Brain disease risk assessment method and device based on regional brain age prediction, electronic equipment and storage medium

Abstract

The present disclosure relates to a brain disease risk assessment method and device based on regional brain age prediction, an electronic device and a storage medium, comprising: dividing brain images of a plurality of healthy subjects to obtain a plurality of brain regions and a brain region mask of each brain region; training a teacher model based on the brain images of the plurality of healthy subjects until convergence, and freezing the teacher model parameters after training convergence; constructing a regional input according to the brain region mask of each brain region, inputting the regional input into the teacher model respectively to obtain a regional brain age, and generating a regional brain age label according to the regional brain age; taking the regional brain age label as a distillation supervision signal of a student model, training the student model based on the brain images of the plurality of healthy subjects; predicting the brain image of a subject to be predicted based on the student model to obtain a regional brain age, and evaluating the risk of a target brain disease according to a prediction result and a related brain region set of the target brain disease. Thus, the risk of the target brain disease can be evaluated in an interpretable manner.

Description

Technical Field

[0001] This disclosure relates to the fields of medical image processing and artificial intelligence technology, and in particular to a method and device, electronic device and storage medium for assessing brain disease risk based on regional brain age prediction. Background Technology

[0002] Neurodegenerative diseases such as Parkinson's disease (PD) and Alzheimer's disease (AD), as well as stroke and its sequelae of functional impairment, are the main causes of severe cognitive and motor impairment in the elderly. These diseases often have an insidious onset and a progressive course; clinically, slow and complex changes have already occurred in brain structure and function before typical symptoms appear.

[0003] Brain age prediction is a quantitative assessment tool for brain health that has emerged in recent years. It typically involves constructing a regression model that maps the brain magnetic resonance imaging (MRI) images of healthy controls to their chronological age. The difference between the predicted brain age and the actual age is used to characterize the degree of brain aging or developmental deviation in an individual. Numerous studies have shown that an increased brain age difference (i.e., predicted brain age minus chronological age) is closely associated with the risk of various brain diseases, cognitive decline, and poor prognosis. Therefore, brain age is considered a potential risk phenotype or imaging biomarker for brain diseases.

[0004] However, most existing brain age-related studies use whole-brain MRI as a uniform input, outputting a global brain age value. While this method can reflect overall aging trends at the population level, it still has many limitations. On the one hand, brain development and aging processes exhibit significant spatial heterogeneity, with different brain regions showing significant differences in maturation time, atrophy rate, and disease susceptibility. Diseases such as PD and AD often manifest as selective damage to specific brain regions or brain networks, making it difficult to locate lesion-related brain regions and characterize inter-regional differences using only a single whole-brain age. On the other hand, different diseases may present different "regional aging fingerprints," making it difficult to distinguish disease types or explain which specific brain regions drive increased risk simply by relying on whole-brain age and its differences. Summary of the Invention

[0005] In view of this, this disclosure proposes a brain disease risk assessment scheme based on regional brain age prediction.

[0006] According to one aspect of this disclosure, a method for assessing the risk of brain diseases based on regional brain age prediction is provided, comprising: dividing brain images of multiple healthy subjects into brain regions to obtain multiple brain regions and a brain region mask for each brain region; training a teacher model based on the brain images of the multiple healthy subjects until convergence, and freezing the parameters of the teacher model after training convergence; constructing a regional input for each brain region based on the brain region mask for each brain region, inputting the regional input for each brain region into the teacher model to obtain the regional brain age for each brain region, and generating a regional brain age label based on the regional brain age for each brain region; using the regional brain age label as a distillation supervision signal for a student model, training the student model based on the brain images of the multiple healthy subjects; predicting the regional brain age of the subject to be predicted based on the student model, and assessing the risk of the target brain disease based on the prediction results and a set of brain regions related to the target brain disease.

[0007] In one possible implementation, training the teacher model until convergence based on the brain images of the multiple healthy subjects includes: using the brain images of the healthy subjects as input and the actual age of the healthy subjects as a supervision signal, training the teacher model to output a whole-brain predicted brain age; optimizing the model parameters by minimizing the error between the whole-brain predicted brain age and the actual age until convergence.

[0008] In one possible implementation, the brain image of the healthy subject is a three-dimensional image; the step of constructing the regional input of each brain region based on the brain region mask of each brain region includes: for each brain region, multiplying the brain image of the healthy subject element-wise with the brain region mask of the corresponding brain region, retaining the voxel signal of the corresponding brain region; replacing voxels of non-brain regions with noise or constant values ​​to obtain the regional input of each brain region.

[0009] In one possible implementation, after obtaining the regional brain age of each brain region, the method further includes: for each brain region, replacing the brain region voxels with noise or constant values, while leaving the other brain regions unchanged, to obtain an occluded brain image; inputting the occluded brain image into the teacher model to obtain the occluded predicted brain age; and obtaining the regional correction signal for each brain region based on the occluded predicted brain age and the whole-brain predicted brain age.

[0010] In one possible implementation, generating a regional brain age label based on the regional brain age of each brain region includes: combining the regional brain age of each brain region with the regional correction signal of each brain region to generate a regional brain age label.

[0011] In one possible implementation, training the student model based on brain images of the plurality of healthy subjects, using the regional brain age label as a distillation supervision signal for the student model, includes:

[0012] With the goal of minimizing distillation loss and functional consistency loss, the student model is trained using brain images from the multiple healthy subjects; the distillation loss is used to make the student model approximate the regional brain age label, and the functional consistency loss is used to suppress non-smooth mutations in regional brain age within the same functional network.

[0013] In one possible implementation, the parameters of the assessment include the neurological disease-related index (NDC), which is expressed as:

[0014]

[0015] in, For regional brain age difference, , where n is the nth subject to be predicted. For the relevant brain regions, For the predicted results, The actual age of the subject to be predicted. A collection of brain regions associated with the target brain disease.

[0016] According to another aspect of this disclosure, a brain disease risk assessment device based on regional brain age prediction is provided, the device comprising:

[0017] The brain image segmentation module is used to segment brain regions from brain images of multiple healthy subjects, resulting in multiple brain regions and a brain region mask for each region.

[0018] The teacher model training module is used to train the teacher model until convergence based on the brain images of the multiple healthy subjects, and freeze the teacher model parameters after the training converges.

[0019] The regional brain age label generation module is used to construct the regional input of each brain region based on the brain region mask of each brain region, input the regional input of each brain region into the teacher model respectively, obtain the regional brain age of each brain region, and generate regional brain age labels based on the regional brain age of each brain region.

[0020] The student model training module is used to train the student model based on the brain images of the multiple healthy subjects by using the brain age labels of the region as distillation supervision signals for the student model.

[0021] The assessment module is used to predict regional brain age based on the student model of the subject to be predicted, and to assess the risk of the target brain disease based on the prediction results and the set of brain regions related to the target brain disease.

[0022] In one possible implementation, the teacher model training module is used to train the teacher model to output a whole-brain predicted brain age by taking the brain image of the healthy subject as input and the actual age of the healthy subject as a supervision signal; and to optimize the model parameters by minimizing the error between the whole-brain predicted brain age and the actual age until convergence.

[0023] In one possible implementation, the brain image of the healthy subject is a three-dimensional image; the regional brain age label generation module is used to multiply the brain image of the healthy subject element-wise with the brain region mask of the corresponding brain region for each brain region, retain the voxel signal of the corresponding brain region; and replace voxels of non-brain regions with noise or constant values ​​to obtain the regional input of each brain region.

[0024] In one possible implementation, the device further includes: a correction module, configured to replace the brain region voxels with noise or constant values ​​for each brain region, while leaving the remaining brain regions unchanged, to obtain an occluded brain image; input the occluded brain image into the teacher model to obtain the occluded predicted brain age; and obtain a region correction signal for each brain region based on the occluded predicted brain age and the whole-brain predicted brain age.

[0025] In one possible implementation, the regional brain age label generation module is used to combine the regional brain age of each brain region with the regional correction signal of each brain region to generate a regional brain age label.

[0026] In one possible implementation, the student model training module is used to train the student model using brain images of the multiple healthy subjects with the goal of minimizing distillation loss and functional consistency loss; the distillation loss is used to make the student model approximate the regional brain age label, and the functional consistency loss is used to suppress non-smooth mutations in regional brain age within the same functional network.

[0027] In one possible implementation, the parameters of the assessment include the neurological disease-related index (NDC), which is expressed as:

[0028]

[0029] in, For regional brain age difference, , where n is the nth subject to be predicted. For the relevant brain regions, For the predicted results, The actual age of the subject to be predicted. A collection of brain regions associated with the target brain disease.

[0030] According to another aspect of this disclosure, a non-volatile computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the above-described method.

[0031] According to another aspect of this disclosure, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the above-described method.

[0032] In this embodiment, brain images of multiple healthy subjects are divided into brain regions to obtain multiple brain regions and a brain region mask for each region. Based on the brain images of the multiple healthy subjects, a teacher model is trained until convergence, and the parameters of the teacher model are frozen after convergence. A regional input for each brain region is constructed based on the brain region mask, and the regional input for each brain region is input into the teacher model to obtain the regional brain age for each brain region. A regional brain age label is generated based on the regional brain age of each brain region. The regional brain age label is used as a distillation supervision signal for a student model, which is then trained based on the brain images of the multiple healthy subjects. Based on the student model, the regional brain age of the subject to be predicted is predicted using the brain images. Based on the prediction results and the set of brain regions related to the target brain disease, the risk of the target brain disease is assessed. This enables interpretable risk assessment of the target brain disease, providing a basis for early risk screening and assisted diagnosis.

[0033] In this embodiment, by training a teacher-student model, the brain age regression ability established by the teacher model from whole-brain data can be transferred to the student model via label distillation. With a significant reduction in the number of parameters and computational cost, the student model's predicted brain age for each brain region still approaches the accuracy of the teacher model, and reasonable physiological differences are observed between different brain regions, avoiding pattern collapse due to output convergence and ensuring the reliability of the prediction. By using the student model to predict regional brain age and assess the risk of target brain diseases from the brain images of the subjects to be predicted, the degree of aging in different brain regions can be reflected, providing an interpretable risk assessment of target brain diseases and offering a basis for early risk screening and assisted diagnosis.

[0034] Other features and aspects of this disclosure will become clear from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description

[0035] The accompanying drawings, which are included in and form part of this specification, illustrate exemplary embodiments, features, and aspects of this disclosure together with the specification and serve to explain the principles of this disclosure.

[0036] Figure 1 A flowchart illustrating a brain disease risk assessment method based on regional brain age prediction according to an embodiment of this disclosure is shown.

[0037] Figure 2 A schematic diagram of a brain disease risk assessment process according to an embodiment of this disclosure is shown.

[0038] Figure 3 A schematic diagram of the overall structure of the teacher model and student model according to an embodiment of the present disclosure is shown.

[0039] Figure 4 A block diagram of a brain disease risk assessment device based on regional brain age prediction according to an embodiment of the present disclosure is shown.

[0040] Figure 5 A block diagram of an apparatus 1900 according to an embodiment of the present disclosure is shown. Detailed Implementation

[0041] Various exemplary embodiments, features, and aspects of this disclosure will now be described in detail with reference to the accompanying drawings. The same reference numerals in the drawings denote elements that have the same or similar functions. Although various aspects of the embodiments are shown in the drawings, they are not necessarily drawn to scale unless specifically indicated otherwise.

[0042] As used herein, the terms “comprising,” “including,” “having,” or variations thereof are open-ended and include one or more of the stated features, integrals, elements, steps, components, or functions, but do not exclude the presence or addition of one or more other features, integrals, elements, steps, components, functions, or groups thereof.

[0043] When an element is referred to as “connected,” “coupled,” “responding,” or a variation thereof relative to another element, it may be directly connected, coupled, or responding to another element, or there may be an intermediate element present.

[0044] Although the terms first, second, third, etc., may be used herein to describe various elements / operations, these elements / operations should not be limited by these terms. These terms are only used to distinguish one element / operation from another. Therefore, without departing from the teachings of the inventive concept, a first element / operation in some embodiments may be referred to as a second element / operation in other embodiments.

[0045] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments.

[0046] Furthermore, to better illustrate this disclosure, numerous specific details are set forth in the following detailed description. Those skilled in the art will understand that this disclosure can be practiced without certain specific details. In some instances, methods, means, components, and circuits well known to those skilled in the art have not been described in detail in order to highlight the main points of this disclosure.

[0047] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, data stored, data displayed, etc.) and signals involved in this application are all authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant regions.

[0048] Against the backdrop of an increasingly aging population, the incidence and disease burden of various brain diseases continue to rise. Among them, neurodegenerative diseases such as Parkinson's disease (PD) and Alzheimer's disease (AD), as well as stroke and its sequelae of functional impairment, are the main causes of severe cognitive and motor function impairment in the elderly. These diseases often have insidious onset and progressive course; clinically, slow and complex changes have already occurred in brain structure and function before typical symptoms appear.

[0049] Brain age prediction is a quantitative assessment tool for brain health that has emerged in recent years. It typically involves constructing a regression model that maps the brain magnetic resonance imaging (MRI) images of healthy controls to their chronological age. The difference between the predicted brain age and the actual age is used to characterize the degree of brain aging or developmental deviation in an individual. Numerous studies have shown that an increased brain age difference (i.e., predicted brain age minus chronological age) is closely associated with the risk of various brain diseases, cognitive decline, and poor prognosis. Therefore, brain age is considered a potential risk phenotype or imaging biomarker for brain diseases.

[0050] However, most existing brain age-related studies use whole-brain MRI as a uniform input, outputting a global brain age value. While this method can reflect overall aging trends at the population level, it still has many limitations. On the one hand, brain development and aging processes exhibit significant spatial heterogeneity, with different brain regions showing significant differences in maturation time, atrophy rate, and disease susceptibility. Diseases such as PD and AD often manifest as selective damage to specific brain regions or brain networks, making it difficult to locate lesion-related brain regions and characterize inter-regional differences using only a single whole-brain age. On the other hand, different diseases may present different "regional aging fingerprints," making it difficult to distinguish disease types or explain which specific brain regions drive increased risk simply by relying on whole-brain age and its differences.

[0051] In this embodiment, brain images of multiple healthy subjects are divided into brain regions to obtain multiple brain regions and a brain region mask for each region. Based on the brain images of the multiple healthy subjects, a teacher model is trained until convergence, and the parameters of the teacher model are frozen after convergence. A regional input for each brain region is constructed based on the brain region mask, and the regional input for each brain region is input into the teacher model to obtain the regional brain age for each brain region. A regional brain age label is generated based on the regional brain age of each brain region. The regional brain age label is used as a distillation supervision signal for a student model, which is then trained based on the brain images of the multiple healthy subjects. Based on the student model, the regional brain age of the subject to be predicted is predicted using the brain images. Based on the prediction results and the set of brain regions related to the target brain disease, the risk of the target brain disease is assessed. This enables interpretable risk assessment of the target brain disease, providing a basis for early risk screening and assisted diagnosis.

[0052] Figure 1 A flowchart illustrating a brain disease risk assessment method based on regional brain age prediction according to an embodiment of this disclosure is shown. Figure 1 As shown, the method includes:

[0053] S11. Divide the brain images of multiple healthy subjects into brain regions to obtain multiple brain regions and brain region masks for each brain region.

[0054] Brain diseases can include Parkinson's disease (PD), Alzheimer's disease (AD), and upper limb motor dysfunction after stroke.

[0055] Regional Brain Age (ReBA) can be achieved by dividing the brain into multiple regions according to a standard brain region map, and then predicting the brain age of each region separately to obtain the regional brain age, in order to characterize the degree of aging and structural abnormality patterns at the brain region level.

[0056] Healthy subjects can be participants without brain diseases and in good overall health, serving as a normal reference standard. Brain images can be magnetic resonance imaging (MRI), specifically T1-weighted imaging (T1WI), and these images of healthy subjects can represent the normal brain aging process. Multiple brain images of healthy subjects can form a training set of healthy controls; in one possible implementation, the brain images of the healthy subjects are three-dimensional images.

[0057] After acquiring brain images from multiple healthy subjects, the brain image set can be preprocessed. Since brain images may come from different centers (such as different hospitals), different equipment, and different scanning protocols, standardized preprocessing can be performed on the brain images to eliminate the interference of technical differences from different sources on subsequent brain region identification and brain age prediction.

[0058] Specifically, preprocessing can first remove the skull from the brain image. A brain extraction tool based on deep learning can be used to remove the skull and non-brain tissue from the original MRI image, leaving only the brain parenchyma region, to obtain a debonded brain image.

[0059] After skull removal, bias field correction can be performed on brain images. Specifically, intensity non-uniformity correction can be performed on the post-bone images to reduce the influence of low-frequency bias fields caused by equipment and acquisition processes, resulting in clearer gray-white matter boundaries.

[0060] Finally, nonlinear registration can be performed. The corrected images can be nonlinearly registered to a standard brain template space, such as the Montreal Neurological Institute space (MNI), to unify voxel resolution and spatial coordinates, aligning the brain structures of different subjects in the same coordinate system, thereby enhancing the compatibility and comparability of multicenter data.

[0061] After preprocessing, standardized whole-brain input can be obtained. .

[0062] In this embodiment of the disclosure, by preprocessing the original brain images, the technical heterogeneity caused by different acquisition centers, devices and scanning protocols can be effectively eliminated, so that the subsequent brain region division and feature extraction are based on unified and comparable data, thereby improving the stability and generalization ability of the brain age prediction model.

[0063] In this embodiment, the acquired brain images can be divided into brain regions. Standard brain region templates can be used to divide standardized MRI into brain regions, dividing the whole brain into multiple anatomical or functional brain regions, and generating a corresponding mask for each brain region for subsequent region input construction and region-level analysis. The standard brain region template can be the Harvard–Oxford Cortical and Subcortical Structural Atlas (HOA), or it can be an Automated Anatomical Labeling atlas (AAL), Schaefer atlas, etc., and this embodiment does not limit the scope of the template.

[0064] Using the Harvard-Oxford brain atlas as an example, the Harvard-Oxford brain atlas can be registered with standardized whole-brain input. Consistent standard space and resolution, assigning brain region labels to each voxel Then, one-hot encoding can be performed on the integer tag to obtain a value of size [size missing]. Multichannel brain region mask Each channel can correspond to a brain region. The number of brain regions can be determined based on the selected brain map, the required level of detail, and the actual needs of those skilled in the art. This disclosure does not impose any limitations on this. In this disclosure, It can be 48. Binary mask of the r-th brain region .

[0065] In one possible implementation, based on the defined brain regions and according to clinical common sense, multiple anatomical brain regions can be grouped into several functional networks according to their shared functions. For example, the motor network can include brain regions related to limb movement; the default mode network can include brain regions active in the resting state; and the memory-related network can include brain regions related to learning and memory functions. The resulting set of functional networks can be represented as follows: .

[0066] In one possible approach, for a specific brain disease, a set of disease-related brain regions can be pre-determined based on clinical common sense. For example, Parkinson's disease (PD) primarily affects the caudate nucleus, putamen, and related motor network brain regions. Structural or functional abnormalities in these brain regions can reflect the pathological characteristics of PD and therefore can be used as a set of disease-related brain regions.

[0067] It should be noted that the functional network division and disease-related brain region set can be adjusted according to different diseases and specific clinical needs, and the embodiments disclosed herein do not impose any restrictions on this.

[0068] In this embodiment of the disclosure, the standardized brain image is registered to a standard brain atlas, and the whole brain is divided into multiple anatomical brain regions according to the brain atlas, thereby generating a corresponding spatial mask for each brain region, so as to realize the independent extraction and subsequent analysis of signals from different brain regions.

[0069] S12. Based on the brain images of the multiple healthy subjects, train the teacher model until convergence, and freeze the parameters of the teacher model after training convergence.

[0070] The teacher model can be a whole-brain age regression model. During the construction of the teacher model, a backbone network can be defined, and a regression head can be set at the tail of the backbone network. Specifically, the backbone network can be a three-dimensional convolutional neural network or other three-dimensional feature extraction networks. In this embodiment, the backbone network can be a three-dimensional densely connected convolutional network (3DDenseNet).

[0071] In this embodiment of the disclosure, the teacher model can be trained until convergence based on the brain images of the multiple healthy subjects. After training is completed, all parameters of the teacher model can be frozen to provide regional brain age labels for the student model.

[0072] In one possible implementation, training the teacher model until convergence based on the brain images of the multiple healthy subjects includes: using the brain images of the healthy subjects as input and the actual age of the healthy subjects as a supervision signal, training the teacher model to output a whole-brain predicted brain age; optimizing the model parameters by minimizing the error between the whole-brain predicted brain age and the actual age until convergence.

[0073] Specifically, the teacher model can be trained using whole-brain MRI scans and actual ages of healthy control subjects. During each training iteration, the whole-brain input X is fed into the 3D DenseNet backbone network. High-dimensional brain image features are extracted through structures such as convolutional layers, dense blocks, and pooling layers, and the raw MRI data is converted into a deep feature representation for brain age regression. The regression head outputs a whole-brain predicted brain age. Then, using the actual age of healthy control subjects as a supervision label, the parameters of the teacher model are optimized by minimizing the error between the predicted brain age and the actual age until convergence, thus completing the training of the teacher model.

[0074] Specifically, L1 or L2 loss can be used to quantify the error between the teacher model's predicted brain age and the actual age into a scalar loss value. Specifically, the absolute value of the difference between the predicted brain age and the actual age can be calculated as the L1 loss; the square of the difference can be calculated as the L2 loss. The L1 or L2 loss can be minimized through backpropagation to optimize the network parameters of the teacher model until the model converges.

[0075] In this embodiment of the disclosure, by constructing and training a teacher model, it is possible to learn the aging patterns of the normal brain from a healthy control population, providing high-quality labels for subsequent student models, thereby achieving effective transfer of regional brain age knowledge.

[0076] S13. Based on the brain region mask of each brain region, construct the region input of each brain region, input the region input of each brain region into the teacher model respectively, obtain the region brain age of each brain region, and generate the region brain age label based on the region brain age of each brain region.

[0077] In this embodiment of the disclosure, a brain region mask can be used. Construct the region input and obtain the initial region brain age prediction based on the frozen teacher model. That is, input the region input of each brain region into the teacher model to obtain the region brain age of each brain region.

[0078] In one possible implementation, constructing the regional input for each brain region based on the brain region mask of each brain region includes: for each brain region, multiplying the brain image of a healthy subject element-wise with the brain region mask of the corresponding brain region, retaining the voxel signal of the corresponding brain region; replacing voxels of non-brain regions with noise or constant values ​​to obtain the regional input for each brain region.

[0079] Because there may be boundary registration errors between the brain region mask and the standardized MRI, the brain region mask can first be expanded by one voxel to increase its range and reduce omissions caused by registration errors. Then, the brain image of a healthy subject can be multiplied element-wise with the corresponding brain region mask to preserve the voxel signal of that brain region. Voxels from non-brain regions can be replaced with low-amplitude noise or constant values ​​to avoid the teacher model over-relying on information from other regions at the whole-brain level, thus obtaining the regional input for each brain region.

[0080] Specifically, the regional input of each brain region The formula can be formula (1):

[0081] (1)

[0082] in, For the first Regional input to each brain region, For whole-brain input, This represents element-wise multiplication. Indicates the first Brain region masking, The noise scaling factor. It is zero-mean, unit-variance Gaussian noise.

[0083] After obtaining the regional input for each brain region, the regional input for each brain region can be... The data are input into the teacher model to obtain the regional brain age of each brain region. , i.e., Initial ReBA. The regional brain age of each brain region can be the initial regional brain age, and the formula for regional brain age can be found in formula (2):

[0084] (2)

[0085] in, The teacher model is frozen.

[0086] In one possible implementation, after obtaining the regional brain age of each brain region, the method further includes: for each brain region, replacing the brain region voxels with noise or constant values, while leaving the other brain regions unchanged, to obtain an occluded brain image; inputting the occluded brain image into the teacher model to obtain the occluded predicted brain age; and obtaining the regional correction signal for each brain region based on the occluded predicted brain age and the whole-brain predicted brain age.

[0087] Since the initial regional brain age may be affected by whole-brain prediction bias and inter-regional averaging effects, a regional correction signal may also be introduced in this embodiment. For each brain region, the regional brain age is calculated... Make corrections.

[0088] Specifically, for each brain region, a whole-brain input after removing that region can be constructed. This involves replacing the voxels of that brain region with noise or constant values ​​while leaving other brain regions unchanged, thus obtaining the occluded whole-brain input. The formula can be formula (3). The occluded brain image can be input into the teacher model to obtain the predicted brain age after occlusion, and the difference between the predicted brain age of the whole brain before and after occlusion can be calculated. The formula for the difference can be formula (4). The average or median of the difference can be calculated from the brain images of multiple healthy subjects to obtain the region correction signal of that brain region. The regional correction signal for this brain region can be expressed by formula (5). This regional correction signal is used to measure the average direction and intensity of the brain region's influence on the prediction of whole-brain age.

[0089] (3)

[0090] (4)

[0091] The meanings of the specific parameters in formulas (3) and (4) are explained above.

[0092] (5)

[0093] in, For the healthy control training set In the middle, for all samples Take the average.

[0094] In one possible implementation, generating a regional brain age label based on the regional brain age of each brain region includes: combining the regional brain age of each brain region with the regional correction signal of each brain region to generate a regional brain age label.

[0095] After obtaining the region correction signal for each brain region, the regional brain age of each brain region can be combined with the region correction signal to generate a regional brain age label. This regional brain age label can be a soft regional brain age label. Specifically, the deviation direction between the regional brain age of each brain region and the predicted brain age of the whole brain can be compared with the direction of the region correction signal for that brain region. When the deviation direction of the regional brain age is consistent with the direction indicated by the region correction signal, the initial regional brain age can be additively corrected along the direction of the region correction signal under the constraint of a preset step size parameter to obtain a soft regional brain age label. When the two directions are inconsistent or the deviation is obviously abnormal, no correction can be performed to avoid introducing unreasonable offsets. The corrected soft regional brain age label, Soft ReBA, can be obtained in this way. The soft regional brain age label for the r-th region... The formula can be formula (6):

[0096] (6)

[0097] in, This is a preset step size coefficient, which can be pre-set. This is an indicator function. When the direction of the deviation between the initial regional brain age and the whole brain age is consistent with the direction of the regional correction signal, an additive correction is performed. The value is 1 when the direction of the deviation between the initial regional brain age and the whole brain age is consistent with the direction of the regional correction signal, and 0 otherwise. When the above conditions are not met, the initial regional brain age may not be corrected or may only undergo a limited fine-tuning.

[0098] In this embodiment of the disclosure, by introducing a regional correction signal to correct the regional brain age, the deviation of the regional brain age can be corrected, thereby improving the accuracy of regional brain age prediction.

[0099] S14. Using the brain age label of the region as a distillation supervision signal for the student model, the student model is trained based on the brain images of the multiple healthy subjects.

[0100] Specifically, for each subject and each brain region r, the student model can employ a three-dimensional backbone network with the same structure as the teacher model's backbone network (i.e., isomorphic) to extract stable region embedding features from the inputs of each brain region. The network parameters of the two models are independent of each other and do not share or inherit parameters. In this embodiment of the disclosure, the backbone network of the student model can be a 3D DenseNet.

[0101] Enter the region Inputting the student network backbone, extracting region-level feature maps, and performing operations such as global average pooling to obtain a fixed-length vector with the same length as the number of channels, which serves as the region embedding feature for that brain region. This embedding feature eliminates spatial location differences in the original feature map while preserving the main semantic information of the brain region.

[0102] In this embodiment of the disclosure, before training the student model, a learnable region cue vector can be set for each brain region, and automatically optimized during the student model training process. When processing a specific brain region, the region cue vector corresponding to that brain region can be input into a feature-wise linear modulation (FiLM) module. The lightweight FiLM module maps the region cue vector to channel-wise scaling factors. and per-channel bias term Modulating the region embedding features allows the model to adaptively adjust feature representations based on the current brain region, thereby improving the accuracy of brain age prediction for each brain region, resulting in a fused embedding vector. See formula (7) for details.

[0103] (7)

[0104] Modulated regional features The data is fed into a lightweight adapter model, which outputs a region-predicted brain age for brain region r. . Specifically, The formula can be formula (8). For each subject, this disclosure embodiment can output the complete regional brain age vector at once, instead of outputting only a single brain region or a single whole brain age.

[0105] (8)

[0106] in, This represents the function mapping for the lightweight adapter model.

[0107] In this embodiment of the disclosure, the regional brain age label output by the teacher model can be used as a distillation supervision signal during the training of the student model to guide the regional prediction results of the student model to approach the output distribution of the teacher model, thereby realizing knowledge transfer.

[0108] In one possible implementation, the step of using the regional brain age label as a distillation supervision signal for the student model and training the student model based on the brain images of the multiple healthy subjects includes: training the student model using the brain images of the multiple healthy subjects with the goal of minimizing distillation loss and functional consistency loss; the distillation loss is used to make the student model approximate the regional brain age label, and the functional consistency loss is used to suppress non-smooth mutations in regional brain age within the same functional network.

[0109] In this embodiment, the total loss of the student model can be composed of two parts: distillation loss and functional consistency loss, used to simultaneously constrain the approximation of regional brain age to the soft label and the spatial smoothness within the functional network. Total Loss This can be expressed as formula (9).

[0110] (9)

[0111] in, For distillation losses, For functional consistency loss, Hyperparameters used to adjust the relative weights of distillation loss and functional consistency loss.

[0112] Distillation loss can be used to approximate the regional brain age predicted by the student model to the soft region brain age label Soft ReBA. In this embodiment of the disclosure, distillation loss... It can be the average L1 loss across the whole brain or all healthy subjects. For the specific formula, please refer to formula (10).

[0113] (10)

[0114] in, This represents the student model's prediction of regional brain age in the r-th brain region for the nth healthy subject sample. This represents the brain age label of the nth healthy subject sample in the rth brain region by the teacher model, where R is the total number of brain regions and N is the number of healthy subject samples.

[0115] By minimizing This allows the student model's predictions in each brain region to be as close as possible to the soft region brain age labels generated by the teacher.

[0116] Loss of functional consistency can be used as a clinical prior to suggest that brain regions within the same functional network age at similar rates. For the k-th functional network in the nth healthy subject sample, the average brain age of each brain region within that network can be calculated first. For the specific formula, please refer to formula (11).

[0117] (11)

[0118] in, For functional network sets, Let K be the set of brain region indices included in the k-th functional network, where K is the number of functional networks.

[0119] The deviation between the predicted values ​​of each brain region and the average value of the corresponding functional network can be used as a penalty term, resulting in functional consistency loss. See formula (12) for details.

[0120] (12)

[0121] By minimizing It can suppress abnormal jumps in brain age within the same functional network, making the spatial distribution of brain age in the functional network level smoother and more consistent with biological priors.

[0122] In one possible implementation, to evaluate the stability of the student model in predicting regional brain age in healthy individuals, the distribution of regional brain age differences can be statistically analyzed in both the healthy control training set and the unseen healthy control test set.

[0123] For the r-th brain region, a set of regional brain age differences can be constructed on the training set. and the set of regional brain age differences on the test set For specific formulas, please refer to formulas (13) and (14).

[0124] (13)

[0125] (14)

[0126] in, For a healthy control training set, No healthy control test set was found, so there is no overlap between the two. The remaining parameters are described above.

[0127] The distribution distance between the two can be measured by the maximum mean difference (MMD) based on the Gaussian kernel. For the specific formula, please refer to formula (15).

[0128] (15)

[0129] Where m is the number of healthy control training set samples, and l is the number of unseen healthy control test set samples. , , This represents the Gaussian kernel function.

[0130] Furthermore, the MMD distribution distance can be converted into a value range of... Health similarity index , .

[0131] Based on the normalized MMD distribution difference function, the healthy control similarity index HCS for the r-th brain region can be found in formula (16).

[0132] (16)

[0133] The formula parameters are described above.

[0134] HCS can be used to evaluate the statistical consistency of regional brain age across different healthy populations. If the distributions are very similar, it indicates that the student model has good generalization ability and will not fluctuate significantly due to changes in the population or data collection conditions. If the differences are large, it suggests that the student model may over-rely on certain specific features in the training set, and the model may be unreliable.

[0135] In this embodiment of the disclosure, the health control similarity index HCS can be obtained by comparing the statistical distance between the brain age difference distribution of corresponding brain regions in the health control training set and the health control test set. The higher the HCS value, the better the statistical consistency of regional brain age among different healthy populations.

[0136] In this embodiment of the disclosure, the trained student model can be applied to different subject populations, including healthy controls, Parkinson's disease patients, Alzheimer's disease patients, and patients with upper limb motor dysfunction after stroke, during the testing phase, thereby achieving brain disease risk assessment under weak supervision of overall age.

[0137] S15. Based on the student model, perform regional brain age prediction on the brain images of the subject to be predicted, and assess the risk of the target brain disease based on the prediction results and the set of brain regions related to the target brain disease.

[0138] The set of brain regions related to the target brain disease can be the set of brain regions related to the target brain disease identified above. In this embodiment of the disclosure, regional brain age prediction can be performed on the brain image of the subject to be predicted based on a trained student model, to obtain the regional predicted brain age of each region of the subject to be predicted. The risk of the target brain disease can be assessed based on the regional predicted brain age of each region and the set of brain regions related to the target brain disease.

[0139] In one possible implementation, the parameters of the assessment include the neurological disease-related index NDC, which is expressed as formula (17):

[0140] (17)

[0141] in, For regional brain age difference, , where n is the nth subject to be predicted. For the relevant brain regions, For the predicted results, The actual age of the subject to be predicted. A collection of brain regions associated with the target brain disease.

[0142] , A higher value indicates a more pronounced pattern of accelerated brain aging in the target disease-related brain regions, suggesting a higher risk of brain disease; conversely, a lower value indicates a lower risk. If the value is close to 0, it indicates that the regional brain age difference in the relevant brain region is generally small or negative, and has a low similarity to the disease characteristic pattern.

[0143] This embodiment can quantify the risk of brain diseases in individual subjects or specific populations, and can visually annotate key brain regions with high NDC on a standard brain template, providing imaging evidence for the localization of high-risk brain regions, individualized follow-up, and clinical auxiliary diagnosis.

[0144] In this embodiment, brain images of multiple healthy subjects are divided into brain regions to obtain multiple brain regions and a brain region mask for each region. Based on the brain images of the multiple healthy subjects, a teacher model is trained until convergence, and the parameters of the teacher model are frozen after convergence. A regional input for each brain region is constructed based on the brain region mask, and the regional input for each brain region is input into the teacher model to obtain the regional brain age for each brain region. A regional brain age label is generated based on the regional brain age of each brain region. The regional brain age label is used as a distillation supervision signal for a student model, which is then trained based on the brain images of the multiple healthy subjects. Based on the student model, the regional brain age of the subject to be predicted is predicted using the brain images. Based on the prediction results and the set of brain regions related to the target brain disease, the risk of the target brain disease is assessed. This enables interpretable risk assessment of the target brain disease, providing a basis for early risk screening and assisted diagnosis.

[0145] Application scenario examples

[0146] In this embodiment, to overcome the shortcomings of existing brain disease auxiliary diagnostic methods, such as only providing whole-brain age or coarse-grained classification results, insufficient generalization ability to multi-center MRI data, and lack of clinical interpretability at the brain region level, a brain disease risk assessment method based on regional brain age prediction is provided, hereinafter referred to as ReBA-Pred-Net. This method uses brain MRI as input and, under the condition of using only the overall age of healthy controls as a weakly supervised signal, automatically predicts brain region-level brain age (regional brain age) through a teacher-student model framework and brain region segmentation strategy. It combines the degree of brain region aging (brain age difference) to construct risk assessment indicators such as the similarity between healthy controls and the correlation with neurological diseases. This allows for interpretable, quantifiable, and multi-center robust risk assessment of various brain diseases, such as Parkinson's disease, Alzheimer's disease, and post-stroke upper limb motor dysfunction, providing a basis for early risk screening and auxiliary diagnosis.

[0147] Figure 2 A schematic diagram of a brain disease risk assessment process according to an embodiment of this disclosure is shown.

[0148] like Figure 2 As shown, the brain disease risk assessment method based on regional brain age prediction in this disclosure includes:

[0149] S1. Acquire and preprocess brain MRI, acquire three-dimensional T1-weighted brain MRI of the subject, and obtain standardized brain images through a unified preprocessing process;

[0150] Specifically, this step involves acquiring raw brain MRI images (preferably T1-weighted structural images) of the subjects and processing the images using a standardized preprocessing procedure to ensure the comparability of data acquired from different centers and devices in terms of space and intensity.

[0151] Specifically, this may include:

[0152] S1.1 Skull Removal: Using deep learning-based brain extraction tools (such as HD-BET), the skull and non-brain tissues are removed from the original T1-weighted MRI, leaving only the brain parenchyma, to obtain a debonded brain image.

[0153] S1.2 Bias Field Correction: The N4 algorithm in the ANTs tool is used to correct the intensity non-uniformity of the boned image, reducing the influence of low-frequency bias field caused by equipment and acquisition process, and making the gray-white matter boundary clearer.

[0154] S1.3 Nonlinear Registration: The SyN algorithm in ANTs is used to nonlinearly register the corrected images to a standard brain template space (e.g., MNI space), unifying voxel resolution and spatial coordinates, so that the brain structures of different subjects are aligned in the same coordinate system, thereby enhancing the compatibility and comparability of multicenter data.

[0155] After the above steps, a standardized MRI image is obtained, which serves as the input for the subsequent teacher-student model regional brain age prediction module.

[0156] S2. On the standardized brain image, the whole brain is divided into brain regions based on a preset brain region atlas to obtain multiple brain regions and corresponding brain region masks, and to determine the functional network division and the disease-related brain region set for different brain diseases.

[0157] Specifically, this step uses a standard brain region template to divide the brain into multiple anatomical or functional brain regions based on standardized MRI, and generates a corresponding mask for each brain region for subsequent regional input construction and regional analysis.

[0158] Specifically, this may include:

[0159] In this embodiment, the Harvard-Oxford brain region template is used to partition the brain on preprocessed images in standard space, dividing the whole brain into several anatomical brain regions, and further constructing functional networks and disease-related brain region priors.

[0160] S2.1 Brain Region Segmentation and Template One-Hot Encoding: The Harvard–Oxford map is registered to a consistent standard space and resolution, and brain region labels are assigned to each voxel. (In this embodiment, R=48). The integer tag is then one-hot encoded to obtain a size of... Multichannel brain region mask Each channel corresponds to a specific brain region.

[0161] S2.2 Functional Networks and Disease-Related Brain Region Priors: Based on the brain region division, and according to previous literature and clinical common sense, anatomical brain regions are further classified into several functional networks (such as motor networks, default mode networks, memory-related networks, etc.), forming a set of functional networks. Simultaneously, for the target brain disease, a set of corresponding disease-related brain regions is provided in advance. For example, PD mainly affects the caudate nucleus, putamen, and related motor network brain regions, which are used to construct subsequent neurological disease-related indicators.

[0162] In other embodiments, the brain region template can be replaced with atlases such as AAL and Schaefer, and the functional network division and disease-related brain region set can be adjusted according to different diseases and specific clinical needs.

[0163] S3. Using standardized brain images and actual ages of each subject in the healthy control training set as training data, construct and train a teacher model, perform brain age regression on whole brain images to obtain a whole brain age prediction model, and freeze the teacher model parameters after training convergence.

[0164] Specifically, this step constructs a teacher model, using the overall actual age of healthy control subjects as the sole supervisory signal, to perform brain age regression modeling on whole-brain MRI, resulting in a whole-brain brain age prediction model. Specifically, the standardized whole-brain MRI is fed into a 3D convolutional neural network or other 3D feature extraction network (e.g., 3D DenseNet). The network backbone automatically extracts structural features at the whole-brain level, and a regression head is set at the network tail to map the whole-brain features into a single scalar output representing the predicted whole-brain brain age. During the training phase, the actual age of healthy control subjects is used as the supervisory label. The teacher model parameters are optimized by minimizing the error between the predicted brain age and the actual age until convergence. After training, the teacher model parameters are frozen, serving as a robust baseline model for subsequently generating regional brain age soft labels and regional correction signals.

[0165] Specifically, this may include:

[0166] The purpose of the teacher model is to convert whole-brain age prediction into soft regional brain age (Soft ReBA) that can be used for training student models under weak supervision with only overall age labels of healthy controls.

[0167] S3.1 Whole-brain age regression model training: On a healthy control training set, a whole-brain age regression model (teacher backbone) is constructed. This embodiment uses a 3D DenseNet backbone structure. Perform feature encoding and output whole-brain predicted brain age Based on the actual age of the subjects For supervision, the network parameters are optimized using L1 or L2 loss until convergence. After training, all parameters of the teacher model are frozen for subsequent region-level soft label generation.

[0168] S3.2 Region Input Construction and Initial Region Brain Age (Initial ReBA): Based on the frozen teacher model, a one-hot brain region mask is used. Construct the region input for each brain region. To reduce boundary registration errors, this embodiment performs one voxel dilation on the brain region mask and replaces non-target brain region voxels with small-amplitude Gaussian noise. Subsequently, construct the region input for each brain region. By sequentially inputting the frozen teacher model, the initial regional brain age prediction for that brain region is obtained. , namely Initial ReBA.

[0169] S3.3 Region Occlusion Correction and Soft ReBA: Given that the initial region brain age may be affected by whole-brain prediction bias and inter-region averaging effects, this embodiment introduces a region correction signal through a "region occlusion – whole-brain prediction change" approach to additively correct the Initial ReBA and generate a soft region brain age (Soft ReBA). The process is as follows:

[0170] Construct "removal of a certain brain region" After the whole brain input, the voxel of that brain region is replaced with noise while other brain regions remain unchanged. The input is then frozen into the teacher model to obtain the whole brain age prediction after occlusion.

[0171] The difference in predicted brain age before and after occlusion was calculated, and the average of this difference was calculated in a healthy control training set to represent the brain region. The area correction signal;

[0172] According to the preset step size coefficient, the region correction signal is added to the Initial ReBA to obtain the brain region. Soft ReBA.

[0173] S4. Construct corresponding region inputs using the mask of each brain region, input the region inputs of each brain region into the frozen teacher model respectively, and obtain the initial region brain age prediction value for each brain region, forming an initial region brain age set;

[0174] Specifically, this step utilizes a brain region mask to construct regional inputs and obtains initial regional brain age predictions based on a frozen teacher model. This includes: for each brain region, multiplying the standardized MRI scan element-wise with the corresponding brain region mask, preserving the voxel signal of that brain region; replacing voxels outside the brain region with low-amplitude noise or constant values ​​to avoid the teacher model over-relying on information from other regions at the whole-brain level; sequentially inputting the corresponding regional inputs for each brain region into the teacher model to obtain the initial regional brain age prediction for each brain region; and forming a set containing the initial regional brain ages of all brain regions, serving as the basis for subsequent soft-region brain age construction. Through this step, the teacher model outputs brain age estimates under the condition of "only looking at a single brain region," providing a preliminary reference for modeling regional aging.

[0175] Specifically, this may include:

[0176] The student model, guided by teacher-generated Soft ReBA, learns to stably output regional brain age predictions on unseen subjects. It mainly comprises three steps: region feature extraction, learnable cues + feature modulation, and lightweight adapter output, which are jointly optimized using distillation loss and functional consistency loss.

[0177] S4.1 Region Feature Extraction and Embedding Representation: For each subject and each brain region r, this embodiment uses a 3D backbone network (parameter-independent) isomorphic to the teacher backbone to input the region. Input the student backbone network, extract region-level feature maps, and obtain fixed-length region embedding features through operations such as global average pooling. .

[0178] S4.2 Learnable Hints and Linear Modulation of Feature Dimensions (FiLM): To enhance the personalized modeling capability of each brain region, this embodiment sets a learnable hint vector for each brain region r, and modulates the hint vector into a channel-wise scaling factor through a lightweight FiLM module. and bias terms And linearly modulate the region embedding features to obtain the fused embedding vector, as shown in formula (7).

[0179] S4.3 Lightweight adapter output region brain age: modulated regional features Feed the data into a lightweight adapter model (e.g., a two-layer fully connected network) and output the brain region. Regional brain age prediction For each subject, this embodiment can obtain the regional brain age vectors of all R brain regions at once.

[0180] S5. Based on the regional occlusion construction, remove the whole-brain input of a certain brain region, calculate the difference in the whole-brain age prediction value before and after occlusion, and perform statistics in the healthy control training set to obtain the regional correction signal of each brain region.

[0181] Specifically, given that the initial regional brain age obtained solely based on regional input is susceptible to noise and data bias, this step estimates the marginal contribution of each brain region to the overall brain age through a "regional occlusion – whole-brain prediction change" approach, constructing a regional correction signal to correct the initial regional brain age. Specifically: for each brain region, an occluded whole-brain input is constructed: the voxel of that brain region is replaced with noise or a constant value, while the original signals for other brain regions remain unchanged; the whole-brain inputs before and after occlusion are input into the frozen teacher model, and the difference in the predicted brain age before and after occlusion is calculated; within the healthy control training set, the above differences for the same brain region are statistically analyzed to obtain the regional correction signal for that brain region, which measures the average direction and intensity of the brain region's influence on the predicted brain age; the regional correction signal can distinguish brain regions that have an increasing or decreasing effect on the overall brain age, providing a basis for subsequent soft-region brain age correction.

[0182] S6. Combine the initial regional brain age of each brain region with the corresponding whole-brain predicted brain age and the regional correction signal of that brain region to generate soft regional brain age labels, and use the set of soft regional brain age labels of all brain regions as the regional distillation supervision signal of the student model.

[0183] This step, based on the initial regional brain age, introduces a regional correction signal to conditionally correct the regional brain age, generating soft regional brain age labels for student model training, thus mitigating the pattern collapse problem caused by overly similar regional prediction results. Specifically, this includes: comparing the deviation direction between the initial regional brain age and the whole-brain predicted brain age for each brain region, and comparing it with the direction of the regional correction signal for that brain region; when the deviation direction of the initial regional brain age is consistent with the direction indicated by the regional correction signal, additively correcting the initial regional brain age along the direction of the regional correction signal under the constraint of a preset step size parameter, to obtain the soft regional brain age label.

[0184] When the two directions are inconsistent or the deviation is significantly abnormal, it is possible to choose not to make corrections or to make only minor adjustments with limited amplitude to avoid introducing unreasonable biases. The set of soft-region brain age labels for each brain region after the above corrections is used to guide the student model's learning. Through this step, regional brain age labels that are more consistent with the whole-brain aging pattern can be obtained under weak supervision of overall age, providing a more reliable teacher signal for the student model's regional predictions.

[0185] S7. Construct a student model, taking the input of each brain region as input, using the shared three-dimensional feature extraction backbone, learnable regional cue vectors and feature modulation module to extract fusion embedding vectors, outputting the regional brain age prediction values ​​of each brain region through a lightweight adapter, and jointly minimizing distillation loss and functional consistency loss on a healthy control training set to train the student model;

[0186] Specifically, it may include:

[0187] S7.1 Student Model Regional Brain Age Prediction: During the construction and training of the student model, the student model is first set as a 3D feature extraction backbone isomorphic to or shares some parameters with the teacher model backbone, used to extract stable regional embedding features from each brain region input. Based on this, a learnable regional cue vector is set for each brain region. This cue vector is input into a feature modulation module (such as a FiLM structure) to generate corresponding scaling and offset parameters. Channel-wise affine modulation is then performed on the regional embedding features of that brain region, thus achieving a regional representation mechanism of "shared feature representation + brain region-specific readout" under the premise of sharing the backbone. Subsequently, the modulated regional embedding features are input into a lightweight adapter module to obtain the student model's regional brain age prediction values ​​for each brain region. During the training phase, the soft regional brain age labels output by the teacher model are used as a supervision signal to define a distillation loss, constraining the student model's regional brain age prediction to approximate the corresponding soft labels. Combined with a functional consistency loss based on functional network partitioning, a smoothing constraint is applied to the regional brain age of each brain region within the same functional network. By jointly optimizing the above loss functions, a student model that can stably output regional brain age on unseen subjects is obtained.

[0188] Figure 3 A schematic diagram of the overall structure of the teacher model and student model according to an embodiment of the present disclosure is shown. Figure 3 This shows the process of building and training teacher and student models.

[0189] S8. For any subject, the difference between the brain age of each brain region predicted by the student model and the subject's actual age is taken as the regional brain age difference. The distribution of regional brain age differences is statistically analyzed in the healthy control training set and the healthy control test set to construct the healthy control similarity index HCS. The regional brain age differences are normalized and averaged in the disease-related brain region set to construct the neurological disease correlation index NDC. The subject's brain disease risk is quantitatively assessed by combining the HCS and NDC.

[0190] Specifically, this may include:

[0191] S8.1 Construction of HCS (Health Control Similarity Index) and NDC (Neurological Disease Correlation Index) as a Similarity Index between Healthy Controls

[0192] To evaluate the stability of the student model in predicting regional brain age in healthy individuals, this embodiment statistically analyzes the distribution of regional brain age differences in both the healthy control training set and the unseen healthy control test set.

[0193] Specifically, this can include: after the student model obtains the predicted regional brain age for each brain region, for any subject, the predicted regional brain age for each brain region is first subtracted from its actual age to obtain the brain region-level brain age difference, which serves as a quantitative indicator of the degree of aging in that brain region deviating from the normal trajectory. In the healthy control training set and the unseen healthy control test set, the distribution of the brain age difference for each brain region is statistically analyzed, and a healthy control similarity index is constructed using statistical distance (e.g., maximum mean difference) to evaluate the statistical consistency of regional brain age across different healthy populations. Furthermore, for target brain diseases such as Parkinson's disease, Alzheimer's disease, and post-stroke upper limb motor dysfunction, corresponding disease-related brain regions are selected based on prior clinical and pathological knowledge. The brain age differences of these brain regions are normalized using a monotonic mapping function and then averaged to construct a neurological disease correlation index, used to measure the overall degree of "accelerated aging" in the corresponding disease-related brain regions of the subjects. Finally, by combining the healthy control similarity and the neurological disease correlation index, the risk of brain disease in the subjects is quantitatively assessed, and the assessment results are used as imaging evidence for early clinical risk screening and auxiliary diagnosis.

[0194] In this embodiment, multi-center T1-weighted MRI images of the brain are used as input. First, the images undergo bone removal, bias field correction, and spatial registration to obtain standardized brain images. Then, the entire brain is partitioned using a standard brain region atlas to obtain masks for each brain region and disease-related brain region sets for different brain diseases. Based on this, a teacher-student regional brain age prediction framework, ReBA-Pred-Net, is constructed: the teacher model performs brain age regression on the whole-brain images and generates soft-region brain age labels with spatial differences by constructing regional inputs according to brain regions and correcting for regional occlusion; the student model extracts regional features for each brain region, introduces learnable cues and FiLM feature modulation, and outputs regional brain age predictions via a lightweight adapter. Furthermore, the regional brain age difference for each brain region is calculated, and a healthy control similarity index is constructed in a healthy control population. A neurological disease correlation index (NDC) is constructed on the disease-related brain region set to quantitatively assess the risk of brain diseases such as Parkinson's disease, Alzheimer's disease, and post-stroke upper limb motor dysfunction. This invention can obtain stable, brain region-specific risk assessment results under multi-center data conditions, providing interpretable imaging evidence for early risk screening and auxiliary diagnosis of brain diseases.

[0195] Dataset preparation and model training configuration:

[0196] In this embodiment, the training set contains only healthy controls (HC) subjects, while the test set includes patients without HC and patients with various brain diseases. Specifically, the HC training set contains 1038 raw T1-weighted MRIs, with data sourced from multi-center publicly available datasets such as GSP, ADNI, and PPMI, and includes only healthy controls, used for training the teacher and student models. The healthy control test set (HC test set) contains 257 HCs who did not participate in training, sourced from the OpenNeuro database and a self-built cohort, used to calculate HCS indices. The Parkinson's disease test set (PD test set) contains 169 PD patients, sourced from PPMI and a self-built cohort, used to calculate NDC indices associated with Parkinson's disease. This embodiment uses the PyTorch framework on a computing platform equipped with an NVIDIA GPU to complete model training. The backbone uses a 3D DenseNet structure to encode T1 MRIs registered to standard space. The optimizer is AdamW, with an initial learning rate set to [missing value]. The weight decay coefficient is set to The learning rate gradually decays using a cosine annealing strategy; the training batch size is 4, the number of training rounds is 60, and key hyperparameters include the teacher's regional brain age correction weight. Regional obstruction noise amplitude coefficient and functional consistency loss weight .

[0197] Experimental Results and Analysis:

[0198] To comprehensively evaluate the ReBA-Pred-Net method of this invention in predicting regional brain age and assessing brain disease risk on multicenter brain MRI data, this embodiment designed several comparative and ablation experiments, and used HCS and NDC as evaluation indicators for quantitative analysis. HCS measures the consistency of regional brain age difference distributions between the healthy control training set and the unseen healthy control test set; NDC reflects the degree of brain age acceleration in individual subjects in target disease-related brain regions, while NDCpd-hc, given in Table 1, is an NDC difference index calculated based on Parkinson's disease-related brain regions, defined as the difference between the average NDC of the PD cohort and the average NDC of the HC cohort, used to quantify the model's ability to distinguish between Parkinson's disease and healthy controls.

[0199] As shown in Table 1, when the teacher module is not used and the student model is directly supervised by the overall actual age (row), the HCS is only 17.9% and the NDCpd-hc is 17.5%, indicating poor statistical consistency of regional brain age in healthy controls and poor discrimination ability in PD-related brain regions. After introducing the teacher module and using the initial regional brain age as the distillation target (row), the two indicators improved to 67.5% and 28.6%, respectively, indicating that the process of converting the weak whole-brain label into a regional soft label at the teacher end can significantly improve the prediction quality of regional brain age. On this basis, further adding regional occlusion correction to obtain soft regional brain age (row), the HCS and NDCpd-hc improved to 73.2% and 37.5%, respectively, indicating that the proposed soft regional brain age construction can further alleviate the collapse of regional prediction patterns and strengthen disease relevance. Furthermore, when the FiLM feature modulation or functional consistency loss on the student side was removed (rows, rows), both HCS and NDCpd-hc decreased to varying degrees. In particular, after removing the functional consistency constraint, HCS dropped to 61.2%, indicating that region-cue-driven feature modulation and smoothing constraints based on functional networks both contribute to obtaining a stable and reasonable regional brain age distribution. Overall, this invention achieved the highest HCS and NDCpd-hc indices (rows) with the teacher module, student module, and their associated constraints all enabled, verifying the necessity and synergistic effect of each key module in improving regional brain age prediction performance and brain disease risk assessment capabilities.

[0200] Table 1 Ablation Experiment (%)

[0201]

[0202] As shown in Table 2, the method of this invention can function normally in terms of HCS and NDCpd-hc when using different 3D backbones, but there are significant differences in performance among different backbone structures. In this embodiment, the HCS of M3T, ViT3D, Swin3D, and SwinUNETR, based on the Transformer structure, is only 17.5%–20.6%, and the NDCpd-hc is mostly below 21.0%, with SwinUNETR's NDCpd-hc being only 1.34%. This indicates that although it has strong modeling capabilities, its improvement on the consistency of healthy control distribution and PD-HC discrimination is limited within the framework of regional brain age prediction and risk assessment of this invention. In comparison, AE-FLOW and S3D, both based on CNN architectures, show significant improvements in HCS and NDCpd-hc. S3D achieves an HCS of 65.3%, while NDCpd-hc reaches 28.6%, indicating that the 3D convolutional backbone is more likely to learn stable and disease-related regional brain age patterns in this task. In particular, when using 3D DenseNet as the backbone, the proposed method achieves the highest HCS of 73.2% and NDCpd-hc of 37.5%, outperforming other backbones in both statistical consistency with healthy controls and risk differentiation between Parkinson's disease and healthy controls. Furthermore, the training time of 3D DenseNet is approximately 94.8 minutes, significantly shorter than some Transformer backbones with fewer parameters but longer training times, demonstrating a good efficiency-performance tradeoff. The above results show that the regional brain age prediction and risk assessment framework proposed in this invention has good adaptability to different 3D backbones, but in the preferred embodiment, 3D DenseNet can achieve better overall performance, and therefore can be used as the preferred network structure for the method of this invention.

[0203] Table 2 Comparison of different backbone networks (%)

[0204]

[0205] In summary, this invention proposes a brain disease risk assessment method based on regional brain age prediction. By constructing a teacher-student regional brain age prediction framework on multi-center brain MRI data, the weakly supervised signal of the only available whole-brain actual age is transformed into soft regional brain age labels with spatial resolution and clinical significance. Furthermore, regional cue-driven feature modulation and functional consistency constraints based on functional networks are introduced at the student end, thereby obtaining stable and interpretable high-resolution regional brain age prediction results. Building upon this, this invention further proposes the healthy control similarity index HCS and the neurological disease correlation index NDC, which are used to measure the statistical consistency of regional brain age across different healthy populations and the degree of accelerated aging in target brain disease-related brain regions, respectively, achieving a unified model from regional brain age to brain disease risk assessment.

[0206] Experimental results show that, on real MRI data from multiple centers and multiple devices, the method of this invention can maintain high regional brain age consistency among healthy controls, while significantly amplifying the brain age differences between patients with Parkinson's disease and healthy controls in disease-related brain regions. Compared with control schemes that only use whole-brain age supervision or lack key modules, it achieves significant improvements in indicators such as HCS and NDC, verifying the effectiveness and synergistic effect of technical features such as soft-region brain age construction for teachers, regional modulation for students, and functional consistency constraints. The method of this invention has good versatility for backbone network structures, and its performance is particularly superior under three-dimensional convolutional backbones represented by 3D DenseNet. It is suitable for early risk screening and auxiliary diagnosis of various brain diseases such as Parkinson's disease, Alzheimer's disease, and upper limb motor disorders after stroke, and can provide spatially directional quantitative imaging evidence for clinical practice, with high application value and promising prospects for promotion.

[0207] Figure 4 A block diagram of a brain disease risk assessment device based on regional brain age prediction according to an embodiment of this disclosure is shown. Figure 4 As shown, block diagram 20 includes:

[0208] The brain image segmentation module 21 is used to segment brain regions from brain images of multiple healthy subjects, resulting in multiple brain regions and a brain region mask for each brain region.

[0209] Teacher model training module 22 is used to train a teacher model until convergence based on the brain images of the multiple healthy subjects, and freeze the teacher model parameters after training convergence.

[0210] The regional brain age label generation module 23 is used to construct the regional input of each brain region based on the brain region mask of each brain region, input the regional input of each brain region into the teacher model respectively, obtain the regional brain age of each brain region, and generate regional brain age labels based on the regional brain age of each brain region.

[0211] The student model training module 24 is used to train the student model based on the brain images of the multiple healthy subjects by using the regional brain age labels as distillation supervision signals for the student model.

[0212] The assessment module 25 is used to predict the regional brain age of the subject to be predicted based on the student model, and to assess the risk of the target brain disease based on the prediction results and the set of brain regions related to the target brain disease.

[0213] In some embodiments, the functions or modules of the apparatus provided in this disclosure can be used to perform the methods described in the above method embodiments. The specific implementation can be referred to the description of the above method embodiments, and for the sake of brevity, it will not be repeated here.

[0214] This disclosure also provides a non-volatile computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described method.

[0215] This disclosure also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described method.

[0216] Figure 5 A block diagram of an apparatus 1900 according to an embodiment of the present disclosure is shown. The apparatus 1900 may be provided as a server or a terminal device. (Refer to...) Figure 5 The apparatus 1900 includes a processing component 1922, which further includes one or more processors, and memory resources represented by memory 1932 for storing instructions, such as application programs, that can be executed by the processing component 1922. The application programs stored in memory 1932 may include one or more modules, each corresponding to a set of instructions. Furthermore, the processing component 1922 is configured to execute instructions to perform the methods described above.

[0217] Device 1900 may also include a power supply component 1926 configured to perform power management of device 1900, a wired or wireless network interface 1950 configured to connect device 1900 to a network, and an input / output interface 1958 (I / O interface). Device 1900 can operate on an operating system, such as Windows Server, stored in memory 1932. TM macOS X TM Unix TM Linux TM FreeBSD TM Or similar.

[0218] In an exemplary embodiment, a non-volatile computer-readable storage medium is also provided, such as a memory 1932 including computer program instructions that can be executed by a processing component 1922 of the device 1900 to perform the above-described method.

[0219] Computer-readable storage media can be tangible devices capable of holding and storing programs / instructions used by instruction execution devices. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0220] The computer program (or computer-readable program instructions) described herein can be downloaded from a computer-readable storage medium to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage medium in the respective computing / processing device.

[0221] The computer program (or computer program instructions) used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing state information from the computer-readable program instructions to implement various aspects of this disclosure.

[0222] Various aspects of this disclosure are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0223] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processor of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner; thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.

[0224] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0225] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0226] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for assessing the risk of brain diseases based on regional brain age prediction, characterized in that, include: Brain images from multiple healthy subjects were divided into brain regions to obtain multiple brain regions and brain region masks for each brain region; Based on the brain images of the multiple healthy subjects, the teacher model was trained until convergence, and the parameters of the teacher model were frozen after training convergence. Based on the brain region mask of each brain region, construct the region input of each brain region, input the region input of each brain region into the teacher model respectively, obtain the region brain age of each brain region, and generate the region brain age label based on the region brain age of each brain region. The brain age labels of the aforementioned regions are used as distillation supervision signals for the student model, and the student model is trained based on the brain images of the multiple healthy subjects. Based on the student model, regional brain age is predicted from the brain images of the subjects to be predicted. The risk of the target brain disease is assessed based on the prediction results and the set of brain regions related to the target brain disease.

2. The method according to claim 1, characterized in that, The process of training the teacher model until convergence based on the brain images of the multiple healthy subjects includes: Using the brain images of the healthy subjects as input and the actual age of the healthy subjects as a supervision signal, the teacher model is trained to output a whole-brain predicted brain age; The model parameters are optimized by minimizing the error between the predicted brain age and the actual age until convergence.

3. The method according to claim 2, characterized in that, The brain images of the healthy subjects are three-dimensional images; the construction of the region input for each brain region based on the brain region mask for each brain region includes: For each brain region, the brain image of the healthy subject is multiplied element-wise with the brain region mask of the corresponding brain region to preserve the voxel signal of the corresponding brain region. Replace non-local brain region voxels with noise or constant values ​​to obtain the regional input for each brain region.

4. The method according to claim 3, characterized in that, After obtaining the regional brain age for each brain region, the method further includes: For each brain region, the brain region voxels are replaced with noise or constant values, while the remaining brain regions remain unchanged, to obtain the occluded brain image. The occluded brain image is input into the teacher model to obtain the predicted brain age after occlusion; Based on the predicted brain age after occlusion and the predicted brain age of the whole brain, a region correction signal for each brain region is obtained.

5. The method according to claim 4, characterized in that, The process of generating regional brain age labels based on the regional brain age of each brain region includes: The regional brain age of each brain region is combined with the regional correction signal of each brain region to generate a regional brain age label.

6. The method according to claim 1, characterized in that, The step of using the regional brain age label as a distillation supervision signal for the student model, and training the student model based on brain images of the multiple healthy subjects, includes: With the goal of minimizing distillation loss and functional consistency loss, the student model is trained using brain images from the multiple healthy subjects; the distillation loss is used to make the student model approximate the regional brain age label, and the functional consistency loss is used to suppress non-smooth mutations in regional brain age within the same functional network.

7. The method according to claim 1, characterized in that, The parameters of the assessment include the neurological disease-related index (NDC), and the formula for NDC is as follows: in, For regional brain age difference, n represents the nth subject to be predicted. For the relevant brain regions, For the predicted results, The actual age of the subject to be predicted. A collection of brain regions associated with the target brain disease.

8. A brain disease risk assessment device based on regional brain age prediction, characterized in that, The device includes: The brain image segmentation module is used to segment brain regions from brain images of multiple healthy subjects, resulting in multiple brain regions and a brain region mask for each region. The teacher model training module is used to train the teacher model until convergence based on the brain images of the multiple healthy subjects, and freeze the teacher model parameters after the training converges. The regional brain age label generation module is used to construct the regional input of each brain region based on the brain region mask of each brain region, input the regional input of each brain region into the teacher model respectively, obtain the regional brain age of each brain region, and generate regional brain age labels based on the regional brain age of each brain region. The student model training module is used to train the student model based on the brain images of the multiple healthy subjects by using the brain age labels of the region as distillation supervision signals for the student model. The assessment module is used to predict regional brain age based on the student model of the subject to be predicted, and to assess the risk of the target brain disease based on the prediction results and the set of brain regions related to the target brain disease.

9. A non-volatile computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer program product comprising a computer program, or a non-volatile computer-readable storage medium carrying a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images