Brain age prediction method and device based on regional segmentation

Abstract

The present invention discloses a brain age prediction method and device based on regional segmentation. The method obtains a diffusion tensor image to be processed corresponding to the target measurement subject, determines multiple diffusion index data corresponding to each voxel in the diffusion tensor image to be processed, and generates multiple diffusion index images based on the multiple diffusion index data. The multiple diffusion index images are processed according to a brain fiber bundle template to obtain a first eigenvector corresponding to each brain region. The eigenvector is used to describe the image features corresponding to each brain region. The first eigenvector corresponding to each brain region is processed using a pre-trained brain age prediction model to obtain a brain age prediction result corresponding to the target measurement subject. The brain age prediction model is used to integrate the image feature information of each brain region in multiple diffusion index images to accurately obtain the white matter information of each brain region, and the brain age is accurately predicted based on the white matter information of each brain region, thereby improving the accuracy of the brain age prediction result.

Description

Technical Field

[0001] The present invention belongs to the field of computer technology, and in particular relates to a brain age prediction method and device based on region segmentation. Background Art

[0002] Aging is a series of changes that occur with the passage of time, characterized by a decline in physiological function and reduced metabolism. These changes encompass a wide range of aspects, including behavior, organs, tissues, and cellular molecular processes. Brain aging is a particularly crucial component, and is often difficult to reverse. Some individuals may even develop degenerative central nervous system diseases (CNS) with a large patient base, with an increasingly younger prevalence. Severe CNS degenerative diseases can negatively impact both ability and quality of life.

[0003] However, given the huge patient base and lack of specialist doctors in primary medical institutions, there is an urgent need for a method that can accurately predict brain age, so as to quickly identify the degree of brain aging and degenerative diseases of the central nervous system based on the brain age prediction results, in order to alleviate the contradiction between supply and demand in primary medical care. Summary of the Invention

[0004] In view of this, the present invention provides a brain age prediction method and device based on region segmentation that solves or partially solves the above technical problems.

[0005] In a first aspect, an embodiment of the present invention provides a brain age prediction method based on region segmentation, the method comprising:

[0006] Obtaining a diffusion tensor image to be processed corresponding to the target measurement person;

[0007] determining a plurality of diffusion index data corresponding to each voxel in the diffusion tensor image to be processed, and generating a plurality of diffusion index images according to the plurality of diffusion index data;

[0008] Processing the plurality of diffusion index images according to the brain fiber bundle template to obtain a first eigenvector corresponding to each brain region, wherein the eigenvector is used to describe the image features corresponding to each brain region;

[0009] The first eigenvectors corresponding to the respective brain regions are processed using a pre-trained brain age prediction model to obtain a brain age prediction result corresponding to the target person being measured.

[0010] In a second aspect, an embodiment of the present invention provides a brain age prediction device, the device comprising:

[0011] An acquisition module, used for acquiring a diffusion tensor image to be processed corresponding to the target measurement subject;

[0012] a determination module, configured to determine a plurality of diffusion index data corresponding to each voxel in the diffusion tensor image to be processed, and generate a plurality of diffusion index images according to the plurality of diffusion index data;

[0013] a first processing module, configured to process the plurality of diffusion index images according to a brain fiber bundle template to obtain a first eigenvector corresponding to each brain region, wherein the eigenvector is used to describe an image feature corresponding to each brain region;

[0014] The second processing module is used to process the first eigenvectors corresponding to the various brain regions using a pre-trained brain age prediction model to obtain a brain age prediction result corresponding to the target person being measured.

[0015] In a third aspect, an embodiment of the present invention provides an electronic device, comprising: a memory and a processor; wherein,

[0016] Memory, used to store programs;

[0017] A processor, coupled to the memory, configured to execute a program stored in the memory for:

[0018] Obtaining a diffusion tensor image to be processed corresponding to the target measurement person;

[0019] determining a plurality of diffusion index data corresponding to each voxel in the diffusion tensor image to be processed, and generating a plurality of diffusion index images according to the plurality of diffusion index data;

[0020] Processing the plurality of diffusion index images according to the brain fiber bundle template to obtain a first eigenvector corresponding to each brain region, wherein the eigenvector is used to describe the image features corresponding to each brain region;

[0021] The first eigenvectors corresponding to the respective brain regions are processed using a pre-trained brain age prediction model to obtain a brain age prediction result corresponding to the target person being measured.

[0022] In a fourth aspect, an embodiment of the present invention provides a computer storage medium for storing a computer program, which, when executed by a computer, implements the following method:

[0023] Obtaining a diffusion tensor image to be processed corresponding to the target measurement person;

[0024] determining a plurality of diffusion index data corresponding to each voxel in the diffusion tensor image to be processed, and generating a plurality of diffusion index images according to the plurality of diffusion index data;

[0025] Processing the plurality of diffusion index images according to the brain fiber bundle template to obtain a first eigenvector corresponding to each brain region, wherein the eigenvector is used to describe the image features corresponding to each brain region;

[0026] The first eigenvectors corresponding to the respective brain regions are processed using a pre-trained brain age prediction model to obtain a brain age prediction result corresponding to the target person being measured.

[0027] In the brain age prediction scheme provided by an embodiment of the present invention, when performing brain age prediction, first, a diffusion tensor image to be processed is obtained, and then the diffusion index data corresponding to each voxel in the diffusion tensor image to be processed is determined. Based on the multiple diffusion index data, multiple diffusion index images are generated. Next, the multiple diffusion index images are processed according to the brain fiber bundle template to obtain the first eigenvector corresponding to each brain region. The eigenvector is used to describe the image characteristics corresponding to each brain region. Finally, using a pre-trained brain age prediction model, the first eigenvector corresponding to each brain region is processed to obtain a brain age prediction result corresponding to the diffusion tensor image to be processed.

[0028] In a fifth aspect, an embodiment of the present invention provides a method for training a brain age prediction model, the method comprising:

[0029] collecting training samples, wherein the training samples include diffusion tensor images corresponding to a plurality of measurers;

[0030] Determining a plurality of diffusion index data corresponding to each voxel in the diffusion tensor image, and generating a plurality of diffusion index image samples based on the plurality of diffusion index data;

[0031] determining a reference brain age corresponding to the plurality of diffusion index image samples;

[0032] Processing the plurality of diffusion index image samples according to the brain fiber bundle template to obtain feature vectors corresponding to the respective brain regions, wherein the feature vectors are used to describe image features corresponding to the respective brain regions;

[0033] Processing the first eigenvectors corresponding to the respective brain regions using a brain age prediction model to obtain a brain age prediction result;

[0034] The brain age prediction model is trained according to the reference brain age and the brain age prediction result.

[0035] In a sixth aspect, an embodiment of the present invention provides a brain age prediction model training device, the device comprising:

[0036] An acquisition module, configured to acquire training samples, wherein the training samples include diffusion tensor images corresponding to a plurality of measurers;

[0037] a generating module, configured to determine a plurality of diffusion index data corresponding to each voxel in the diffusion tensor image, and generate a plurality of diffusion index image samples according to the plurality of diffusion index data;

[0038] a determination module, configured to determine a reference brain age corresponding to the plurality of diffusion index image samples;

[0039] A first processing module is configured to process the plurality of diffusion index image samples according to a brain fiber bundle template to obtain a feature vector corresponding to each brain region, wherein the feature vector is used to describe the image features corresponding to each brain region;

[0040] A second processing module is configured to process the first eigenvectors corresponding to the respective brain regions using a brain age prediction model to obtain a brain age prediction result;

[0041] A training module is used to train the brain age prediction model based on the reference brain age and the brain age prediction result.

[0042] In a seventh aspect, an embodiment of the present invention provides an electronic device, comprising: a memory and a processor; wherein,

[0043] Memory, used to store programs;

[0044] A processor, coupled to the memory, configured to execute a program stored in the memory for:

[0045] collecting training samples, wherein the training samples include diffusion tensor images corresponding to a plurality of measurers;

[0046] Determining a plurality of diffusion index data corresponding to each voxel in the diffusion tensor image, and generating a plurality of diffusion index image samples based on the plurality of diffusion index data;

[0047] determining a reference brain age corresponding to the plurality of diffusion index image samples;

[0048] Processing the plurality of diffusion index image samples according to the brain fiber bundle template to obtain feature vectors corresponding to the respective brain regions, wherein the feature vectors are used to describe image features corresponding to the respective brain regions;

[0049] Processing the first eigenvectors corresponding to the respective brain regions using a brain age prediction model to obtain a brain age prediction result;

[0050] The brain age prediction model is trained according to the reference brain age and the brain age prediction result.

[0051] In an eighth aspect, an embodiment of the present invention provides a computer storage medium for storing a computer program, which, when executed by a computer, implements the following method:

[0052] collecting training samples, wherein the training samples include diffusion tensor images corresponding to a plurality of measurers;

[0053] Determining a plurality of diffusion index data corresponding to each voxel in the diffusion tensor image, and generating a plurality of diffusion index image samples based on the plurality of diffusion index data;

[0054] determining a reference brain age corresponding to the plurality of diffusion index image samples;

[0055] Processing the plurality of diffusion index image samples according to the brain fiber bundle template to obtain feature vectors corresponding to the respective brain regions, wherein the feature vectors are used to describe image features corresponding to the respective brain regions;

[0056] Processing the first eigenvectors corresponding to the respective brain regions using a brain age prediction model to obtain a brain age prediction result;

[0057] The brain age prediction model is trained according to the reference brain age and the brain age prediction result.

[0058] In the brain age prediction scheme provided by an embodiment of the present invention, when performing brain age prediction, first, a diffusion tensor image to be processed is obtained, and then the diffusion index data corresponding to each voxel in the diffusion tensor image to be processed is determined. Based on the multiple diffusion index data, multiple diffusion index images are generated. Next, the multiple diffusion index images are processed according to the brain fiber bundle template to obtain the first eigenvector corresponding to each brain region. The eigenvector is used to describe the image characteristics corresponding to each brain region. Finally, using a pre-trained brain age prediction model, the first eigenvector corresponding to each brain region is processed to obtain a brain age prediction result corresponding to the diffusion tensor image to be processed.

[0059] In the above scheme, the brain fiber bundle template is used to segment the multiple diffusion index images corresponding to the diffusion tensor image to be processed to obtain image features corresponding to each brain region. This image feature can accurately describe the white matter information corresponding to each brain region. The pre-trained brain age prediction model is then used to process the image features corresponding to each brain region to obtain a brain age prediction result corresponding to the diffusion tensor image to be processed, thus achieving brain age prediction. In addition, the brain age prediction model also incorporates more image feature information corresponding to each brain region in the multiple diffusion index images to more accurately obtain the white matter information corresponding to each brain region. Based on the white matter information corresponding to each brain region, brain age is accurately predicted, improving the accuracy of brain age prediction results and assisting doctors in completing brain aging examinations. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work. In the drawings:

[0061] Figure 1 A schematic flow chart of a brain age prediction method based on region segmentation provided by an embodiment of the present invention;

[0062] Figure 2 A schematic diagram of a process for obtaining a brain age prediction result corresponding to a target subject using a brain age prediction model provided by an embodiment of the present invention;

[0063] Figure 3 A flowchart of a brain age prediction model training method provided by an embodiment of the present invention;

[0064] Figure 4 A schematic diagram of a Transformer encoder structure provided by an embodiment of the present invention;

[0065] Figure 5 A schematic structural diagram of a brain age prediction device provided by an embodiment of the present invention;

[0066] Figure 6 The embodiment of the present invention provides Figure 5 A schematic diagram of the structure of the corresponding electronic equipment;

[0067] Figure 7 A schematic diagram of the structure of a brain age prediction model training device provided by an embodiment of the present invention;

[0068] Figure 8 The embodiment of the present invention provides Figure 7 Schematic diagram of the structure of the corresponding electronic equipment. DETAILED DESCRIPTION

[0069] Before introducing the technical solutions provided by the embodiments of the present invention, a brief introduction to the technical terms involved in this document is first given.

[0070] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0071] The terms used in the embodiments of the present invention are for the purpose of describing specific embodiments only and are not intended to limit the present invention. The singular forms "a," "the," and "the" used in the embodiments of the present invention and the appended claims are also intended to include plural forms, unless the context clearly indicates otherwise. "A plurality" generally includes at least two, but does not exclude the inclusion of at least one.

[0072] It should be understood that the term "and / or" as used herein is merely a description of the relationship between associated objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A exists alone, A and B exist simultaneously, or B exists alone. Furthermore, the character " / " in this document generally indicates that the associated objects are in an "or" relationship.

[0073] It should also be noted that the terms "include," "comprises," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a product or system comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such product or system. In the absence of further limitations, an element defined by the phrase "comprises a..." does not exclude the presence of other identical elements in the product or system comprising the element.

[0074] First, it is necessary to explain the background of the present invention. Aging affects the human body in many ways, with brain aging being a particularly important and critical factor. It is well known that neurons in the brain have difficulty replicating and regenerating in large quantities, making brain damage generally difficult to repair. Because aging often leads to permanent brain damage, memory and computational abilities often decline with age, and people may even develop degenerative diseases of the central nervous system, posing a significant threat to human health.

[0075] Brain age can be used as a biomarker to quantitatively assess the extent of brain aging, providing a quantitative analytical basis for assisting in the diagnosis and progression assessment of neurological disorders. Brain age prediction (BAP) involves extracting and analyzing biological features associated with brain aging to predict brain age. On the one hand, brain age prediction can provide early warning of potential brain disorders. If a subject's predicted brain age is greater than their actual biological age, it indicates that the brain may have lesions that accelerate brain aging. Conversely, if the subject's predicted brain age is greater than their actual biological age, it indicates that the subject's brain may be developing slowly. This intuitive metric can provide researchers with a deeper understanding of the mechanisms linking brain aging and brain disorders. On the other hand, brain age prediction has guiding significance for clinical diagnosis and treatment, and can be used to assess the risk of brain disease. In clinical practice, brain age prediction can assist doctors in identifying the causes of a patient's brain abnormalities and provide timely intervention and treatment.

[0076] Existing brain age prediction methods based on regional segmentation are mostly based on structural magnetic resonance imaging (MRI). The morphological features extracted are relatively limited and fail to capture neural information. Commonly used features for brain age prediction include cerebral cortical surface area, cortical thickness, gray matter volume, and white matter volume. Algorithms for predicting brain age based on brain MRI are generally divided into two categories: traditional machine learning methods and deep learning methods. Traditional machine learning-based brain age prediction relies on complex processes such as tissue segmentation and feature selection. The prediction results are susceptible to artifacts and suffer from poor generalization and low accuracy. Compared with traditional machine learning algorithms, deep learning can automatically and effectively capture the relationship between brain imaging features and brain age. However, due to the traditional architecture of most current convolutional neural networks, the accuracy and efficiency of feature extraction need to be improved.

[0077] Therefore, in order to solve the above technical problems, the present invention proposes a new brain age prediction method based on regional segmentation. By processing the to-be-processed diffusion tensor image corresponding to the target measurement subject to obtain multiple diffusion index images reflecting brain white matter information, and processing the multiple diffusion index images according to the brain fiber bundle template to obtain feature vectors reflecting the brain white matter information of each brain region, and using a pre-trained brain age prediction model, the feature vectors corresponding to each brain region are processed to obtain the brain age prediction result corresponding to the target measurement subject. That is, the brain age corresponding to the target measurement subject is accurately predicted based on the brain white matter information corresponding to each brain region. In addition, the brain prediction model can better capture the local and global features corresponding to each brain region to more accurately determine the degree of brain aging of the target measurement subject, and then predict the brain age corresponding to the target measurement subject, thereby improving the accuracy of the brain age prediction result.

[0078] The technical solutions provided in the embodiments of the present invention may be implemented by a single device or multiple devices. These devices may include, but are not limited to, devices integrated into any terminal device, such as a smartphone, tablet computer, PDA (Personal Digital Assistant), smart TV, laptop computer, desktop computer, smart wearable device, or medical device.

[0079] The specific implementation of the technical solution is introduced below in conjunction with specific embodiments.

[0080] Figure 1 The following is a flow chart of a method for predicting brain age based on regional segmentation provided by an embodiment of the present invention. The method may include the following steps:

[0081] 101. Obtain a diffusion tensor image to be processed corresponding to the target measurement person.

[0082] 102. Determine a plurality of diffusion index data corresponding to each voxel in the diffusion tensor image to be processed, and generate a plurality of diffusion index images based on the plurality of diffusion index data.

[0083] 103. Based on the brain fiber bundle template, multiple diffusion index images are processed to obtain first eigenvectors corresponding to each brain region. The eigenvectors are used to describe image features corresponding to each brain region.

[0084] 104. Using the pre-trained brain age prediction model, the first eigenvector corresponding to each brain region is processed to obtain the brain age prediction result corresponding to the target measurement subject.

[0085] Among them, brain aging is difficult to judge quantitatively. Usually, in order to describe a person's brain aging, it is often said that his brain volume is shrinking, white matter integrity is damaged, or functional activity is declining. Studies have shown that aging changes in brain white matter occur earlier than changes in the structure of the cerebral cortex. Even for the aging process of the hippocampus, white matter damage in the hippocampus-cingulate gyrus can lead to hippocampal atrophy. Therefore, for brain aging, white matter may be more sensitive and change earlier. Then, the real age of the brain can be more accurately predicted by the white matter information of the target measurement.

[0086] In practical applications, magnetic resonance technology is an important technical means for non-invasive evaluation of brain structure and function. Magnetic resonance diffusion tensor imaging (DTI) technology is an imaging and post-processing technology developed in recent years based on DWI. It is the development and deepening of DWI. It is currently the only non-invasive examination method that can effectively observe and track brain white matter fiber bundles. In addition to being able to clearly display white matter fiber bundles, it can also assess brain development levels and brain cognitive functions, and reveal pathological changes in brain diseases. Therefore, in an embodiment of the present invention, when predicting the brain age of a target person being measured, the diffusion tensor image (DTI image) of the target person being measured is analyzed and processed to obtain the white matter information corresponding to the target person being measured, and then based on the white matter information, the brain age corresponding to the target person being measured is further predicted.

[0087] Specifically, first, a diffusion tensor image to be processed corresponding to the target measurement subject is obtained, then a plurality of diffusion index data corresponding to each voxel in the diffusion tensor image to be processed is determined, and a plurality of diffusion index images are generated based on the plurality of diffusion index data.

[0088] Furthermore, during implementation, the acquired diffusion tensor images to be processed often contain noise, and head movement may occur when acquiring the target subject's diffusion tensor images. To ensure more accurate prediction results, the acquired diffusion tensor images can be preprocessed after acquisition. This preprocessing includes scalp removal, head motion correction, and noise reduction. Subsequently, the preprocessed diffusion tensor images are processed to generate multiple corresponding diffusion index images.

[0089] Specifically, a DTI model is fitted voxel by voxel based on the preprocessed diffusion tensor image. Multiple diffusion index data are then calculated for each voxel in the diffusion tensor image based on the fitted DTI model. Based on the multiple diffusion index data, a diffusion image corresponding to multiple diffusion index parameters is obtained. The diffusion index data include mean diffusion coefficient (MD), axial diffusion coefficient (AD), radial diffusion coefficient (RD), and diffusion anisotropy (FA).

[0090] The diffusion tensor image includes a diffusion weighted image. In an optional embodiment, a plurality of diffusion index data corresponding to each voxel in the diffusion tensor image to be processed is determined, and a plurality of diffusion index images are generated based on the plurality of diffusion index data. The specific implementation method may be: determining the diffusion tensor based on the diffusion weighted image; determining the diffusion eigenvalue based on the diffusion tensor; determining the diffusion index data based on the diffusion eigenvalue; and generating the diffusion index image based on the diffusion index data.

[0091] After obtaining multiple diffusion index images corresponding to the target person being measured, the multiple diffusion index images are processed according to the brain fiber bundle template to obtain the first eigenvector corresponding to each brain region. The eigenvector is used to describe the image features corresponding to each brain region. Among them, the brain fiber bundle template can be a unified white matter segmentation template, a white matter fiber bundle anatomical template, etc. In addition, different brain regions have different effects on brain aging. Therefore, when predicting brain age, in order to improve the accuracy of the prediction results, the diffusion index image is segmented to obtain the various brain regions contained in the diffusion index image, and then the image features corresponding to each brain region are determined respectively. Based on the image features corresponding to each brain region, the brain age result corresponding to the target person being measured is determined.

[0092] Specifically, based on the brain fiber bundle template, the target diffusion index image is segmented to obtain a segmented image corresponding to the target diffusion index image, where the target diffusion index image is any one of multiple diffusion index images, and the segmented image includes multiple brain regions; voxel extraction is performed on the multiple brain regions to obtain first sub-eigenvectors corresponding to each of the multiple brain regions in the target diffusion index image; the first sub-eigenvectors corresponding to each brain region in the multiple diffusion index images are spliced to obtain the first eigenvector corresponding to each brain region.

[0093] For example, based on the JHU-ICBM label atlas, data from 48 brain regions, including the cerebellar pedicles and pons chiasmata, were extracted from the AD, FA, MD, and RD diffusion index images to obtain feature vectors corresponding to each brain region in multiple diffusion index images. The four vectors extracted from the same brain region (based on AD, FA, MD, and RD data, respectively) were then horizontally spliced to obtain the image features corresponding to the target subject. In other words, each brain region can be represented by a feature vector, and the white matter information corresponding to each target subject can be represented by 48 one-dimensional feature vectors of varying lengths.

[0094] The voxels corresponding to each brain region marked in the white matter fiber bundle template can be first extracted separately and the voxels corresponding to each brain region can be spliced together to obtain a one-dimensional vector. Based on this one-dimensional voxel, the feature vector corresponding to each brain region is extracted from the AD, FA, MD, and RD images respectively. Then, the four feature vectors corresponding to the same brain region are concatenated in the order of AD, FA, MD, and RD to synthesize a new one-dimensional vector. That is, the first feature vector corresponding to each brain region contains the four local diffusion index data corresponding to each brain region.

[0095] Finally, the first eigenvectors corresponding to each brain region are processed using the pre-trained brain age prediction model to obtain the brain age prediction result for the target subject. Alternatively, the first eigenvectors corresponding to each brain region are concatenated and then input into the pre-trained brain age prediction model to obtain the brain age prediction result for the target subject.

[0096] Among them, the brain age prediction model is trained to determine the brain age prediction result corresponding to the target measurer. The brain age prediction model can be a Transformer deep learning model, which may include at least one cascaded encoder and decoder, and each encoder includes a multi-head attention layer and a feedforward neural network layer.

[0097] In addition, for diffusion index images, each brain region has more image feature information. Therefore, when using a pre-trained brain age prediction model to process the first eigenvector representing the image features corresponding to each brain region, the image features can be first extracted to obtain the main feature information corresponding to each brain region, and then the main feature information corresponding to each brain region can be processed using the brain prediction model to obtain the brain age prediction result corresponding to the target measurer.

[0098] Optionally, using a pre-trained brain age prediction model, the first eigenvector corresponding to each brain region is processed to obtain the brain age prediction result corresponding to the target measurement person. The specific implementation process may include: performing feature extraction on the first eigenvector corresponding to each brain region to obtain the second eigenvector corresponding to each brain region; inputting the second eigenvector corresponding to each brain region into the pre-trained brain age prediction model to obtain the brain age prediction result corresponding to the target measurement person.

[0099] Among them, the second eigenvector corresponding to each brain region can reflect the white matter information corresponding to each brain region from multiple aspects. In subsequent processing, the aging degree of each brain region can be determined based on the white matter information corresponding to each brain region, and then the brain age prediction result can be determined based on the aging degree of each brain region.

[0100] In addition, an image feature extraction network can be used to perform feature extraction processing on the first feature vectors corresponding to each brain region to obtain the second feature vectors corresponding to each brain region, thereby ensuring the accuracy of the extracted feature information. In an optional embodiment, multiple image feature extraction networks are used to perform feature extraction processing on the first feature vectors corresponding to each brain region to obtain the depth features corresponding to each brain region. Specifically, the first feature vectors corresponding to each brain region are input into the image feature extraction networks corresponding to each brain region to obtain the second feature vectors corresponding to each brain region.

[0101] In other words, each brain region corresponds to a trained image feature extraction network. When performing feature extraction processing, the image feature extraction networks corresponding to multiple brain regions are used for feature extraction processing. For example, assuming there are 48 corresponding brain regions, 48 independent image feature extraction networks can be used for feature extraction processing. This can make the extracted image feature information, namely the second eigenvector, have richer information content and more abstract high-order feature representation, which can better reflect the essence of the data. Among them, the second eigenvector can include texture features, shape features, spatial features, etc. corresponding to each brain region.

[0102] Alternatively, multiple independent fully connected networks can be used as image feature extraction networks corresponding to each brain region. The image feature extraction network includes a linear layer, a Tanh activation function, and a dropout random inactivation layer. The input dimension of the linear layer is the length corresponding to each first eigenvector, and the output dimension is 1024. The activation function Tanh is calculated as follows: Where e is the base of the natural logarithm and x is the first eigenvector. Finally, the second eigenvectors corresponding to the 48 brain regions extracted from the multiple first eigenvectors corresponding to the target subject can be concatenated together to form a feature matrix of size (48, 1024).

[0103] After obtaining the second eigenvectors corresponding to each brain region, these are input into a pre-trained brain age prediction model to obtain a brain age prediction result for the target subject. Alternatively, a feature matrix obtained by concatenating multiple second eigenvectors is input into the brain age prediction model to obtain a brain age prediction result for the target subject. The specific implementation method can be set according to actual needs and is not limited in this embodiment.

[0104] The brain age prediction model analyzes and processes the second eigenvectors corresponding to each brain region to determine the corresponding degree of aging in each brain region. Based on this degree of aging, the predicted brain age of the target individual is then determined. In other words, the brain age prediction model learns the mapping between different eigenvector data and the degree of white matter aging, and between this degree of white matter aging and the preset brain age. This allows for more accurate determination of the target individual's brain age when using the pre-trained brain age prediction model.

[0105] In summary, in the brain age prediction scheme provided by the embodiment of the present invention, the brain fiber bundle template is used to segment the multiple diffusion index images corresponding to the diffusion tensor image to be processed to obtain image features corresponding to each brain region. The image features can accurately describe the white matter information corresponding to each brain region, and the pre-trained brain age prediction model is used to process the image features corresponding to each brain region to obtain a brain age prediction result corresponding to the diffusion tensor image to be processed, thereby achieving brain age prediction. In addition, the brain prediction model can also integrate more image feature information corresponding to each brain region in the multiple diffusion index images to more accurately obtain the white matter information corresponding to each brain region, and accurately predict brain age based on the white matter information corresponding to each brain region, thereby improving the accuracy of the brain age prediction results and assisting doctors in completing brain aging examinations.

[0106] The above embodiment introduces the specific implementation process of predicting the brain age of the target subject. To facilitate understanding of the specific implementation process of analyzing and processing the second eigenvectors corresponding to each brain region using a pre-trained brain age prediction model in the above embodiment, the processing process of the brain age prediction model is exemplified in conjunction with the following embodiment.

[0107] Figure 2 A flowchart of a method for obtaining a brain age prediction result corresponding to a target subject by using a brain age prediction model is provided in an embodiment of the present invention; based on the above embodiment, further reference is made to the attached Figure 2 As shown, this embodiment provides a method for using a pre-trained brain age prediction model to analyze and process the second eigenvectors corresponding to each brain region to obtain a brain age prediction result corresponding to the target measurement subject. The specific method may include the following steps:

[0108] 201. Perform splicing processing on the second eigenvectors corresponding to the respective brain regions to obtain a spliced first eigenmatrix.

[0109] 202. Perform position encoding processing on the first feature matrix to obtain a second feature matrix, where the second feature matrix is used to describe brain region position information corresponding to each feature vector.

[0110] 203. Input the second feature matrix into the brain age prediction model to obtain a brain age prediction result.

[0111] In order to enable the brain age prediction model to better distinguish the eigenvectors (second eigenvectors) of different brain regions and help the brain age prediction model perform differential processing on the second eigenvectors corresponding to each brain region, each second eigenvector can be subjected to brain region encoding processing. Specifically, the second eigenvectors corresponding to each brain region are spliced to obtain a spliced first eigenmatrix, and the first eigenmatrix is positionally encoded to obtain a second eigenmatrix. The second eigenmatrix is used to describe the brain region position information corresponding to each eigenvector.

[0112] Among them, the spatial position information corresponding to each brain region can be obtained, the spatial position information can be encoded to obtain a position vector, and then processed with the first feature matrix, and the position information corresponding to each brain region can be added to each second feature vector in the first feature matrix, so that each second feature vector carries not only the depth features corresponding to each brain region but also the position information corresponding to each brain region.

[0113] In addition, in order to more accurately predict the brain age of each target measurement subject, in specific implementation, each diffusion index image obtained can be divided into 48 brain regions to more accurately obtain the white matter information corresponding to each region in the image, thereby improving the accuracy of the brain age prediction results. Then, when performing position encoding processing on the second eigenvectors corresponding to each of the 48 brain regions, in order to facilitate encoding of each brain region and avoid encoding errors, the second eigenvectors corresponding to the 48 brain regions can be first spliced together in a certain order to obtain the spliced first eigenmatrix. The first eigenmatrix is then position-encoded so that the second eigenvectors corresponding to each brain region carry the corresponding position information, helping the brain prediction model to better understand the position of each brain region, perform differential processing on the second eigenvectors of each brain region, and obtain the dependency relationship between each brain region, so that the dependency relationship between each brain region can be better combined in subsequent processing to determine the brain age prediction result corresponding to each target measurement subject.

[0114] From the above description, it can be seen that the second eigenvector contains the local diffusion index data corresponding to each brain region, that is, the second eigenvector is used to describe the eigenvector of the information of each brain region. Therefore, in order to better integrate the information of each brain region, before using the brain prediction model to analyze and process the second eigenvector, it is also possible to first obtain the eigenvector describing the overall information of the whole brain, and then analyze and process the second eigenvector of the information of each brain region and the eigenvector describing the overall information of the whole brain to obtain the brain age prediction result corresponding to the target measurement. The brain age prediction result obtained in this way is more accurate and can more realistically feedback the actual brain age status of the current target measurement.

[0115] Specifically, in an optional embodiment, the implementation method of performing position encoding processing on the first feature matrix to obtain the second feature matrix may include: inputting the first feature matrix into a global information extraction encoder to obtain a third feature vector corresponding to the whole brain, and the third feature vector is used to characterize the overall information of the whole brain; splicing the first feature matrix and the third feature vector to obtain a third feature matrix; and inputting the third feature matrix into a position encoder to obtain a second feature matrix.

[0116] Among them, the global information extraction encoder is trained to extract the overall information of the whole brain. The global information extraction encoder includes an attention layer and a feedforward neural network layer. The second eigenvectors of each brain region (the deep feature information of each brain region) are fused through the attention layer to obtain the third eigenvector used to describe the overall information of the whole brain.

[0117] After obtaining the third eigenvector corresponding to the whole brain, the third eigenvector can be concatenated with the first eigenmatrix to obtain a third eigenmatrix. The third eigenvector has the same length as the second eigenvector, and a new row can be added to the end of the first eigenmatrix. The third eigenvector can be concatenated with the newly added row of the first eigenmatrix to obtain the concatenated third eigenvector.

[0118] For example, assuming the length of each second eigenvector is 1024, the length of the third eigenvector is 1024, and the diffusion index image is segmented into 48 brain regions, the first eigenmatrix is a eigenmatrix with a shape of (48, 1024). A third eigenvector with a length of 1024 is concatenated to the first eigenmatrix to obtain a third eigenmatrix with a shape of (49, 1024).

[0119] Next, the third feature matrix is input into the position encoder to obtain a second feature matrix. The position encoder can be used to perform position encoding processing on the third feature matrix to obtain a second feature matrix carrying position information. The position encoder is trained to perform position encoding processing on each eigenvector in the feature matrix to embed the encoding information of each brain region into the eigenvector corresponding to each brain region, so that when the brain age prediction model is used for analysis and processing, differential analysis of each brain region can be carried out in a targeted manner, and the dependency relationship between each brain region can be better captured.

[0120] The position encoder can be implemented by a three-layer fully connected neural network, which includes an input layer, a hidden layer, and an output layer. The position encoder is used to encode each row in the third feature matrix to obtain an encoded second feature matrix.

[0121] In an optional embodiment, the specific implementation process of inputting the third characteristic matrix into the position encoder to obtain the second characteristic matrix may be: obtaining an encoding matrix corresponding to the third characteristic matrix, concatenating the third characteristic matrix and the encoding matrix to obtain a concatenated matrix, and encoding each row of the concatenated matrix using the position encoder to obtain an encoded second characteristic matrix. The encoding matrix includes position information corresponding to each brain region and position information corresponding to overall brain information.

[0122] Continuing with the above example, the third feature matrix is a feature matrix with a shape of (49, 1024), and the encoding matrix is an encoding matrix with a shape of (49, 1024). The third feature matrix and the encoding matrix are concatenated to obtain a concatenated matrix with a shape of (49, 2048). A position encoder or encoding layer is then used to encode each row of the concatenated matrix. The position information in the encoding matrix is embedded into each eigenvector in the third feature matrix to obtain an encoded second feature matrix with a shape of (49, 1024). The encoded feature matrix obtained in this way not only contains the layout diffusion index data corresponding to each brain region (the second eigenvector corresponding to each brain region) and the global diffusion index data (the third eigenvector corresponding to the entire brain), but also contains the spatial position information corresponding to each diffusion index data.

[0123] From the above description, it can be seen that: by embedding the position coding information of each brain region into the second eigenvector of each brain region, the spatial position information is integrated into each second eigenvector, so that when the second eigenvector of each brain region interacts with the second eigenvectors of other brain regions in the self-attention layer of the brain prediction model, the dependency between different brain regions can be captured. This helps the brain age prediction model to capture local features within the brain region while paying attention to the overall structure of the whole brain, thereby improving the regression performance.

[0124] After obtaining the encoded second feature matrix, the second feature matrix is input into the brain age prediction model to obtain a brain age prediction result. The second feature matrix is analyzed and processed using the brain age prediction model to obtain a brain age prediction result.

[0125] In an optional embodiment, the brain age prediction model includes a feature fusion encoder and a decoder. Inputting the second feature matrix into the brain age prediction model to obtain a brain age prediction result may include: inputting the second feature matrix into the feature fusion encoder to obtain a fused target feature vector; and inputting the target feature vector into the decoder to obtain a brain age prediction result corresponding to the target feature vector. The feature fusion encoder includes at least one cascaded encoder, each of which includes an attention layer and a feedforward neural network layer.

[0126] After the second feature matrix passes through the multi-layer encoder, the feature vectors corresponding to each brain region in the second feature matrix can be fused to calculate the correlation between the brain regions, so as to obtain the target feature vector used to characterize all image information corresponding to the entire diffusion tensor image to be processed.

[0127] Finally, the target feature vector is input into the decoder of the brain age prediction model to obtain the brain age prediction result corresponding to the target feature vector. The decoder is used to determine the brain age prediction result corresponding to the target feature vector. The decoder includes an attention layer and a feedforward neural network layer.

[0128] In an optional embodiment, a specific implementation method of inputting the target feature vector into a decoder to obtain a brain age prediction result corresponding to the target feature vector may include: obtaining the real age of the target measurer; determining reference feature vectors corresponding to multiple brain regions corresponding to the real age, the reference feature vectors being used to characterize the aging degree of each brain region corresponding to multiple measurers of age; splicing the reference feature vectors corresponding to multiple brain regions to obtain a spliced fourth feature matrix; splicing the fourth feature matrix and the target feature vector to obtain a spliced fifth feature matrix; and inputting the fifth feature matrix into the decoder to obtain a brain age prediction result.

[0129] Among them, the target feature vector is mainly used to represent the feature information of each brain region corresponding to the target measurer as an individual and the feature information of the whole brain at the overall level. Therefore, in order to obtain more accurate brain age prediction results, when making brain age predictions, it is also possible to combine the group-level information corresponding to each brain region of multiple measurers of the same age as the target measurer, and use the reference feature vectors of multiple brain regions corresponding to multiple measurers as reference standards to further determine the brain age prediction results corresponding to the target measurer.

[0130] In addition, in order to more intuitively understand the degree of white matter aging in each brain region corresponding to the target measurement subject, the pre-trained brain prediction model can not only output the brain age prediction results, but also output the brain structure corresponding to the degree of white matter aging in each brain region. Based on the output brain structure, the degree of aging corresponding to each brain region can be determined intuitively.

[0131] In an embodiment of the present invention, position encoding processing is performed on the first eigenvector corresponding to the second eigenvector corresponding to each brain region to obtain a second eigenmatrix for describing the brain region position information corresponding to each eigenvector. When the second eigenmatrix containing the position information corresponding to each brain region is analyzed and processed by the brain age prediction model, the eigenvectors of different brain regions can be distinguished, which helps the brain age prediction model to perform differential processing on the eigenvectors of each brain region. Moreover, when the eigenvector of each brain region interacts with the eigenvectors of other brain regions, the correlation between different brain regions can be captured, so that the brain age prediction model not only pays attention to the local characteristics of each brain region, but also manages the global characteristics at the overall level of the whole brain. In this way, the brain age prediction result is more accurate and can more realistically reflect the corresponding brain age of the target measurer.

[0132] After introducing the network models that may be used in the present invention and their usage, in order to facilitate understanding of the working principles of the above-mentioned network models, the specific process of training the brain age prediction model is exemplified below.

[0133] Figure 3 A flowchart of a brain age prediction model training method provided by an embodiment of the present invention. Specifically, the method includes the following steps:

[0134] 301. Collect training samples, where the training samples include diffusion tensor images corresponding to multiple measurers.

[0135] 302. Determine a plurality of diffusion index data corresponding to each voxel in the diffusion tensor image, and generate a plurality of diffusion index image samples based on the plurality of diffusion index data.

[0136] 303. Determine a reference brain age corresponding to the plurality of diffusion index image samples.

[0137] 304. Based on the brain fiber bundle template, multiple diffusion index image samples are processed to obtain feature vectors corresponding to each brain region. The feature vectors are used to describe the image features corresponding to each brain region.

[0138] 305. Through the brain age prediction model, the feature vectors corresponding to each brain region are processed to obtain the brain age prediction results.

[0139] 306. Train a brain age prediction model based on the reference brain age and brain age prediction results.

[0140] In the pair Figure 1To train the brain age prediction model used in the illustrated embodiment, a large number of training samples are first collected. These training samples include diffusion tensor images corresponding to multiple subjects. To prevent variations in imaging quality across different scanning devices and differences in data preprocessing from influencing the prediction results of the brain age prediction model, a standardized data processing process is used to preprocess the diffusion tensor image samples used for modeling and testing, starting with the DICOM data corresponding to the acquired diffusion tensor images.

[0141] Specifically, multiple DICOM files acquired from the same acquisition were merged and converted to .nii format. The converted files include a .nii file containing image data and head information, a .bval file containing the b-value information used for the acquired images, a .bvec file containing the gradient directions used for the acquired images, and a .json file containing acquisition-related metadata. The voxel size of the image data in the .nii file was resampled to 1 mm × 1 mm × 1 mm. All resampled diffusion tensor images were subjected to head motion correction, using the b-value 0 image as the standard, to obtain motion-corrected diffusion tensor images. The b-value 0 images were segmented into head and background using the median-based maximum interclass contrast method, with all segmentation parameters set to default values. This step ultimately yielded a head mask for the b-value 0 diffusion tensor image, which was used as the head mask for all motion-corrected diffusion tensor images. Using the Probabilistic Identification and Estimation of Noise (PIESNO) method, we estimate the noise of the DTI data after head motion removal, taking into account the number of coils used in the scan. This noise sigma value is then calculated. Using the estimated sigma value and the head mask obtained through head segmentation, we perform noise reduction on the head region to obtain a denoised DTI image.

[0142] Then, a plurality of diffusion index data corresponding to each voxel in the diffusion tensor image corresponding to each of the plurality of measurers is determined, and based on the diffusion index data, a plurality of diffusion index image samples corresponding to each of the plurality of measurers are generated, and a reference brain age corresponding to the plurality of diffusion index image samples is determined.

[0143] Specifically, the DTI model is fitted voxel by voxel according to the denoised diffusion tensor image, and then the mean diffusion coefficient (MD), axial diffusion coefficient (AD), radial diffusion coefficient (RD) and diffusion anisotropy (FA) of each voxel are calculated based on the fitted DTI model. Based on the diffusion index data, the diffusion index image samples corresponding to the four diffusion index data are obtained.

[0144] Next, multiple diffusion index image samples are processed according to the brain fiber bundle template to obtain feature vectors corresponding to each brain region. The feature vectors are used to describe the image features corresponding to each brain region.

[0145] Specifically, in an optional embodiment, for each sample, according to the JHU-ICBM label atlas, data of 48 brain regions such as the cerebellar pedicles, pons and chiasma are extracted from the acquired AD, FA, MD and RD diffusion index image samples, and the four vectors extracted from the same brain region (the four vectors are obtained based on AD, FA, MD and RD data) are horizontally spliced (connected in sequence in the order of AD, FA, MD, RD to synthesize a new one-dimensional vector). Then, each brain region can be represented by a vector, and each sample can be represented by 48 one-dimensional vectors of different lengths. The 48 one-dimensional vector representations of different lengths are determined as the feature vectors corresponding to each of the 48 brain regions.

[0146] The brain age prediction model then processes the feature vectors corresponding to each brain region to obtain a brain age prediction result. Because the feature vectors corresponding to each brain region contain various image information corresponding to that region, the accuracy of the prediction result may be affected during the analysis and processing of the feature vectors. Therefore, in an optional embodiment, feature extraction can be performed on the feature vectors corresponding to each brain region before analysis and processing of the extracted feature vectors.

[0147] Continuing with the above example, specifically, the feature vectors corresponding to each brain region are input into a feature extraction network to extract deep features corresponding to each brain region. Deep features provide richer, more abstract, and higher-order feature representations that better reflect the essence of the data. Deep features include texture features, shape features, spatial features, and more.

[0148] Among them, 48 independent fully connected networks are used as feature extraction networks corresponding to each brain region, that is, each brain region corresponds to a specific feature extraction network, and the feature vectors corresponding to each brain region obtained in the above steps are used as input, and the deep features corresponding to each brain region are used as output. In addition, the feature extraction network consists of a linear layer, a Tanh activation function, and a dropout (random inactivation) layer. The input dimension of the linear layer matches the length of each vector of the sample, and the output dimension is 1024. The deep features of the 48 brain regions extracted from a single sample are spliced together to form a feature matrix of size (48, 1024).

[0149] A trainable parameter vector of length 1024 is concatenated with the feature matrix, and the shape of the feature matrix becomes (49, 1024). This parameter vector is a feature vector used to describe the overall level of information of the whole brain. In the subsequent Transformer deep learning model (multiple encoders), the deep features of 48 brain regions can be fused through the attention mechanism to obtain a feature vector that describes the overall level of information of the whole brain.

[0150] In order to enable the brain age prediction model to distinguish the eigenvectors of different brain regions and help the brain age prediction model to perform differential processing on the eigenvectors of each brain region, it is necessary to perform brain region encoding on each eigenvector. Among them, the specific implementation method of brain region encoding is as follows: first, a trainable (49, 1024) encoding matrix is spliced on the feature matrix to obtain a splicing matrix with a shape of (49, 2048); then, an embedding layer is used to encode each row of the splicing matrix to obtain a coding feature matrix with a shape of (49, 1024). Among them, the encoding matrix refers to the encoding matrix containing the spatial position information corresponding to each brain region, and the embedding layer is implemented by a three-layer fully connected neural network.

[0151] By embedding the code of each brain region into the feature vector of that brain region, the dependency between different brain regions can be captured when the feature vector of each brain region interacts with the feature vectors of other brain regions in the self-attention mechanism. This helps the model to capture local features within the brain region while paying attention to the overall structure of the whole brain, thereby improving the regression performance.

[0152] A brain age prediction model was trained based on the reference brain age and brain age prediction results. To achieve brain age prediction and explore the amount of information provided by each brain region in this prediction, a Transformer deep learning model based on brain region segmentation imaging data was constructed and trained to obtain a trained brain age prediction model. The brain age prediction model consists of a stack of several Transformer encoders and Transformer decoders. Each encoder and decoder includes a self-attention layer and a feedforward neural network layer. The attention mechanism in the self-attention layer captures the relationship between different brain regions, achieving efficient image information encoding and deep feature extraction.

[0153] Specifically, the encoded feature matrix obtained above is input into a network structure composed of several stacked Transformer encoders. The attention mechanism captures the relationship between various brain regions, achieving efficient image information encoding and deep feature extraction to obtain the target feature vector. Each Transformer Encoder consists of a multi-head self-attention structure and a feedforward neural network.

[0154] The multi-head self-attention structure is as follows: the encoding feature matrix is multiplied by three different trainable matrices of size (1024, 1024) to obtain matrices Q, K, and V. Then Q, K, and V are divided into 16 matrices of size (49, 64) respectively (i = 1, 2, ..., 16), and the attention sub-matrix is calculated according to the following formula:

[0155]

[0156] Among them, d k is the vector length, i.e., 64. The 16 attention matrices are then concatenated horizontally to obtain an attention matrix of size (49, 1024). This result is then added to the matrix of the input self-attention structure. This allows information from the previous layer to be passed to the next layer without error, solving the problem of difficult neural network training.

[0157] The structure of the feedforward neural network is a linear layer (input size is 1024, output size is 2048), activation function Tanh, dropout (p = 0.9), linear layer (input size is 2048, output size is 1024), and dropout (p = 0.9). Finally, the output is added to the structure of the feedforward neural network and normalized to accelerate the convergence of the brain age prediction model. For example Figure 4 The Transformer encoder structure shown.

[0158] After passing through the multi-layer Transformer encoder network structure, the target feature vector has obtained the information of the entire image. Normalizing the target feature vector layer by layer can improve the stability of deep learning network training. Then, the normalized target feature vector is input into the Transformer decoder of the network structure used to predict brain age, predict brain age and calculate the loss function. Among them, the Transformer decoder network structure is: linear layer (input size is 1024, output size is 32), activation function Tanh, dropout layer, linear layer (input size is 32, output size is 1) and Sigmoid. The Sigmoid calculation method is:

[0159] Where x is the target feature vector.

[0160] Finally, the brain age prediction model was trained based on the reference brain age and predicted brain age results. The training process used the Adam optimizer and the Mean Squared Error (MSE) loss function. The learning rate was set to 1e-6 and decreased by 30% per round, the batch size was set to 2, the number of epochs was set to 20, and the transformer encoder depth was set to 3. The training process continued until the loss converged or the Mean Squared Error (MAE) reached the expected level. This resulted in a trained brain age prediction model.

[0161] After building the brain age prediction model, we input diffusion tensor image samples into the model and extract the attention matrix output by the last attention layer in the Transformer encoder. This matrix represents the interactions between different brain regions and has a size of (48, 48). The element at (i, j) represents the degree of association between brain regions i and j. The resulting vector, averaged across the columns of this matrix, can be considered the degree of association between each brain region and the entire brain, indicating its contribution to the prediction.

[0162] It is worth noting that the brain age prediction model training method is similar to the implementation of the brain age prediction method based on regional segmentation provided in the above embodiment. The similarities are mentioned above and will not be repeated here.

[0163] Figure 5 This is a schematic diagram of the structure of a brain age prediction device provided by an embodiment of the present invention. Figure 5 As shown, the device includes: an acquisition module 11, a determination module 12, a first processing module 13, and a second processing module 14; wherein,

[0164] The acquisition module 11 is used to acquire the diffusion tensor image to be processed corresponding to the target person being measured.

[0165] The determination module 12 is configured to determine a plurality of diffusion index data corresponding to each voxel in the diffusion tensor image to be processed, and generate a plurality of diffusion index images according to the plurality of diffusion index data.

[0166] The first processing module 13 is configured to process the plurality of diffusion index images according to the brain fiber bundle template to obtain a first eigenvector corresponding to each brain region, wherein the eigenvector is used to describe image features corresponding to each brain region.

[0167] The second processing module 14 is configured to process the first eigenvectors corresponding to the respective brain regions using a pre-trained brain age prediction model to obtain a brain age prediction result corresponding to the target subject.

[0168] Optionally, the first processing module 13 is specifically used to: segment the target diffusion index image according to the brain fiber bundle template to obtain a segmented image corresponding to the target diffusion index image, where the target diffusion index image is any one of the multiple diffusion index images, and the segmented image includes multiple brain regions; perform voxel extraction on the multiple brain regions to obtain first sub-eigenvectors corresponding to each of the multiple brain regions in the target diffusion index image; and perform splicing on the first sub-eigenvectors corresponding to each brain region in the multiple diffusion index images to obtain a first eigenvector corresponding to each brain region.

[0169] Optionally, the second processing module 14 is specifically used to: perform feature extraction on the first eigenvectors corresponding to the respective brain regions to obtain second eigenvectors corresponding to the respective brain regions; input the second eigenvectors corresponding to the respective brain regions into a pre-trained brain age prediction model to obtain a brain age prediction result corresponding to the target measuree.

[0170] Optionally, the second processing module 14 is specifically configured to: input the first feature vectors corresponding to the respective brain regions into the image feature extraction networks corresponding to the respective brain regions, so as to obtain the second feature vectors corresponding to the respective brain regions.

[0171] Optionally, the second processing module 14 is specifically used to: perform splicing processing on the second eigenvectors corresponding to the various brain regions to obtain a spliced first eigenmatrix; perform position encoding processing on the first eigenmatrix to obtain a second eigenmatrix, and the second eigenmatrix is used to describe the brain region position information corresponding to each eigenvector; and input the second eigenmatrix into the brain age prediction model to obtain a brain age prediction result.

[0172] Optionally, the second processing module 14 is specifically used to: input the first feature matrix into the global information extraction encoder to obtain a third feature vector corresponding to the whole brain, and the third feature vector is used to characterize the overall information of the whole brain; splice the first feature matrix and the third feature vector to obtain a third feature matrix; input the third feature matrix into the position encoder to obtain a second feature matrix.

[0173] Optionally, the brain age prediction model includes a feature fusion encoder and a decoder; the second processing module 14 is specifically used to: input the second feature matrix into the feature fusion encoder to obtain a fused target feature vector; input the target feature vector into the decoder to obtain a brain age prediction result corresponding to the target feature vector.

[0174] Optionally, the second processing module 14 is specifically used to: obtain the real age of the target measurer; determine the reference eigenvectors corresponding to multiple brain regions corresponding to the real age, and the reference eigenvectors are used to characterize the aging degree of each brain region corresponding to multiple measurers of the age; splice the reference eigenvectors corresponding to the multiple brain regions to obtain a spliced fourth eigenmatrix; splice the fourth eigenmatrix and the target eigenvector to obtain a spliced fifth eigenmatrix; input the fifth eigenmatrix into the decoder to obtain a brain age prediction result.

[0175] Optionally, the feature fusion encoder includes at least one cascaded encoder, and each of the encoders includes an attention layer and a feedforward neural network layer.

[0176] Figure 5 The device shown can execute the steps introduced in the aforementioned embodiments. For detailed execution process and technical effects, please refer to the description in the aforementioned embodiments and will not be repeated here.

[0177] In one possible design, the above Figure 5 The structure of the brain age prediction device shown can be realized as an electronic device, such as Figure 6 As shown, the electronic device may include: a memory 21, a processor 22, and a communication interface 23. The memory 21 stores executable code, and when the executable code is executed by the processor 22, the processor 22 can at least implement the brain age prediction method based on region segmentation as provided in the aforementioned embodiment.

[0178] In addition, an embodiment of the present invention provides a non-transitory machine-readable storage medium, on which executable code is stored. When the executable code is executed by a processor of an electronic device, the processor can at least implement the brain age prediction method based on regional segmentation as provided in the aforementioned embodiment.

[0179] Figure 7 This is a schematic diagram of the structure of a brain age prediction model training device provided by an embodiment of the present invention. Figure 7 As shown, the device includes: an acquisition module 31, a generation module 32, a determination module 33, a first processing module 34, a second processing module 35, and a training module 36; wherein,

[0180] The acquisition module 31 is used to acquire training samples, wherein the training samples include diffusion tensor images corresponding to multiple measurers.

[0181] The generating module 32 is configured to determine a plurality of diffusion index data corresponding to each voxel in the diffusion tensor image, and generate a plurality of diffusion index image samples according to the plurality of diffusion index data.

[0182] The determination module 33 is configured to determine the reference brain age corresponding to the plurality of diffusion index image samples.

[0183] The first processing module 34 is configured to process the plurality of diffusion index image samples according to the brain fiber bundle template to obtain a feature vector corresponding to each brain region, where the feature vector is used to describe the image features corresponding to each brain region.

[0184] The second processing module 35 is configured to process the first eigenvectors corresponding to the respective brain regions using a brain age prediction model to obtain a brain age prediction result.

[0185] The training module 36 is used to train the brain age prediction model according to the reference brain age and the brain age prediction result.

[0186] Figure 7 The device shown can execute the steps introduced in the aforementioned embodiments. For detailed execution process and technical effects, please refer to the description in the aforementioned embodiments and will not be repeated here.

[0187] In one possible design, the above Figure 7 The structure of the brain age prediction model training device shown can be implemented as an electronic device, such as Figure 8 As shown, the electronic device may include: a memory 41, a processor 42, and a communication interface 43. The memory 41 stores executable code, and when the executable code is executed by the processor 42, the processor 42 can at least implement the brain age prediction model training method provided in the aforementioned embodiment.

[0188] In addition, an embodiment of the present invention provides a non-transitory machine-readable storage medium, on which executable code is stored. When the executable code is executed by a processor of an electronic device, the processor can at least implement the brain age prediction model training method provided in the aforementioned embodiment.

[0189] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0190] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, or of course, by hardware. Based on this understanding, the essence of the above technical solution or the part that contributes to the existing technology can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or certain parts of the embodiments.

[0191] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A brain age prediction method based on regional segmentation, characterized in that: include: Obtaining a diffusion tensor image to be processed corresponding to the target measurement person; determining a plurality of diffusion index data corresponding to each voxel in the diffusion tensor image to be processed, and generating a plurality of diffusion index images according to the plurality of diffusion index data; Processing the plurality of diffusion index images according to the brain fiber bundle template to obtain a first eigenvector corresponding to each brain region, wherein the eigenvector is used to describe the image features corresponding to each brain region; performing feature extraction on the first eigenvectors corresponding to the respective brain regions to obtain second eigenvectors corresponding to the respective brain regions; performing splicing processing on the second eigenvectors corresponding to the respective brain regions to obtain a spliced first eigenmatrix; Performing position encoding processing on the first feature matrix to obtain a second feature matrix, where the second feature matrix is used to describe brain region position information corresponding to each eigenvector; The second feature matrix is input into the brain age prediction model to obtain a brain age prediction result.

2. The method according to claim 1, characterized in that The processing of the plurality of diffusion index images according to the brain fiber bundle template to obtain a first eigenvector corresponding to each brain region includes: performing segmentation processing on a target diffusion index image according to the brain fiber bundle template to obtain a segmented image corresponding to the target diffusion index image, wherein the target diffusion index image is any one of the plurality of diffusion index images, and the segmented image includes a plurality of brain regions; performing voxel extraction processing on the multiple brain regions respectively to obtain first sub-feature vectors corresponding to the multiple brain regions in the target diffusion index image; The first sub-eigenvectors corresponding to the respective brain regions in the plurality of diffusion index images are spliced to obtain the first eigenvectors corresponding to the respective brain regions.

3. The method according to claim 1, characterized in that The step of extracting features from the first eigenvectors corresponding to the respective brain regions to obtain second eigenvectors corresponding to the respective brain regions includes: The first eigenvectors corresponding to the respective brain regions are respectively input into the image feature extraction networks corresponding to the respective brain regions to obtain the second eigenvectors corresponding to the respective brain regions.

4. The method according to claim 1, wherein The performing position encoding processing on the first feature matrix to obtain a second feature matrix includes: Inputting the first feature matrix into a global information extraction encoder to obtain a third feature vector corresponding to the whole brain, wherein the third feature vector is used to represent the overall information of the whole brain; Concatenate the first characteristic matrix and the third characteristic vector to obtain a third characteristic matrix; The third characteristic matrix is input to a position encoder to obtain a second characteristic matrix.

5. The method according to claim 4, characterized in that The brain age prediction model includes a feature fusion encoder and a decoder; Inputting the second feature matrix into the brain age prediction model to obtain a brain age prediction result includes: Inputting the second feature matrix into the feature fusion encoder to obtain a fused target feature vector; The target feature vector is input into the decoder to obtain a brain age prediction result corresponding to the target feature vector.

6. The method according to claim 5, characterized in that Inputting the target feature vector into the decoder to obtain a brain age prediction result corresponding to the target feature vector includes: Obtaining the real age of the target measurement person; Determining reference feature vectors corresponding to multiple brain regions corresponding to the true age, wherein the reference feature vectors are used to represent the aging degree of each brain region corresponding to multiple measurers of the age; performing splicing processing on the reference eigenvectors corresponding to the multiple brain regions to obtain a spliced fourth eigenmatrix; Concatenating the fourth characteristic matrix and the target characteristic vector to obtain a concatenated fifth characteristic matrix; The fifth feature matrix is input into the decoder to obtain a brain age prediction result.

7. The method according to claim 5, characterized in that The feature fusion encoder includes at least one cascaded encoder, and each of the encoders includes an attention layer and a feedforward neural network layer.

8. An electronic device, characterized in that: include: A memory, a processor, and a communication interface; wherein the memory stores executable code, and when the executable code is executed by the processor, the processor executes the brain age prediction method based on regional segmentation as described in any one of claims 1 to 7.

Images