0017] S14. Perform whole-brain voxel normalization, the normalization formula is:

[0018]

[0019] In the formula, These are the normalized voxels; x is the actual input voxel. xstd is the average value of the voxels; xstd is the standard deviation of the voxels.

[0020] S15. Obtain the standard brain image after standard brain registration and normalization calculation;

[0021] S16. Perform AAL brain region template registration on the standard whole-brain images to obtain standard images of each brain region;

[0022] S17. Output the standard images of the whole brain and each brain region obtained in S15 and S16.

[0023] In step S2, the spatial feature analysis method includes the following steps:

[0024] S21. Input standard whole-brain imaging;

[0025] S22. Calculate the volume of cerebrospinal fluid, gray matter, white matter, and total intracranial volume;

[0026] S23. Input standard images of each brain region;

[0027] S24. Extract key brain regions and calculate the absolute volume of each key brain region. The key brain regions include the left hippocampus, right hippocampus, left parahippocampal gyrus, right parahippocampal gyrus, left cingulate gyrus, and right cingulate gyrus, for a total of six brain regions.

[0028] S25. Calculate the relative spatial indicators of the whole brain and key brain regions, including the ratio of the total intracranial volume of the input whole brain to that of a normal person, the ratio of the volume of each key brain region to that of the same brain region in a normal person, and the ratio of the volume of each key brain region to that of the input whole brain.

[0029] S26. Integration to form spatial features of the whole brain and key brain regions: The total intracranial volume of the whole brain, the ratio of the total intracranial volume of the input whole brain to that of a normal person, the absolute volume of the six key brain regions, the ratio of the volume of each key brain region to that of the same brain region in a normal person, and the ratio of the volume of each key brain region to the volume of the input whole brain are connected in series to form 20-dimensional spatial features of the whole brain and key brain regions.

[0030] In step S3, the method for whole-brain image feature analysis includes the following steps:

[0031] S31. Input standard whole-brain imaging;

[0032] S32. Perform radiomics feature extraction. Based on the 24 features of gray-level co-occurrence matrix, 14 features of gray-level dependence matrix, 16 features of gray-level run matrix, and 16 features of gray-level region size matrix, perform principal component analysis and dimensionality reduction to obtain 30-dimensional final effective radiomics features.

[0033] S33. Divide the whole brain region into multiple patches. Preferably, the whole brain can be divided into 64 patches according to a 4*4*4 space.

[0034] S34. Construct a neural network based on 3D spatial convolution, train the network using multi-instance learning technology, and obtain an AD risk probability prediction model;

[0035] S35. Use the AD risk probability prediction model to predict the risk value of each patch, and fuse and splice them to obtain a 64-dimensional whole-brain AD risk probability feature;

[0036] S36. The radiomics features in step S32 and the whole-brain AD risk probability features in step S35 are fused and stitched together to obtain 94-dimensional whole-brain imaging features.

[0037] In step S34, the AD risk probability prediction model consists of model data input, a backbone feature extraction network, a head prediction branch network, and a final classification function cascade. The model data input uses a single-channel black and white image with a size of 1*16*16*16. All patches are resized to this size before being fed into the network. The backbone feature extraction network uses 3D convolutional blocks with variable kernels as feature extraction unit modules, which includes four cascaded 3D convolutional layers, one 3D pooling layer, one batch normalization layer, and one non-linear activation unit. The backbone feature extraction network contains three 3D convolutional block modules. The head prediction branch network includes one 3D convolutional block module and two fully connected layers, which are used to aggregate and further compress the features output by the backbone to output the risk features of the current image converting to AD. Finally, the Sigmoid function is used as the activation function to convert the output of the head prediction branch into a risk probability between 0 and 1. In the model training process, the model is trained using a multi-instance learning approach, which gives the algorithm the ability to find the optimal solution in multiple patches and finally obtains the trained AD risk probability prediction model.

[0038] In step S4, the method for analyzing the cognitive features of the whole brain and key brain regions includes the following steps:

[0039] S41. Input standard whole-brain imaging;

[0040] S42. Divide the whole brain region into multiple patches. Preferably, the whole brain can be divided into 64 patches according to a 4*4*4 space.

[0041] S43. Input standard images of key brain regions;

[0042] S44. Normalize images of individual key brain regions and construct the region of interest of the largest outer cube of the brain region;

[0043] S45. Divide the brain region R0l into multiple patches, and divide the specified brain region into 64 patches according to a 4*4*4 space.

[0044] S46. Calculate the correlation between any two patches in S42 and S45 respectively to form a brain region association map;

[0045] S47. Calculate the eigenvector centrality of each node in the brain region association map to obtain the center-weighted association map of the brain regions;

[0046] S48. Based on graph convolution and graph attention mechanisms, a neural network is established to obtain a brain region cognitive index prediction model;

[0047] S49. Using a brain region cognitive index prediction model, predict cognitive indicators for four whole brain regions and six key brain regions.

[0048] S410. Integrate to form whole-brain and key brain region cognitive features. The four whole-brain cognitive indicators and 24 key brain region cognitive indicators obtained in step S49 are spliced ​​together to obtain the final 28-dimensional whole-brain and key brain region cognitive features.

[0049] In step S47, the eigenvector centrality represents an index of the degree values ​​of a node's neighboring nodes. The eigenvector centrality of the i-th node is the sum of the eigenvector centralities of all its neighbors, and its calculation formula is as follows:

[0050]

[0051] In the formula, e i Represents the eigenvector centrality of node i; a ij Indicates the connection state between node i and node j; e j The eigenvector centrality of node j is represented by λ; λ represents the scaling constant.

[0052] In step S48, the brain region cognitive index prediction model uses four graph convolutional blocks to extract the center-weighted association graph features of the brain region. Each graph convolutional block contains four cascaded graph convolutional layers and four graph attention layers interleaved, as well as a pooling layer, a batch normalization layer, and a non-linear activation unit. Two fully connected layers are connected after the graph processing module for graph feature mining and compression, and then another fully connected layer is connected to output the model result as a 1*4 dimensional vector. The brain region cognitive index prediction model is trained using the center-weighted association graph of the patient's brain region as input, and brain age, logical memory score, visual memory score, and long-term delayed memory score as annotations, so as to obtain a brain region cognitive index prediction model that can regress and predict four numerical indicators.

[0053] In step S5, the method for forming interpretive data includes the following steps:

[0054] S51. Aggregate and stitch together the spatial features of the whole brain and key brain regions, the imaging features of the whole brain, and the cognitive features of the whole brain and key brain regions to obtain a 142-dimensional aggregated feature vector. Then, perform logarithmic and normalization operations on the aggregated feature vector to further reduce the numerical differences within the vector, and finally obtain 142-dimensional features of the whole brain and key brain regions.

[0055] S52. Aggregate the whole brain analysis results, including the whole brain spatial sign analysis results of cerebrospinal fluid, gray matter, white matter and total intracranial volume, the whole brain cognitive feature analysis results of brain age, logical memory score, visual memory score and long-term delayed memory score, and the whole brain imaging feature analysis results after visualization.

[0056] S53. Aggregate the analysis results of key brain regions, including the spatial feature analysis results and cognitive feature analysis results of six key brain regions: left hippocampus, right hippocampus, left parahippocampal gyrus, right parahippocampal gyrus, left cingulate gyrus, and right cingulate gyrus.

[0057] S54. The aggregated analysis results from steps S51-S53 are output in a structured manner to form explanatory data for reference.

[0058] In step S52, the method for visualizing the whole-brain image feature analysis results is as follows: the AD risk probability of each patch of the 64-dimensional whole brain obtained in step S35 is restored to a 4*4*4 dimension vector, and multiplied with the original 64 whole-brain image patches to obtain a high-risk distribution map of the whole-brain tissue space. This map is used as the visualization result of the whole-brain image feature analysis.

[0059] On the other hand, an analysis system for analyzing AD feature information based on brain medical imaging includes:

[0060] The data preprocessing module preprocesses the brain MRI images to be analyzed to obtain standard images of each brain region.

[0061] The spatial feature analysis module is used to analyze the spatial features of the whole brain and key brain regions;

[0062] The image feature analysis module is used to analyze the features of whole-brain images;

[0063] The cognitive feature analysis module is used to analyze the cognitive features of the whole brain and key brain regions.

[0064] The analysis results aggregation module is used to aggregate the above analysis results to form corresponding explanatory data.

[0065] The AD feature information analysis method and system based on brain medical imaging of the present invention has the following advantages:

[0066] Beneficial effects:

[0067] 1. A method using deep learning technology to analyze cognitive indicators such as brain age, logical memory score, visual memory score, and long-term delayed memory score from brain MRI images can provide doctors with analysis results of brain MRI images and other important medical indicators, which is helpful in assisting them in the diagnosis of Alzheimer's disease (AD) and also provides effective arguments and explanations for their subsequent scientific research analysis and interpretation.

[0068] 2. By calculating and analyzing the spatial features, imaging features, and cognitive features of brain regions, the features of the whole brain and key brain regions are finally formed, providing effective data for subsequent model interpretation and quantitative analysis.

[0069] 3. By using the patch analysis results to calculate and visualize the image features, high-risk areas of AD are identified in brain MRI images. This can help doctors quickly and accurately locate abnormal brain regions, assist doctors in focusing on the manifestations of symptoms and lesions, and provide a visualized analysis process and evidence for the subsequent diagnostic process. This further enriches the effective data types for subsequent scientific research and interpretation.

[0070] 4. Based on image analysis, this invention extracts a variety of quantitative, visual, and related image signs and clinical values, which can provide doctors with more comprehensive and contextually logical references in the diagnostic interpretation process and clinical research, and can effectively solve the problem of lack of result interpretation in traditional analysis processes. Attached Figure Description

[0071] Figure 1 This is a flowchart of the AD feature information analysis method based on brain medical imaging of the present invention;

[0072] Figure 2This is a flowchart of the spatial feature analysis method of the present invention.

[0073] Figure 3 This is a flowchart of the data preprocessing method of the present invention;

[0074] Figure 4 This is a flowchart of the whole-brain imaging feature analysis method of the present invention;

[0075] Figure 5 This is a flowchart of the method for analyzing cognitive features of the whole brain and key brain regions according to the present invention;

[0076] Figure 6 This is a block diagram illustrating the principle of the method for aggregating analysis results and forming interpretive data according to the present invention;

[0077] Figure 7 This is a schematic diagram of the final output result of the method of the present invention;

[0078] Figures 8a-8c yes Figure 7 Enlarged views of the three high-risk distribution maps in the image. Detailed Implementation

[0079] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0080] This invention relates to an AD feature information analysis method based on brain medical imaging, as follows: Figure 1 As shown, the specific steps include:

[0081] S1. Perform data preprocessing on the brain MRI images to be analyzed to obtain standard images of each brain region;

[0082] S2. Spatial feature analysis of the whole brain and key brain regions;

[0083] S3. Whole-brain imaging feature analysis;

[0084] S4. Analysis of cognitive characteristics of the whole brain and key brain regions;

[0085] S5. Aggregate the above analysis results to form corresponding explanatory data.

[0086] In step S1 above, a data preprocessing method is performed on the brain MRI images to be analyzed, specifically as follows: Figure 2 As shown, it includes the following steps:

[0087] S11. Acquire the input raw MRI image;

[0088] S12. Offset Field Correction: Reconstructs MRI images using the current image offset field to eliminate grayscale fluctuations;

[0089] S13. Register the corrected MRI images with the MNI152 standard brain template and remove non-brain regions. Specifically, register the MRI to the publicly available MNI152 standard brain template and remove non-brain regions such as the patient's face, neck, and skull from the original MRI images.

[0090] S14. Whole-brain voxel normalization, the formula for which voxel normalization is:

[0091]

[0092] In the formula, These are the normalized voxels; x is the actual input voxel. It is the average value of voxels; x std It is the standard deviation of voxels.

[0093] S15. Obtain the standard brain image after standard brain registration and normalization calculation;

[0094] S16. Register the AAL brain region templates of the whole brain standard images to obtain standard images of each brain region;

[0095] S17. Output the processed standard images of the whole brain and each brain region.

[0096] In step S2 above, the method for spatial feature analysis of the whole brain and key brain regions is as follows: Figure 3 As shown, the specific steps include:

[0097] S21. Input standard whole-brain imaging;

[0098] S22. Calculate the volume of cerebrospinal fluid, gray matter, white matter, and total intracranial volume;

[0099] S23. Input standard images of each brain region;

[0100] S24. Extract key brain regions and calculate the absolute volume of each key brain region. The key brain regions include the left hippocampus, right hippocampus, left parahippocampal gyrus, right parahippocampal gyrus, left cingulate gyrus, and right cingulate gyrus, a total of 6 brain regions;

[0101] S25. Calculation of relative spatial indices for the whole brain and key brain regions. This includes the ratio of the input whole brain volume to the total intracranial volume of a normal person, the ratio of the input volume of each key brain region to the volume of the same brain region in a normal person, and the ratio of the input volume of each key brain region to the input whole brain volume;

[0102] S26. Integration to form spatial features of the whole brain and key brain regions: The two spatial features of the whole brain (including the total intracranial volume of the whole brain and the ratio of the input whole brain to the total intracranial volume of the normal person) and the 18 spatial features of the six key brain regions (including absolute volume, the ratio of the input volume of each key brain region to the volume of the same brain region in the normal person, and the ratio of the input volume of each key brain region to the input whole brain volume) are connected to form 20-dimensional spatial features of the whole brain and key brain regions.

[0103] In step S3 above, the method for whole-brain imaging feature analysis is as follows: Figure 4 As shown, the specific steps include:

[0104] S31. Input standard images of the whole brain.

[0105] S32. Radiomics feature extraction, specifically: based on 24 features of gray-level co-occurrence matrix, 14 features of gray-level dependency matrix, 16 features of gray-level run matrix, and 16 features of gray-level region size matrix, principal component analysis and dimensionality reduction are performed to obtain 30-dimensional final effective features;

[0106] S33. Divide the whole brain region into multiple patches (spatial blocks). Preferably, the whole brain can be divided into 64 patches according to a 4*4*4 spatial pattern.

[0107] S34. A neural network is constructed based on 3D spatial convolution, and the network is trained using multi-instance learning technology to obtain an AD risk probability prediction model. This AD risk probability prediction model consists of model data input, a backbone feature extraction network, a head prediction branch network, and a cascaded final classification function. The model data input uses a single-channel black and white image with a size of 1*16*16*16; all patches are resized to this size before being fed into the network. The backbone feature extraction network uses 3D convolutional blocks with variable kernels as feature extraction unit modules, which include four cascaded 3D convolutional layers, one 3D pooling layer, one batch normalization layer (BN), and one non-linear activation unit (ReLU). The backbone feature extraction network contains three 3D convolutional block modules. The head prediction branch network includes one 3D convolutional block module and two fully connected layers (Full... Connected (FC) is used to aggregate and further compress the features of the backbone output to output the risk features of the current image to AD conversion; finally, Sigmoid is used as the activation function to convert the output of the head prediction branch into a risk probability between 0 and 1; in the training process of the AD risk probability prediction model, the model is trained using a multi-instance learning method, which gives the algorithm the ability to find the optimal solution in multiple patches, and finally obtains the trained AD risk probability prediction model.

[0108] S35. Predict the risk value of each patch and fuse them to obtain the whole-brain AD risk probability feature. In this case, the AD risk probability prediction model obtained in step S34 is used to predict the patch obtained in step S33, and the predicted risk values ​​of all patches are fused and spliced ​​into a 64-dimensional whole-brain AD risk probability feature.

[0109] S36. Fuse to form whole-brain imaging features: The radiomics features from step S32 and the whole-brain AD risk probability features from step S35 are fused and spliced ​​to obtain 94-dimensional whole-brain imaging features.

[0110] In step S4 above, the method for analyzing the cognitive characteristics of the whole brain and key brain regions is as follows: Figure 5 As shown, the specific steps include:

[0111] Step 4-1: Input standard images of the whole brain.

[0112] Step 4-2: Divide the whole brain region into multiple patches. Divide the whole brain into 64 patches according to a 4*4*4 spatial arrangement.

[0113] Step 4-3: Input standard images of key brain regions.

[0114] Step 4-4: Normalize the images of individual key brain regions and construct the largest circumscribed cube of the brain region (Region of Interest, ROI).

[0115] Steps 4-5: Divide the brain region ROI into multiple patches. Divide the specified brain region into 64 patches according to a 4*4*4 spatial arrangement.

[0116] Steps 4-6: Calculate the correlation between any two patches and form a brain region association map. Convert the voxels in the patch into vectors and calculate the Pearson correlation coefficient between any two patch vectors. Define this coefficient as the edge between the two nodes.

[0117] Steps 4-7: Calculate the eigenvector centrality of each node in the brain region association map to obtain the center-weighted association map of the brain regions. Eigenvector centrality represents the degree value of a node's neighboring nodes. The eigenvector centrality of the i-th node is the sum of the eigenvector centralities of all its neighbors, calculated using the following formula:

[0118]

[0119] In the formula, e i Represents the eigenvector centrality of node i; a ij Indicates the connection state between node i and node j; ej The eigenvector centrality of node j is represented by λ; λ represents the scaling constant.

[0120] Steps 4-8: A neural network is established based on graph convolution and graph attention mechanisms to obtain a brain region cognitive index prediction model. The model uses four graph convolutional blocks (GCNBlock) to extract the center-weighted association graph features of brain regions. Each graph convolutional block contains four cascaded graph convolutional layers (GCN) and four graph attention layers (GAT) interleaved, as well as one pooling layer, one batch normalization layer, and one non-linear activation unit. Two fully connected layers are connected after the graph processing module for graph feature mining and compression, and then another fully connected layer is connected to output the model results as a 1*4 dimensional vector. Using the center-weighted association graph of the patient's brain region as input, and brain age, logical memory score, visual memory score, and long-term delayed memory score as annotations, the model is trained so that it can regress and predict four numerical indicators, ultimately obtaining the brain region cognitive index prediction model.

[0121] Steps 4-9: Predict multidimensional cognitive indicators of brain regions. Using the brain region cognitive indicator prediction model obtained in Step 4-8, predict cognitive indicators for the whole brain and six key brain regions.

[0122] Steps 4-10: Integrate and form cognitive features of the whole brain and key brain regions. The four whole-brain cognitive indicators and 24 key brain region cognitive indicators obtained in steps 4-9 are spliced ​​together to obtain the final 28-dimensional cognitive features of the whole brain and key brain regions.

[0123] The aggregation analysis results described in step 5 above form the data interpretation method as follows: Figure 6 As shown, the specific steps include:

[0124] S51. Obtain the outputs of all the above methods and aggregate them to form 142-dimensional whole-brain and key brain region features. Specifically, the 20-dimensional whole-brain and key brain region spatial features obtained in step S26, the 94-dimensional whole-brain image features obtained in step S36, and the 28-dimensional whole-brain and key brain region cognitive features obtained in step S410 are aggregated and spliced ​​to obtain a 142-dimensional aggregated feature vector. Logarithmic and normalization operations are performed on the aggregated feature vector to further reduce the numerical differences within the vector, and finally, 142-dimensional whole-brain and key brain region features are obtained.

[0125] S52. Aggregate Whole-Brain Analysis Results: Aggregate the whole-brain spatial feature analysis results obtained in step S22 (analysis of cerebrospinal fluid, gray matter, white matter, and total intracranial volume), the whole-brain cognitive feature analysis results obtained in step S49 (analysis of brain age, logical memory score, visual memory score, and long-term delayed memory score), and the visualization results based on the whole-brain imaging feature analysis results obtained in step S35. The visualization method based on the whole-brain imaging feature analysis results obtained in step S35 is as follows: the AD risk probability of each patch of the 64-dimensional whole-brain image obtained in step S35 is restored to a 4*4*4 dimension vector, and multiplied with the original 64 whole-brain image patches to obtain a high-risk distribution map of the whole-brain tissue space (see...). Figures 8a-8c (The white lines in the brain in the image indicate high-risk areas with abnormalities). This image is used as the visualization result of whole-brain imaging feature analysis.

[0126] S53. Aggregate the analysis results of key brain regions, including the spatial feature analysis results and cognitive feature analysis results of six key brain regions: left hippocampus, right hippocampus, left parahippocampal gyrus, right parahippocampal gyrus, left cingulate gyrus, and right cingulate gyrus.

[0127] S54. Structured output of analysis results to form interpretive data, that is, structuring and outputting the analysis results obtained in steps S51, S52, and S53 above to form final interpretive data for reference, such as... Figure 7 As shown.

[0128] The analysis system of the AD feature information analysis method based on brain medical imaging of the present invention includes:

[0129] The data preprocessing module preprocesses the brain MRI images to be analyzed to obtain standard images of each brain region.

[0130] The spatial feature analysis module is used to analyze the spatial features of the whole brain and key brain regions;

[0131] The image feature analysis module is used to analyze the features of whole-brain images;

[0132] The cognitive feature analysis module is used to analyze the cognitive features of the whole brain and key brain regions.

[0133] The analysis results aggregation module is used to aggregate the above analysis results to form corresponding explanatory data.

[0134] Since the content and principles of this system are essentially the same as those of the aforementioned analysis method, they will not be described in detail again.

[0135] In summary, the AD feature information analysis method and system based on brain medical imaging of the present invention can analyze the spatial, imaging and cognitive indicators of the whole brain and key brain regions, which can provide effective data support and analysis process explanation for subsequent clinical research analysis. It can effectively solve the problem that existing technical means lack explanation of the reasoning process and can effectively assist doctors in conducting subsequent research.

[0136] Those skilled in the art should recognize that the above embodiments are only used to illustrate the purpose of the present invention and are not intended to limit the present invention. Any changes or modifications to the above embodiments that are within the scope of the present invention will fall within the scope of the claims of the present invention.

Claims

1. A method for analyzing AD feature information based on brain medical imaging, characterized in that, Includes the following steps: S1. Perform data preprocessing on the brain MRI images to be analyzed to obtain standard images of each brain region; S2. Perform spatial feature analysis of the whole brain and key brain regions; S3. Perform whole-brain imaging feature analysis; S4. Conduct cognitive feature analysis of the whole brain and key brain regions; S5. Aggregate the analysis results of S2-S4 above to form corresponding explanatory data; In step S2, the spatial feature analysis method includes the following steps: S21. Input standard whole-brain imaging; S22. Calculate the volume of cerebrospinal fluid, gray matter, white matter, and total intracranial volume; S23. Input standard images of each brain region; S24. Extract key brain regions and calculate the absolute volume of each key brain region. The key brain regions include the left hippocampus, right hippocampus, left parahippocampal gyrus, right parahippocampal gyrus, left cingulate gyrus, and right cingulate gyrus, for a total of six brain regions. S25. Calculate the relative spatial indicators of the whole brain and key brain regions, including the ratio of the total intracranial volume of the input whole brain to that of a normal person, the ratio of the volume of each key brain region to that of the same brain region in a normal person, and the ratio of the volume of each key brain region to that of the input whole brain. S26. Integrating to form spatial features of the whole brain and key brain regions: The total intracranial volume of the whole brain and the ratio of the total intracranial volume of the input whole brain to that of a normal person, as well as the absolute volume of the six key brain regions, the ratio of the volume of each key brain region to that of the same brain region in a normal person, and the ratio of the volume of each key brain region to the volume of the input whole brain, a total of 18 spatial features, are connected to form 20-dimensional spatial features of the whole brain and key brain regions. In step S3, the method for whole-brain image feature analysis includes the following steps: S31. Input standard whole-brain imaging; S32. Perform radiomics feature extraction. Based on the 24 features of gray-level co-occurrence matrix, 14 features of gray-level dependence matrix, 16 features of gray-level run matrix, and 16 features of gray-level region size matrix, perform principal component analysis and dimensionality reduction to obtain 30-dimensional final effective radiomics features. S33. Divide the entire brain region into multiple patches; S34. Construct a neural network based on 3D spatial convolution, train the network using multi-instance learning technology, and obtain an AD risk probability prediction model; S35. Use the AD risk probability prediction model to predict the risk value of each patch, and fuse and splice them to obtain a 64-dimensional whole-brain AD risk probability feature; S36. The radiomics features in step S32 and the whole-brain AD risk probability features in step S35 are fused and stitched together to obtain 94-dimensional whole-brain imaging features. In step S4, the method for analyzing the cognitive features of the whole brain and key brain regions includes the following steps: S41. Input standard whole-brain imaging; S42. Divide the entire brain region into multiple patches; S43. Input standard images of key brain regions; S44. Normalize images of individual key brain regions and construct the region of interest of the largest outer cube of the brain region; S45. Divide the brain region ROI into multiple patches, and divide the specified brain region into 64 patches according to a 4*4*4 space; S46. Calculate the correlation between any two patches in S42 and S45 respectively to form a brain region association map; S47. Calculate the eigenvector centrality of each node in the brain region association map to obtain the center-weighted association map of the brain regions; S48. Based on graph convolution and graph attention mechanisms, a neural network is established to obtain a brain region cognitive index prediction model; S49. Using a brain region cognitive index prediction model, predict cognitive indices for the whole brain and six key brain regions; S410. Integrate to form whole-brain and key brain region cognitive features. The four whole-brain cognitive indicators and 24 key brain region cognitive indicators obtained in step S49 are spliced ​​together to obtain the final 28-dimensional whole-brain and key brain region cognitive features. In step S5, the method for forming the corresponding interpretation data includes the following steps: S51. Aggregate and stitch together the spatial features of the whole brain and key brain regions, the imaging features of the whole brain, and the cognitive features of the whole brain and key brain regions to obtain a 142-dimensional aggregated feature vector. Then, perform logarithmic and normalization operations on the aggregated feature vector to further reduce the numerical differences within the vector, and finally obtain 142-dimensional features of the whole brain and key brain regions. S52. Aggregate the whole brain analysis results, including the whole brain spatial sign analysis results of cerebrospinal fluid, gray matter, white matter and total intracranial volume, the whole brain cognitive feature analysis results of brain age, logical memory score, visual memory score and long-term delayed memory score, and the whole brain imaging feature analysis results after visualization. S53. Aggregate the analysis results of key brain regions, including the spatial feature analysis results and cognitive feature analysis results of six key brain regions: left hippocampus, right hippocampus, left parahippocampal gyrus, right parahippocampal gyrus, left cingulate gyrus, and right cingulate gyrus. S54. The aggregated analysis results from steps S51-S53 are output in a structured manner to form explanatory data for reference.

2. The method for analyzing AD feature information based on brain medical imaging as described in claim 1, characterized in that, In step S1, the data preprocessing method includes the following steps: S11. Obtain the raw MRI images to be analyzed; S12. Perform bias field correction on the MRI image; S13. Register the corrected MRI images with the MNI152 standard brain template and remove non-brain regions; S14. Perform whole-brain voxel normalization, the normalization formula is: In the formula, These are normalized voxels; It is the actual input voxel; It is the average value of voxels; It is the standard deviation of voxels; S15. Obtain the standard brain image after standard brain registration and normalization calculation; S16. Perform AAL brain region template registration on the standard whole-brain images to obtain standard images of each brain region; S17. Output the standard images of the whole brain and each brain region obtained in S15 and S16.

3. The method for analyzing AD feature information based on brain medical imaging as described in claim 1, characterized in that: In step S34, the AD risk probability prediction model consists of model data input, a backbone feature extraction network, a head prediction branch network, and a final classification function cascade. The model data input uses a single-channel black and white image with a size of 1*16*16*16. All patches are resized to this size before being fed into the network. The backbone feature extraction network uses 3D convolutional blocks with variable kernels as feature extraction unit modules, which includes four cascaded 3D convolutional layers, one 3D pooling layer, one batch normalization layer, and one non-linear activation unit. The backbone feature extraction network contains three 3D convolutional block modules. The head prediction branch network includes one 3D convolutional block module and two fully connected layers, which are used to aggregate and further compress the features output by the backbone to output the risk features of the current image converting to AD. Finally, the Sigmoid function is used as the activation function to convert the output of the head prediction branch into a risk probability between 0 and 1. In the model training process, the model is trained using a multi-instance learning approach, which gives the algorithm the ability to find the optimal solution in multiple patches and finally obtains the trained AD risk probability prediction model.

4. The method for analyzing AD feature information based on brain medical imaging as described in claim 1, characterized in that: In step S47, the eigenvector centrality represents an index of the degree values ​​of a node's neighboring nodes. The eigenvector centrality of the i-th node is the sum of the eigenvector centralities of all its neighbors, and its calculation formula is as follows: In the formula, Indicate the eigenvector centrality of node i; This indicates the connection state between node i and node j; Indicate the eigenvector centrality of node j; Represents a proportionality constant; In step S48, the brain region cognitive index prediction model uses four graph convolutional blocks to extract the center-weighted association graph features of the brain region. Each graph convolutional block contains four cascaded graph convolutional layers and four graph attention layers interleaved, as well as a pooling layer, a batch normalization layer, and a non-linear activation unit. Two fully connected layers are connected after the graph processing module for graph feature mining and compression, and then another fully connected layer is connected to output the model result as a 1*4 dimensional vector. The brain region cognitive index prediction model is trained using the center-weighted association graph of the patient's brain region as input, and brain age, logical memory score, visual memory score, and long-term delayed memory score as annotations, so as to obtain a brain region cognitive index prediction model that can regress and predict four numerical indicators.

5. The method for analyzing AD feature information based on brain medical imaging as described in claim 1, characterized in that: In step S52, the method for visualizing the whole-brain image feature analysis results is as follows: the AD risk probability of each patch of the whole brain obtained in step S35 is restored to a 4*4*4 dimension vector, and multiplied with the original 64 whole-brain image patches to obtain a high-risk distribution map of the whole-brain tissue space. This map is used as the visualization result of the whole-brain image feature analysis.

6. The analysis system of the AD feature information analysis method based on brain medical imaging according to claim 1, characterized in that, include: The data preprocessing module preprocesses the brain MRI images to be analyzed to obtain standard images of each brain region. The spatial feature analysis module is used to analyze the spatial features of the whole brain and key brain regions; The image feature analysis module is used to analyze the features of whole-brain images; The cognitive feature analysis module is used to analyze the cognitive features of the whole brain and key brain regions. The analysis results aggregation module is used to aggregate the above analysis results to form corresponding explanatory data.

Images