Diffusion magnetic resonance imaging method and device for personalized human brain microenvironment brain injury detection

Abstract

The invention provides a diffusion magnetic resonance imaging method and device for personalized human brain microenvironment brain injury detection, and the method comprises the steps: selecting an embedded mask automatic encoder of a state space; the method is a new method for enhancing global context modeling and keeping more local details in the three-dimensional diffusion magnetic resonance image reconstruction process. A priori knowledge base is established by utilizing reconstruction errors of healthy diffusion magnetic resonance image data, and an embedded mask automatic encoder in a state space is applied to personalized brain injury detection. According to the method, three advanced diffusion models including a neurite direction dispersion and density imaging model, an average signal dispersion kurtosis model and a free water elimination double-chamber model are integrated, multi-dimensional microstructure measurement is constructed, and detection of cross-health-condition heterogeneity brain injury is achieved.

Description

Technical Field

[0001] The present invention relates to the technical fields of medical image analysis and artificial intelligence, and particularly relates to a diffusion magnetic resonance imaging method and device for personalized detection of brain injury in the human brain microenvironment. Background Art

[0002] The tissues and microstructure of the brain are highly interconnected, and changes in one area can affect the entire network. Therefore, anatomical and microstructure evaluation is crucial for detecting brain injury. Traditional structural magnetic resonance imaging can identify macroscopic lesions, but it lacks sensitivity to microstructure changes such as axonal integrity, resulting in "occult injuries" in approximately 30% of patients. Diffusion magnetic resonance imaging non-invasively characterizes the microstructure by quantifying the diffusion of water molecules. Diffusion tensor imaging measures anisotropic diffusion and provides insights into myelination and fiber density, but it has limitations in specificity. Multi-compartment models such as neurite orientation dispersion and density imaging, spherical mean technique, and diffusion kurtosis imaging white matter characterization technique have been developed to address this issue. These models are applicable to brain and spinal cord research and can be achieved through double-shell acquisition. Each model has its own assumptions that affect diffusion measurement, and combining them may help identify personalized changes associated with various health conditions.

[0003] Statistical evaluation in brain injury detection involves managing large datasets and hypothesis testing to evaluate the p-value of statistical significance. Standard methods include voxel-wise analysis or correlation analysis for group comparison, which may be sensitive to noise. Channel-based spatial statistics provides higher sensitivity and objectivity than traditional voxel-based methods for studying neurological device diseases that affect white matter integrity, such as multiple sclerosis and traumatic brain injury. Region of interest analysis focuses on detailed examination of specific brain regions, including registering anatomical landmarks or functional criteria, followed by statistical analysis. In recent years, machine learning has revolutionized medical imaging, bringing many innovative applications. Machine learning methods can effectively analyze brain images, extract stable patterns from neuroimaging data, and identify biomarkers. For example, support vector machines can distinguish children with medial temporal lobe epilepsy from healthy peers. Convolutional neural networks and models with residual and attention modules have shown strong adaptability. However, predictive modeling that compares patient groups with matched controls is unrealistic for clinically heterogeneous populations such as neurological or psychiatric diseases. Although diffusion magnetic resonance imaging is mainly used for diagnosing acute ischemic stroke or monitoring tumor invasion, studies have shown its potential to identify significant microstructure changes at the individual level, and the lack of a diffusion magnetic resonance imaging technology framework for single-patient analysis hinders personalized diagnosis. The shift from group comparison to personalized assessment is urgently needed to strengthen research on diverse clinical conditions and promote accurate diagnosis of rare diseases. Summary of the Invention

[0004] To overcome the challenges of the existing technology, the present invention provides a diffusion magnetic resonance imaging method and device for personalized detection of brain injury in the human brain microenvironment, which improves the sensitivity and specificity of injury detection through multi-dimensional microstructure parameter integration; optimizes the reconstruction accuracy of three-dimensional diffusion magnetic resonance imaging data, constructs a prior knowledge base of healthy samples; and realizes individualized abnormal scoring to assist early and accurate diagnosis.

[0005] The technical solution for achieving the object of the present invention is as follows:

[0006] A diffusion magnetic resonance imaging method for personalized detection of brain injury in the human brain microenvironment, comprising:

[0007] Step 1, based on the neurite orientation dispersion and density imaging model, the mean signal diffusion kurtosis model, and the free water elimination two-compartment model, construct a multi-dimensional microstructure index device;

[0008] Step 2, based on the multi-dimensional microstructure index device, design a multi-modal diffusion magnetic resonance image reconstruction model, which is pre-trained using the three-dimensional diffusion magnetic resonance imaging data of healthy individuals, and then uses the three-dimensional diffusion magnetic resonance imaging abnormal data of individuals, and optimizes the reconstruction process of the model through a non-repeating random masking strategy to generate a prior knowledge base of the reconstruction error of healthy samples;

[0009] Step 3, use the multi-modal diffusion magnetic resonance image reconstruction model to reconstruct the diffusion magnetic resonance imaging data of the individual to be detected, calculate the cosine distance between the pixel-level reconstruction error and the prior knowledge base, and generate an abnormal scoring map for locating abnormal brain regions.

[0010] Further, the multi-modal diffusion magnetic resonance image reconstruction model includes a selective state space embedded masked autoencoder framework constructed using a residual convolutional neural network, a selective state space, and a self-attention mechanism.

[0011] Further, the process of the multi-modal diffusion magnetic resonance image reconstruction model for image processing is as follows:

[0012] Step 2.1, convert the input diffusion magnetic resonance image data into a low-dimensional long sequence X through patch embedding by a two-layer residual convolutional neural network;

[0013] Step 2.2, the selective state space uses convolution and symmetric branches to preserve local information;

[0014] Step 2.3, perform image processing through a multi-head self-attention mechanism to avoid limited local receptive fields;

[0015] Step 2.4, loop steps 2.2 and 2.3 m times to achieve global context modeling and multi-scale feature fusion.

[0016] Furthermore, for the encoder in the selective state-space embedded masked autoencoder framework, m equals 8, and for the decoder, m equals 4.

[0017] Furthermore, the patch embedding is specifically as follows: segment the three-dimensional diffusion magnetic resonance image data and map it to a low-dimensional space.

[0018] Furthermore, the output calculation of the selective state space is:

[0019]

[0020] where Linear(Cin, Cout)(.) represents a linear layer, where Cin and Cout are the input and output embedding dimensions, Scan is the selective scan operation, σ is the activation function using the sigmoid linear unit, and in addition, Conv and Concat represent one-dimensional convolution and concatenation operations.

[0021] Furthermore, optimizing the reconstruction process through the non-repeating random masking strategy specifically includes:

[0022] Let the sequence processed by the multi-modal diffusion magnetic resonance image reconstruction model be Generate a random permutation matrix The total sorting is

[0023] Π n = argsort(η n ), η n ~ U(0, 1) L

[0024] Divide the permuted indices into n non-overlapping sub-blocks, and define the length of the i-th sub-block as:

[0025]

[0026] For the i-th round of masking, construct a temporary mask in the permuted order:

[0027]

[0028] In the formula is the start position of the mask, and e i = s i + l i is the end position;

[0029] Restore the mask in the original order through inverse permutation:

[0030] M (i) = gather(M' (i) , Π -1 , dim = 1)

[0031] In the formula, the role of the gather function is to collect the input tensor M' according to the given index Π -1 tensor, and collect the values of the input tensor M' on the specified dimension dim (i) value.

[0032] Through n rounds of non-repetitive masks, each round retains a different subset, and outputs the structural information degree of diffusion magnetic resonance imaging data from different mask perspectives.

[0033] Furthermore, a prior knowledge base of reconstruction errors for healthy samples is generated, specifically including: using a pre-trained multi-modal diffusion magnetic resonance image reconstruction model to generate healthy diffusion magnetic resonance image data, and calculating the reconstruction error of each image patch to obtain an error vector; storing all error vectors of healthy samples as a prior knowledge base where e i represents the error feature of the i-th regular block.

[0034] Furthermore, the specific steps of step 3 include:

[0035] Input the test image into the multi-modal diffusion magnetic resonance image reconstruction model, and calculate the similarity between the pixel-level reconstruction error of the model and the errors in the prior knowledge base in the error:

[0036]

[0037] where d is the cosine similarity, and e testi is the error of the i-th test block;

[0038] Generate an image-level anomaly score based on the cosine similarity:

[0039]

[0040] where q is defined as the average distance of the 1% most abnormal pixels.

[0041] A diffusion magnetic resonance imaging device for personalized detection of brain injury in the human brain microenvironment, comprising:

[0042] A multi-dimensional microstructural index device construction unit, which constructs a multi-dimensional microstructural index device based on the neurite orientation dispersion and density imaging model, the mean signal diffusion kurtosis model, and the free water elimination two-compartment model;

[0043] The multi-modal diffusion magnetic resonance image reconstruction model training unit designs a multi-modal diffusion magnetic resonance image reconstruction model based on a multi-dimensional microstructural index device. This model is pre-trained using three-dimensional diffusion magnetic resonance imaging data of healthy individuals, and then uses the three-dimensional diffusion magnetic resonance imaging abnormal data of an individual. Furthermore, through a non-repeating random masking strategy, the reconstruction process of the model is optimized to generate a prior knowledge base of reconstruction errors for healthy samples.

[0044] The detection unit uses the multi-modal diffusion magnetic resonance image reconstruction model to reconstruct the diffusion magnetic resonance imaging data of the individual to be detected, calculates the cosine distance between the pixel-level reconstruction error and the prior knowledge base, and generates an abnormal score map for locating abnormal brain regions.

[0045] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention adopts the neurite orientation dispersion and density imaging model, the mean signal diffusion kurtosis model, and the free water elimination two-compartment model, which enhances tissue characterization and improves the sensitivity and specificity of detection; adopts a selective state space embedded masked autoencoder framework, which solves the limitations of the pure selective state space in global modeling and local detail capture, making the framework more suitable for high-dimensional tasks; segments the three-dimensional diffusion magnetic resonance image data and maps it to a low-dimensional space, combines the high-level semantic information and local detail information in the image, and improves the performance of visual tasks; applies a random masking strategy during training. This can prevent the model from relying on visible adjacent blocks and ensure that it learns more robust and generalizable features. The masking strategy is improved in the abnormal detection stage to avoid the instability of the results caused by random masking in the abnormal detection stage; the present invention is particularly suitable for the identification of early subtle brain injuries and the differentiation of heterogeneous pathological features. Description of the Drawings

[0046] Figure 1 It is a flowchart of a diffusion magnetic resonance imaging method for personalized human brain microenvironment brain injury detection in an embodiment.

[0047] Figure 2 It is a schematic diagram of the summary of all processes in an embodiment.

[0048] Figure 3 It is a result diagram for 3 tumor patients under different masking times n and the number of healthy samples m for constructing the prior knowledge base.

[0049] Figure 4 It is a reconstruction image, a reconstruction error heat map, and a distribution difference map for 2 tumor patients.

[0050] Figure 5 It is a reconstruction image, a reconstruction error heat map, and a distribution difference map for 2 patients with mental diseases.

[0051] Figure 6Box plots of abnormal voxels and healthy voxels of patients with neurological diseases marked by the methods in multiple embodiments. DETAILED DESCRIPTION

[0052] Figure 1 The flow chart of a diffusion magnetic resonance imaging method for personalized human brain microenvironment brain injury detection in one embodiment is shown, comprising the following steps:

[0053] Step 1: A multidimensional microstructure index system was constructed by integrating three advanced diffusion magnetic resonance models: neurite orientation dispersion and density imaging model, average signal diffusion kurtosis model, and free water elimination two-compartment model.

[0054] The free water elimination two-compartment model eliminates cerebrospinal fluid contamination by introducing free water, and its derived parameters - axial diffusivity, radial diffusivity, mean diffusivity and fractional anisotropy - characterize the macroscopic diffusion properties of white matter fibers. The mean signal diffusion kurtosis model extracts parameters such as intrinsic diffusivity, axonal volume fraction, mean signal diffusivity, mean signal kurtosis and microscopic anisotropy, revealing the microscopic complexity and heterogeneity of the tissue. In addition, the neurite orientation dispersion and density imaging model supplements the intracellular volume fraction, isotropic volume fraction and orientation diffusion index to depict the directional distribution of neurites. Combining model parameters enhances tissue characterization and improves the sensitivity and specificity of brain lesion detection.

[0055] Step 2: Design a multimodal diffusion magnetic resonance image reconstruction model based on the multidimensional microstructure indicator device.

[0056] The multimodal diffusion magnetic resonance image reconstruction model adopts a selective state space embedded mask autoencoder framework, which integrates residual convolutional neural network, selective state space and self-attention mechanism, solves the limitations of simple selective state space in global modeling and local detail capture, and makes the framework more suitable for high-dimensional tasks. The steps of data processing in this framework are as follows:

[0057] Step 2.1, the residual convolutional neural network has efficient local perception ability, which can quickly extract basic features and reduce the computational complexity of subsequent modules while retaining the structural details of the brain as much as possible. Only two layers of residual convolution are used in the architecture. The input diffusion magnetic resonance image data is converted into a low-dimensional long sequence X through patch embedding. The calculation process of X in a single layer of residual convolution is as follows:

[0058]

[0059]

[0060] GELU and BN stand for Gaussian Error Linear Unit activation function and batch normalization, respectively.

[0061] Step 2.2. To overcome the limitations of traditional selective state - space causal convolution, traditional convolution and symmetric branches are adopted to preserve local information and better capture the microscopic features found in brain diffusion magnetic resonance data. The output of the improved selective state - space is calculated as:

[0062]

[0063] where Linear(Cin,Cout)(.) represents a linear layer, where Cin and Cout are the input and output embedding dimensions, Scan is the selective scan operation, σ is the activation function using the sigmoid linear unit. In addition, Conv and Concat represent one - dimensional convolution and concatenation operations.

[0064] Step 2.3. Explicitly model long - range dependencies through the self - attention mechanism to make up for the deficiency of the selective state - space in global interaction, which is beneficial to better maintaining the global information in multi - modal diffusion magnetic resonance image data. The feature map is divided into non - overlapping windows to calculate local attention to reduce the computational complexity. And cross - region information interaction is achieved through window offset to avoid the limitation of local receptive fields. This is a general multi - head self - attention mechanism:

[0065]

[0066] where Q, K, V represent query, key, and value respectively, and d h is the number of attention heads.

[0067] Step 2.4. Loop m times for the relevant calculations in Steps 2.2 and 2.3 to achieve global context modeling and multi - scale feature fusion. To maintain an appropriate balance in the design of the encoder and decoder, the encoder provides sufficient feature extraction ability, while the decoder efficiently reconstructs the image, thus achieving a good balance between model complexity and performance. m equals 8 in the encoder and m equals 4 in the decoder.

[0068] In this study, patches and position embeddings are re - defined by dividing the data into non - overlapping three - dimensional patches, the three - dimensional diffusion magnetic resonance image data is segmented and mapped to a low - dimensional space, combining the high - level semantic information and local detail information in the image, and improving the performance of visual tasks.

[0069] And a random masking strategy is applied during training. This can prevent the model from relying on visible adjacent blocks, ensuring that it learns more robust and generalizable features. The masking strategy is improved in the anomaly detection stage to avoid the instability of results caused by random masking in the anomaly detection stage. n rounds of non - repeating random masking are adopted.

[0070] Let the sequence processed by the encoder be Generate a random permutation matrix The total sorting is

[0071] Π n = argsort(η n ), η n ~ U(0, 1) L

[0072] The permuted indices are divided into n non - overlapping sub - blocks. Define the length of the i - th sub - block as:

[0073]

[0074] For the i - th round of masking, in the permuted order, construct a temporary mask:

[0075]

[0076] Here is the start position of the mask, and e i = s i + l i is the end position.

[0077] In subsequent steps, restore the mask in the original order through inverse permutation:

[0078] M (i) = gather(M' (i) , Π -1 , dim = 1)

[0079] Through n rounds of non - repeating masks, each round retains a different subset (with a ratio of 1 / n), ensuring that each position has the opportunity to participate in the reconstruction process, and outputting the structural information of diffusion - weighted magnetic resonance imaging data from different mask perspectives, thereby improving the reconstruction accuracy.

[0080] Step 3: Use the pre - trained selective state - space embedded masked auto - encoder model to generate a reconstruction of healthy diffusion - weighted magnetic resonance image data. Calculate the reconstruction error of each image block to obtain an error vector. Store all the error vectors of healthy samples as a prior knowledge base where e i represents the error feature of the i - th regular block. Perform selective state - space embedded masked auto - encoder reconstruction on the test image, and calculate the reconstruction error e test for each block. For the error e testi of the i - th test block, calculate its similarity with the errors in the memory bank :

[0081]

[0082] where d(e1, e2) is the cosine similarity

[0083] Generating image-level anomaly scores based on cosine similarity:

[0084]

[0085] Among them, q is defined as the average distance of the 1% most abnormal pixels. The training and detection processes are shown in Figure 2 .

[0086] Regions with higher final anomaly scores, that is, the damaged regions detected by the method, are used for subsequent auxiliary analysis of pathological features based on the distribution differences of damaged voxels and healthy voxels in various microscopic indicators.

[0087] The present invention also provides a diffusion magnetic resonance imaging device for personalized detection of brain injury in the human brain microenvironment, including:

[0088] A multi-dimensional microstructural index device construction unit that constructs a multi-dimensional microstructural index device based on the neurite orientation dispersion and density imaging model, the mean signal diffusion kurtosis model, and the free water elimination two-compartment model;

[0089] A multi-modal diffusion magnetic resonance image reconstruction model training unit that designs a multi-modal diffusion magnetic resonance image reconstruction model based on the multi-dimensional microstructural index device. This model is pre-trained using three-dimensional diffusion magnetic resonance imaging data of healthy individuals, and then uses three-dimensional diffusion magnetic resonance imaging anomaly data of individuals. Through a non-repetitive random masking strategy, the reconstruction process of the model is optimized to generate a prior knowledge base of reconstruction errors for healthy samples;

[0090] A detection unit that uses the multi-modal diffusion magnetic resonance image reconstruction model to reconstruct the diffusion magnetic resonance imaging data of the individual to be detected, calculates the cosine distance between the pixel-level reconstruction error and the prior knowledge base, and generates an anomaly score map for locating abnormal brain regions.

[0091] In summary, the proposed method based on microstructural information and selective state space masked autoencoder enhances voxel-level feature reconstruction, allows early diagnosis of various neurological device diseases, and captures microstructural changes in brain injury. The reconstruction error from healthy dMRI data is used to establish a knowledge base for personalized detection. Figure 3 The results of 3 tumor patients under different masking times n and the number of healthy samples m for constructing the prior knowledge base are shown, Figure 4 The reconstructed images, reconstruction error heat maps, and distribution difference maps of 2 tumor patients are shown. Figure 5 The reconstructed images, reconstruction error heat maps, and distribution difference maps of 2 patients with mental diseases are shown. Figure 6Box plots of abnormal and healthy voxels of patients with neurological diseases marked by the method in multiple embodiments are shown. The experimental results show that based on microstructure information and selective state space masked autoencoders, a sufficient knowledge base can be established on the basis of limited healthy control data, facilitating personalized detection of brain damage in the human brain microenvironment. This study promotes the shift towards individualized analysis and provides innovative methods for early detection of subtle injuries and advancement of precision medicine.

[0092] The above-described embodiments merely represent several embodiments of the present invention. The description thereof is relatively specific and detailed, but should not be construed as a limitation on the scope of the patent for the present invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent for the present invention should be subject to the appended claims.

Claims

1. A diffusion magnetic resonance imaging method for personalized detection of brain injury in the human brain microenvironment, characterized in that, Including: Step 1: Based on the neurite orientation dispersion and density imaging model, the mean signal diffusion kurtosis model, and the free water elimination two-compartment model, construct a multi-dimensional microstructure index device; Step 2: Based on the multi-dimensional microstructure index device, design a multi-modal diffusion magnetic resonance image reconstruction model. This model is pre-trained using three-dimensional diffusion magnetic resonance imaging data of healthy individuals, and then uses the three-dimensional diffusion magnetic resonance imaging abnormal data of an individual. Through a non-repeating random masking strategy, optimize the reconstruction process of the model to generate a prior knowledge base of reconstruction errors for healthy samples; Step 3: Use the multi-modal diffusion magnetic resonance image reconstruction model to reconstruct the diffusion magnetic resonance imaging data of the individual to be detected, calculate the cosine distance between the pixel-level reconstruction error and the prior knowledge base, and generate an abnormal score map for localizing abnormal brain regions.

2. A diffusion magnetic resonance imaging method for personalized detection of brain injury in the human brain microenvironment according to claim 1, characterized in that, The multi-modal diffusion magnetic resonance image reconstruction model includes a selective state space embedded masked autoencoder framework built using a residual convolutional neural network, a selective state space, and a self-attention mechanism.

3. A diffusion magnetic resonance imaging method for personalized detection of brain injury in the human brain microenvironment according to claim 2, characterized in that, The process of the multi-modal diffusion magnetic resonance image reconstruction model for image processing is as follows: Step 2.1: Through a two-layer residual convolutional neural network, convert the input diffusion magnetic resonance image data into a low-dimensional long sequence X through patch embedding; Step 2.2: The selective state space uses convolution and symmetric branches to preserve local information; Step 2.3: Perform image processing through a multi-head self-attention mechanism to avoid limited local receptive fields; Step 2.4: Repeat steps 2.2 and 2.3 m times to achieve global context modeling and multi-scale feature fusion.

4. A diffusion magnetic resonance imaging method for personalized detection of brain injury in the human brain microenvironment according to claim 3, characterized in that, For the encoder in the selective state space embedded masked autoencoder framework, m is equal to 8, and for the decoder, m is equal to 4.

5. A diffusion magnetic resonance imaging method for personalized human brain microenvironment brain injury detection according to claim 3, wherein The specific process of patch embedding is as follows: Segment the three-dimensional diffusion magnetic resonance image data and map it to a low-dimensional space.

6. A diffusion magnetic resonance imaging method for personalized human brain microenvironment brain injury detection according to claim 2, wherein The output calculation of the selective state space is as follows: Where, Linear(Cin,Cout)(.) represents a linear layer, where Cin and Cout are used as input and output embedding dimensions, Scan is a selective scanning operation, σ is an activation function using a sigmoid linear unit. In addition, Conv and Concat represent one-dimensional convolution and concatenation operations.

7. A diffusion magnetic resonance imaging method for personalized human brain microenvironment brain injury detection according to claim 1, characterized in that, Optimizing the reconstruction process through a non-repeating random masking strategy specifically includes: Let the sequence processed by the multi-modal diffusion magnetic resonance image reconstruction model be Generate a random permutation matrix The total sorting is ∏ n = argsoft(η n ), η n ~ U(0,1) L Divide the permuted indices into n non-overlapping sub-blocks, and define the length of the i-th sub-block as: For the i-th round of masking, in the permuted order, construct a temporary mask: where is the mask start position, and e i = s i + l i is the end position; Restore the mask in the original order through inverse permutation: M (i) = gather(M' (i) , ∏ -1 , dim = 1) In the formula, the role of the gather function is to collect the input tensor M' -1 according to the given index ∏ (i) tensor at the specified dimension dim. Through n rounds of non-repeating masking, each round retains a different subset, and outputs the structural information degree of the diffusion magnetic resonance imaging data from different masking perspectives.

8. A diffusion magnetic resonance imaging method for personalized detection of brain injury in the human brain microenvironment according to claim 1, characterized in that, Generate a prior knowledge base for the reconstruction error of healthy samples, specifically including: using a pre-trained multi-modal diffusion magnetic resonance image reconstruction model to generate healthy diffusion magnetic resonance image data, and calculating the reconstruction error X of each image patch e = to obtain an error vector; store all the error vectors of the healthy samples as a prior knowledge base where e i represents the error feature of the i-th regular patch.

9. A diffusion magnetic resonance imaging method for personalized human brain microenvironment brain injury detection according to claim 8, characterized in that, The specific content of step 3 is as follows: Input the test image into the multi-modal diffusion magnetic resonance image reconstruction model, and calculate the cosine similarity between the pixel-level reconstruction error of the model and the error in the prior knowledge base in the error: where d is the cosine similarity, and e testi is the error of the i-th test block; Generate an image-level abnormal score based on cosine similarity: Where, q is defined as the average distance of the 1% most abnormal pixels.

10. A diffusion magnetic resonance imaging apparatus for implementing the method according to any one of claims 1-9, characterized in that, Including: A multi-dimensional microstructure index device construction unit, which constructs a multi-dimensional microstructure index device based on the neurite orientation dispersion and density imaging model, the mean signal diffusion kurtosis model, and the free water elimination two-compartment model; The multi-modal diffusion magnetic resonance image reconstruction model training unit designs a multi-modal diffusion magnetic resonance image reconstruction model based on a multi-dimensional microstructural index device. This model is pre-trained using three-dimensional diffusion magnetic resonance imaging data of healthy individuals, and then uses the three-dimensional diffusion magnetic resonance imaging abnormal data of an individual. Furthermore, through a non-repeating random masking strategy, the reconstruction process of the model is optimized to generate a prior knowledge base of reconstruction errors for healthy samples. The detection unit uses the multi-modal diffusion magnetic resonance image reconstruction model to reconstruct the diffusion magnetic resonance imaging data of the individual to be detected, calculates the cosine distance between the pixel-level reconstruction error and the prior knowledge base, and generates an abnormal score map for locating abnormal brain regions.

Images