Brain age prediction and mental disease evaluation method and system based on multi-modal constraint

Abstract

The application discloses a brain age prediction and mental disease evaluation method and system based on multi-modal constraints, and belongs to the technical field of cross artificial intelligence and medical image analysis. The application constructs an attention gate and a TransformerU-Net model, including a U-Net encoder-decoder network with an attention gate structure and an age prediction module. Multi-scale attention maps are generated by fusing the encoder and decoder features through a skip connection. The age prediction module outputs a predicted brain age through pooling alignment, Transformer global dependence modeling and linear regression. The application is trained by adopting a five-fold cross-validation strategy, an AdamW optimizer and multi-dimensional data enhancement, and performs benchmark testing, simulation training, reliability analysis, attention map atlas and pathological image alignment verification and cross-diagnosis evaluation in sequence. The application realizes high-precision brain age prediction, has cross-modal robustness and strong biological interpretability, and can assist in mental disease risk evaluation.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and medical image analysis, and in particular to a method and system for predicting brain age and assessing mental illness based on multimodal constraints. Background Technology

[0002] Brain age prediction is a technique that estimates the biological age of the brain based on structural magnetic resonance imaging (MRI) images of the brain using machine learning models. The difference between the predicted value and the actual age has been shown to be significantly associated with various neuropsychiatric diseases and neurodegenerative diseases. Kaufmann et al. used the XGBoost algorithm to build a brain age prediction model based on manually extracted low-dimensional structural features, revealing the heritability of brain age differences and their pleiotropic genetic effects on common brain diseases, but it cannot directly learn high-dimensional spatial information from 3D images. Peng et al. proposed a lightweight 3D fully convolutional network, SFCN, which achieves high-precision prediction through end-to-end training. However, this model is based on a simple stacked architecture and lacks an efficient cross-layer feature fusion mechanism, limiting its modeling ability when capturing subtle brain structural degeneration patterns. Rajabli et al. built the SFCN-reg model based on VGG-16 and introduced a deep data augmentation strategy, reducing the MAE of the external dataset by 47%. However, its blocks are only connected by simple sequential connections, lacking dense connections or feature reuse mechanisms, resulting in a bottleneck in high-order nonlinear modeling capabilities.

[0003] Furthermore, while the Grad-CAM visualization technique proposed by Selvaraju et al. can provide spatial localization explanations for CNN models, its depth relies heavily on the feature map of the last convolutional layer. Its spatial resolution is inherently limited by the dimensionality of this layer, making it difficult to capture pixel-level details or subtle pathological features. Moreover, it lacks evidence directly linking attention mechanisms to molecular pathology. Current brain age prediction models face significant challenges in clinical deployment: research-grade high-resolution models applied to routine clinical scans can increase prediction errors by up to 11.5 years; models exhibit high variance and insufficient stability under scanner fluctuations; and the "black box" nature of deep learning hinders clinical trust. Therefore, there is an urgent need for a brain age prediction model that can effectively generalize to low-resolution routine clinical scans and possesses high robustness and reliability. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for brain age prediction and mental illness assessment based on multimodal constraints, in order to solve the technical problems of existing deep learning brain age prediction models having high accuracy on high-resolution scientific research data but failing to generalize on low-resolution clinical images, having poor model stability, and limited biological interpretability, and to construct an AI model that can effectively generalize to low-resolution routine clinical scans and has high robustness and reliability.

[0005] To achieve the above objectives, this invention provides a method for predicting brain age and assessing mental illness based on multimodal constraints, comprising the following steps: S1. Collect multicenter brain MRI research datasets, and sequentially perform bias field correction, skull dissection, nonlinear spatial standardization, unified resampling, and intensity normalization preprocessing to obtain standardized brain image data. S2. Based on standardized brain imaging data, an attention-gated and Transformer U-Net model is constructed, including a U-Net encoder-decoder network with an attention-gated structure and an age prediction module. The U-Net encoder-decoder network with an attention-gated structure fuses encoder features and upsampled decoder features through skip connections via an attention-gated mechanism to generate a multi-scale attention map. The age prediction module performs pooling alignment, Transformer global dependency modeling, and linear regression on the multi-stage features of the decoder to output the predicted brain age. S3. The attention gate and TransformerU-Net model are trained using a five-fold cross-validation strategy with mean squared error as the loss function, and the AdamW optimizer is used to configure the learning rate strategy. Multidimensional data augmentation is introduced, and the model is trained on standardized brain image data to obtain the trained brain age prediction model. S4. Based on the trained brain age prediction model, perform benchmark tests, simulation training, reliability analysis, attention map and pathological image alignment verification, and cross-diagnostic assessment, and output brain age prediction values ​​and mental illness risk assessment results.

[0006] Preferably, the dataset in step S1 includes: The dataset integrates a multicenter publicly available research database and a single-site dataset, along with a model training dataset containing no fewer than 10,000 healthy control subjects; a reliability assessment dataset including 676 scans from 11 independent test-retest datasets; a cross-diagnostic validation dataset integrating the LEMON, ADNI, and Anding datasets; and a biointerpretation cross-modal dataset containing 3D T1-weighted images, amyloid β-PET, and fluorodeoxyglucose PET images.

[0007] Preferably, the preprocessing in step S1 specifically includes: The N4ITK algorithm was used for bias field correction; the ANTS tool was used for skull dissection; nonlinear spatial normalization was performed to the ICBM2009c asymmetric MNI space using OASIS30ANTS as the target template; uniform resampling was performed to a 91×109×91 voxel matrix; and the image gray values ​​were normalized to the [0,1] interval.

[0008] Preferably, the workflow of the attention gate mechanism in step S2 is as follows: The encoded features and upsampled decoder features are respectively subjected to 3D convolution, followed by element-wise summation, ReLU activation, 3D convolutional compression, batch normalization, and Sigmoid activation to obtain attention coefficients. These attention coefficients are then multiplied element-wise with the encoder features to obtain a refined feature representation. Finally, the attention maps from all levels are upsampled to a uniform spatial resolution and summed to obtain an attention map aggregating multi-scale regions of interest. It is represented as: ; in, This represents the attention weight tensor of the k-th layer. This represents the trilinear interpolation operation, and N represents the total number of skip connection levels.

[0009] Preferably, in step S2, the age prediction module first achieves alignment and reconstruction of heterogeneous feature maps in terms of spatial resolution and channel dimension through adaptive average pooling and 3D convolution; then, it uses a dimension permutation operation to map the spatial tensor into a serialized representation and performs cross-level feature concatenation along the channel dimension; the fused representation is modeled globally for dependencies through a Transformer architecture, and features are dynamically filtered and long-range context enhanced through a self-attention mechanism; the enhanced feature vector is restored to the original spatial topology configuration and mapped to the linear regression space to complete physiological age estimation.

[0010] Preferably, in step S3, the five-fold cross-validation strictly isolates the training set and the validation set at the subject level; the AdamW optimizer has an initial learning rate of 0.0007 and a weight decay coefficient of 0.01; the learning rate is warmed up for the first 5 epochs, and then the ReduceLROnPlateau scheduler monitors the validation set loss. If the performance stagnates for 5 consecutive epochs, the learning rate decays in a stepwise manner at a rate of 0.95, and the lower limit of the learning rate is set to 0.0000001.

[0011] Preferably, in step S3, the multidimensional data augmentation scheme includes spatial dimension transformation and intensity dimension perturbation. Spatial dimension transformation includes random flipping, rotation, and affine transformation, while intensity dimension perturbation includes noise injection and intensity shift. The model performance is evaluated using mean squared error to measure prediction accuracy. , represented as: ; in, and Let represent the actual age and predicted brain age of the i-th subject, respectively, and N represent the number of subjects. Pearson correlation coefficient was used to assess linear association, coefficient of determination was used to quantify the proportion of variance explained by the regression model, and within-group correlation coefficient was used to assess Test-Retest reliability.

[0012] Preferably, in step S4, the clinical style simulation training simulates the volume of thick-slice MRI in clinical practice by averaging slices of high-resolution T1 images, artificially reducing transillumination resolution, and constructing an ATUN-S variant model optimized for low-quality images. End-to-end learning is then performed on paired simulation datasets. Reliability analysis uses intra-group correlation coefficients to quantify the consistency of brain age estimation for the same subject across different imaging sessions. Represented as: ; in, Indicates the mean square among subjects. The mean square between scans is represented by , and k represents the number of repeated scans. Spatial alignment validation of attention maps and molecular pathological images was performed using a subset of ADNI images. Spatial correlation analysis was conducted between the attention maps generated by the model and paired amyloid β-PET and fluorodeoxyglucose PET images. The Dice similarity coefficient and Jaccard index were used to quantify the spatial overlap between the model's region of interest and the neuropathological lesion. The Dice coefficient is expressed as: ; The Jaccard index is expressed as: ; Where A represents the attention map and B represents the region of interest derived from PET.

[0013] Preferably, in step S4, the cross-diagnostic assessment and modal adaptability validation perform adaptation experiments on two heterogeneous sequences, T1-weighted images and FLAIR. Cross-site migration is achieved by performing finite parameter fine-tuning through unfreezing the bottom convolutional layers. Weighted binary cross-entropy loss is used to address class imbalance. Continuous age regression outputs are mapped to disease risk probabilities to validate the model's ability to assist in decision-making in the diagnosis of Alzheimer's disease and major depressive disorder.

[0014] Preferred, a multimodal constraint-based brain age prediction and mental illness assessment system includes: The data acquisition module is used to collect multicenter brain MRI research datasets and sequentially performs bias field correction, skull dissection, nonlinear spatial standardization, unified resampling, and intensity normalization preprocessing to obtain standardized brain image data. The model building module is used to construct attention-gated and Transformer U-Net models based on standardized brain imaging data. It includes a U-Net encoder-decoder network with attention-gated structure and an age prediction module. The U-Net encoder-decoder network with attention-gated structure fuses encoder features and upsampled decoder features through skip connections via the attention-gated mechanism to generate a multi-scale attention map. The age prediction module performs pooling alignment, Transformer global dependency modeling, and linear regression on the multi-stage features of the decoder to output the predicted brain age. The model training module is used to train the attention gate and TransformerU-Net model using a five-fold cross-validation strategy, with mean squared error as the loss function, and the AdamW optimizer to configure the learning rate strategy. Multidimensional data augmentation is introduced to train the model on standardized brain image data to obtain the trained brain age prediction model. The prediction and evaluation module is used to perform benchmark tests, simulation training, reliability analysis, attention map and pathological image alignment verification, and cross-diagnostic assessment based on the trained brain age prediction model, and output brain age prediction values ​​and mental illness risk assessment results.

[0015] The advantages and beneficial effects of this invention compared to the prior art are: 1. The attention gate and Transformer U-Net model proposed in this invention achieve extremely low mean absolute error in a large-scale multicenter study cohort. The predicted values ​​have a very high Pearson correlation coefficient with the actual age, and the coefficient of determination is close to complete explanation. It is significantly better than the existing benchmark model and effectively overcomes the nonlinear bias problem that traditional models are prone to when predicting across age groups. It provides a highly reliable quantitative benchmark for subsequent clinical auxiliary diagnosis.

[0016] 2. This invention constructs an ATUN-S variant model through clinical style simulation training based on simulated degeneration. On conventional clinical FLAIR images, the predicted values ​​maintain a strong linear correlation with actual age and are highly consistent with the predicted values ​​of high-resolution models. It successfully bridges the performance gap between research and clinical levels. The model can still fully retain the predictive validity and biological sensitivity of brain aging trajectory, and has the technical prerequisites for large-scale deployment in real medical environments.

[0017] 3. In the extensive scanning validation of multiple independent test-retest datasets, the intragroup correlation coefficient of the vast majority of datasets exceeded the high confidence threshold. ATUN-S achieved a high Pearson correlation coefficient in two-time point retests and maintained a very high pairwise correlation in three-time point longitudinal tracking. It effectively suppressed the prediction fluctuations caused by scanner noise and imaging randomness, and became a highly reliable tool for tracking brain health trajectory and identifying subtle pathological aging trends.

[0018] 4. During training, the attention focus of the model in this invention significantly converges from a non-specific diffuse state to a specific anatomical region characterized by high amyloid protein load and low glucose metabolism. The continuous reduction in prediction error is highly synchronized with the improvement in the model's ability to capture PET-derived pathological patterns. This confirms that the improvement in model accuracy is based on the precise alignment of aging-related molecular pathological regions, providing clinicians with an intuitive basis for model judgment and greatly enhancing the credibility of deep learning models in medical scenarios.

[0019] 5. The ATUN-S model of this invention exhibits a high area under the curve (AUC) on FLAIR images for distinguishing between Alzheimer's disease and healthy controls. It also demonstrates good AUCs for distinguishing between mild cognitive impairment and healthy controls, as well as between diagnosed major depressive disorder and depressive state. This proves that brain age difference is not only sensitive to severe brain atrophy but can also capture subtle and hidden neuroanatomical features in the progression of mental illness. It verifies the model's ability to distinguish between transitional and chronic mental states in real hospital settings, providing objective imaging assessment indicators for clinical personalized treatment planning and disease progression evaluation.

[0020] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0021] Figure 1 This is a flowchart of a brain age prediction and mental illness assessment method based on multimodal constraints in an embodiment of the present invention; Figure 2 This is a structural diagram of the brain age prediction and mental illness assessment system based on multimodal constraints according to an embodiment of the present invention; Figure 3 This is a data distribution diagram of the main dataset in an embodiment of the present invention; Figure 4 This is an overall flowchart of an embodiment of the present invention; Figure 5 This is a performance comparison chart of ATUN-O (research T1w) and ATUN-S (clinical FLAIR) on the LEMON dataset in this embodiment of the invention; Figure 6 This is an evolution diagram of the attention map and error index in an embodiment of the present invention.

[0022] Specific implementation method: Transformer U-Net model In the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product is in use. They are used only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," and "connect" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0023] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0024] like Figure 1 As shown, this invention provides a method for predicting brain age and assessing mental illness based on multimodal constraints, comprising the following steps: S1. Collect multicenter brain MRI research datasets, and sequentially perform bias field correction, skull dissection, nonlinear spatial standardization, unified resampling, and intensity normalization preprocessing to obtain standardized brain image data. S2. Based on standardized brain imaging data, an attention-gated and Transformer U-Net model is constructed, including a U-Net encoder-decoder network with an attention-gated structure and an age prediction module. The U-Net encoder-decoder network with an attention-gated structure fuses encoder features and upsampled decoder features through skip connections via an attention-gated mechanism to generate a multi-scale attention map. The age prediction module performs pooling alignment, Transformer global dependency modeling, and linear regression on the multi-stage features of the decoder to output the predicted brain age. S3. The attention gate and TransformerU-Net model are trained using a five-fold cross-validation strategy with mean squared error as the loss function, and the AdamW optimizer is used to configure the learning rate strategy. Multidimensional data augmentation is introduced, and the model is trained on standardized brain image data to obtain the trained brain age prediction model. S4. Based on the trained brain age prediction model, perform benchmark tests, simulation training, reliability analysis, attention map and pathological image alignment verification, and cross-diagnostic assessment, and output brain age prediction values ​​and mental illness risk assessment results.

[0025] Preferably, the dataset in step S1 includes: The dataset integrates a multicenter publicly available research database and a single-site dataset, along with a model training dataset containing no fewer than 10,000 healthy control subjects; a reliability assessment dataset including 676 scans from 11 independent test-retest datasets; a cross-diagnostic validation dataset integrating the LEMON, ADNI, and Anding datasets; and a biointerpretation cross-modal dataset containing 3D T1-weighted images, amyloid β-PET, and fluorodeoxyglucose PET images.

[0026] In some embodiments, the model training dataset integrates a multi-center publicly available research database and a single-site dataset, containing no fewer than 10,000 healthy control subjects, to establish a reference standard for normal brain aging; the reliability assessment dataset includes 676 scans from 11 independent test-retest datasets to assess model stability; the cross-diagnostic validation dataset integrates the LEMON, ADNI, and Anding datasets, covering different age groups and disease spectrums; and the biointerpretation cross-modal dataset collects three-dimensional T1-weighted images, amyloid β-PET, and fluorodeoxyglucose PET images to validate the biological effectiveness of the model's attention map.

[0027] Preferably, the preprocessing in step S1 specifically includes: The N4ITK algorithm was used for bias field correction; the ANTS tool was used for skull dissection; nonlinear spatial normalization was performed to the ICBM2009c asymmetric MNI space using OASIS30ANTS as the target template; uniform resampling was performed to a 91×109×91 voxel matrix; and the image gray values ​​were normalized to the [0,1] interval.

[0028] In some embodiments, the N4ITK algorithm is used for bias field correction to eliminate low-frequency intensity variations in MRI images caused by radio frequency field inhomogeneity; the ANTS tool is used for skull dissection to automatically identify and remove non-brain tissues such as skull and scalp; nonlinear spatial normalization is performed to ICBM2009c asymmetric MNI space using OASIS30ANTS as the target template, and elastic deformation is used to ensure that the corresponding brain regions of different subjects are located in the same voxel coordinates; uniform resampling is performed to a 91×109×91 voxel matrix with an isotropic spatial resolution of 2mm to ensure consistent spatial accuracy in all directions; and the image grayscale values ​​are normalized to the [0,1] range to eliminate intensity differences caused by different scanners.

[0029] Preferably, the workflow of the attention gate mechanism in step S2 is as follows: The encoded features and upsampled decoder features are respectively subjected to 3D convolution, followed by element-wise summation, ReLU activation, 3D convolutional compression, batch normalization, and Sigmoid activation to obtain attention coefficients. These attention coefficients are then multiplied element-wise with the encoder features to obtain a refined feature representation. Finally, the attention maps from all levels are upsampled to a uniform spatial resolution and summed to obtain an attention map aggregating multi-scale regions of interest. It is represented as: ; in, This represents the attention weight tensor of the k-th layer. This represents the trilinear interpolation operation, and N represents the total number of skip connection levels.

[0030] In some embodiments, the encoded features and upsampled decoder features are respectively convolved in three dimensions and then the corresponding elements are added together. Then, ReLU activation, three-dimensional convolution compression, batch normalization, and Sigmoid activation are performed to obtain attention coefficients. These coefficients are spatial weight tensors with values ​​ranging from 0 to 1, representing the degree of attention of the model at each spatial location. The attention coefficients are multiplied element-wise with the encoder features to obtain a refined feature representation, realizing spatial attention weighting. The attention maps of all levels are upsampled to a uniform spatial resolution and then summed to obtain an attention map that aggregates multi-scale attention regions.

[0031] Preferably, in step S2, the age prediction module first achieves alignment and reconstruction of heterogeneous feature maps in terms of spatial resolution and channel dimension through adaptive average pooling and 3D convolution; then, it uses a dimension permutation operation to map the spatial tensor into a serialized representation and performs cross-level feature concatenation along the channel dimension; the fused representation is modeled globally for dependencies through a Transformer architecture, and features are dynamically filtered and long-range context enhanced through a self-attention mechanism; the enhanced feature vector is restored to the original spatial topology configuration and mapped to the linear regression space to complete physiological age estimation.

[0032] In some embodiments, firstly, the heterogeneous feature maps are aligned and reconstructed in terms of spatial resolution and channel dimension through adaptive average pooling and 3D convolution; then, the spatial tensor is mapped to a serialized representation using a dimension permutation operation, and cross-level feature concatenation is performed along the channel dimension; the fused representation is modeled globally for dependencies through a Transformer architecture, and dynamic feature selection and long-range context enhancement are achieved through a self-attention mechanism; the enhanced feature vector is restored to the original spatial topology configuration and mapped to the linear regression space to complete physiological age estimation.

[0033] Preferably, in step S3, the five-fold cross-validation strictly isolates the training set and the validation set at the subject level; the AdamW optimizer has an initial learning rate of 0.0007 and a weight decay coefficient of 0.01; the learning rate is warmed up for the first 5 epochs, and then the ReduceLROnPlateau scheduler monitors the validation set loss. If the performance stagnates for 5 consecutive epochs, the learning rate decays in a stepwise manner at a rate of 0.95, and the lower limit of the learning rate is set to 0.0000001.

[0034] In some embodiments, five-fold cross-validation strictly isolates the training set and validation set at the subject level to avoid data leakage; the AdamW optimizer has an initial learning rate of 0.0007 and a weight decay coefficient of 0.01. The learning rate is warmed up for the first 5 epochs, and then the ReduceLROnPlateau scheduler monitors the validation set loss. If the performance stagnates for 5 consecutive epochs, it decays in a stepwise manner at a rate of 0.95, with a lower limit of 0.0000001, to ensure that the model parameters converge optimally under massive heterogeneous data.

[0035] Preferably, in step S3, the multidimensional data augmentation scheme includes spatial dimension transformation and intensity dimension perturbation. Spatial dimension transformation includes random flipping, rotation, and affine transformation, while intensity dimension perturbation includes noise injection and intensity shift. The model performance is evaluated using mean squared error to measure prediction accuracy. , represented as: ; in, and Let represent the actual age and predicted brain age of the i-th subject, respectively, and N represent the number of subjects. Pearson correlation coefficient was used to assess linear association, coefficient of determination was used to quantify the proportion of variance explained by the regression model, and within-group correlation coefficient was used to assess Test-Retest reliability.

[0036] In some embodiments, multidimensional data augmentation encompasses spatial dimension transformation and intensity dimension perturbation. The former includes random flipping, rotation, and affine transformation, while the latter includes noise injection and intensity shift, to improve the model's robustness to heterogeneous clinical environments. Model performance is evaluated using mean squared error to measure prediction accuracy, Pearson correlation coefficient to assess linear association, coefficient of determination to quantify the proportion of variance explained by the regression model, and within-group correlation coefficient to assess Test-Retest reliability.

[0037] Preferably, in step S4, the clinical style simulation training simulates the volume of thick-slice MRI in clinical practice by averaging slices of high-resolution T1 images, artificially reducing transillumination resolution, and constructing an ATUN-S variant model optimized for low-quality images. End-to-end learning is then performed on paired simulation datasets. Reliability analysis uses intra-group correlation coefficients to quantify the consistency of brain age estimation for the same subject across different imaging sessions. Represented as: ; in, Indicates the mean square among subjects. The mean square between scans is represented by , and k represents the number of repeated scans. Spatial alignment validation of attention maps and molecular pathological images was performed using a subset of ADNI images. Spatial correlation analysis was conducted between the attention maps generated by the model and paired amyloid β-PET and fluorodeoxyglucose PET images. The Dice similarity coefficient and Jaccard index were used to quantify the spatial overlap between the model's region of interest and the neuropathological lesion. The Dice coefficient is expressed as: ; The Jaccard index is expressed as: ; Where A represents the attention map and B represents the region of interest derived from PET.

[0038] In some embodiments, clinical style simulation training simulates the volume of thick-slice MRI in clinical settings by averaging slices of high-resolution T1 images, artificially reducing transillumination resolution, and constructing an ATUN-S variant model optimized for low-quality images. End-to-end learning is performed on paired simulated datasets, enabling the model to master the ability to extract aging features from mathematically downsampled research data to adapt to the low-resolution clinical environment. Reliability analysis uses intragroup correlation coefficients to quantify the consistency of brain age estimates for the same subject across different imaging sessions. This coefficient assesses reliability by comparing inter-subject variability with measurement error.

[0039] Preferably, in step S4, the cross-diagnostic assessment and modal adaptability validation perform adaptation experiments on two heterogeneous sequences, T1-weighted images and FLAIR. Cross-site migration is achieved by performing finite parameter fine-tuning through unfreezing the bottom convolutional layers. Weighted binary cross-entropy loss is used to address class imbalance. Continuous age regression outputs are mapped to disease risk probabilities to validate the model's ability to assist in decision-making in the diagnosis of Alzheimer's disease and major depressive disorder.

[0040] In some embodiments, using a subset of ADNI images, the attention map generated by the model is spatially correlated with paired amyloid β-PET images and fluorodeoxyglucose PET images. The Dice similarity coefficient and Jaccard index are used to quantify the spatial overlap between the model’s focus area and neuropathological lesions. The model’s focus on areas with real pathological significance is verified by comparing the spatial distribution of the two.

[0041] Adaptation experiments were conducted on two heterogeneous sequences, T1-weighted images and FLAIR. Cross-site transfer was achieved by finite parameter fine-tuning through unfreezing the bottom convolutional layers. Weighted binary cross-entropy loss was used to address class imbalance. Continuous age regression outputs were mapped to disease risk probabilities to verify the model's ability to assist in decision-making in the diagnosis of Alzheimer's disease and major depressive disorder. T1-weighted images clearly show anatomical structures, while FLAIR makes lesions clearer by suppressing cerebrospinal fluid signals. The contrast mechanisms of the two are different.

[0042] Preferred, such as Figure 2 As shown, the brain age prediction and mental illness assessment system based on multimodal constraints includes: The data acquisition module is used to collect multicenter brain MRI research datasets and sequentially performs bias field correction, skull dissection, nonlinear spatial standardization, unified resampling, and intensity normalization preprocessing to obtain standardized brain image data. The model building module is used to construct attention-gated and Transformer U-Net models based on standardized brain imaging data. It includes a U-Net encoder-decoder network with attention-gated structure and an age prediction module. The U-Net encoder-decoder network with attention-gated structure fuses encoder features and upsampled decoder features through skip connections via the attention-gated mechanism to generate a multi-scale attention map. The age prediction module performs pooling alignment, Transformer global dependency modeling, and linear regression on the multi-stage features of the decoder to output the predicted brain age. The model training module is used to train the attention gate and TransformerU-Net model using a five-fold cross-validation strategy, with mean squared error as the loss function, and the AdamW optimizer to configure the learning rate strategy. Multidimensional data augmentation is introduced to train the model on standardized brain image data to obtain the trained brain age prediction model. The prediction and evaluation module is used to perform benchmark tests, simulation training, reliability analysis, attention map and pathological image alignment verification, and cross-diagnostic assessment based on the trained brain age prediction model, and output brain age prediction values ​​and mental illness risk assessment results.

[0043] The following verification is based on a specific embodiment.

[0044] To verify the effectiveness of the method of the present invention, the following practical application scenarios were used for testing: Application Scenario: The neurology department of a top-tier hospital plans to deploy a brain age prediction model in its clinical auxiliary diagnostic system for early screening of Alzheimer's disease and auxiliary identification of severe depression. The hospital's routine clinical MRI scans use a 1.5T MRI scanner with a T1-weighted slice thickness of 4mm and a FLAIR sequence slice thickness of 5mm, which is significantly different from research-grade 3T high-resolution scans.

[0045] S1. Data Preparation and Preprocessing: Model training data were integrated from multi-center databases, including publicly available databases such as ABIDEI&II, ADHD-200, CORR, and FCP, as well as single-site data from OASIS-3, BGSP, and NKI-RS, totaling 10,748 healthy control subjects aged 5.1–97 years. Eleven independent test-retest datasets (DS1–DS11) containing 676 scans were included for reliability assessment. Cross-diagnostic validation used the LEMON dataset (228 subjects), the ADNI dataset (261 subjects, including 76 healthy controls, 63 MCIc subjects, 72 MCInc subjects, and 50 AD subjects), and the Anding dataset (280 subjects, including 143 MDD subjects and 137 DEP subjects). Biointerpretation data were obtained from 335 subjects in the ADNI dataset who simultaneously had T1, amyloid β-PET, and FDG-PET images. All images underwent N4ITK bias field correction, ANTS craniotomy, and nonlinear spatial normalization to the ICBM2009c asymmetric MNI space using OASIS30ANTS as the target template. All images were uniformly resampled to a 91×109×91 voxel matrix, and intensity was normalized to the [0,1] interval. The age distribution of the data is as follows: Figure 3 As shown.

[0046] S2. Model Construction: Construct an attention-gated and Transformer U-Net model. The U-Net encoder-decoder network with attention gates takes a 91×109×91 3D image as input, performs convolution and downsampling to output an 8×8×8 feature representation, and then reconstructs it to the original 91×109×91 spatial resolution. Encoder features and upsampled decoder features are fused through skip connections using an attention gate mechanism. The attention coefficient α is obtained through 3D convolution, ReLU activation, 3D convolution compression, batch normalization, and Sigmoid activation, and is element-wise multiplied with the encoder features to obtain refined features. All levels of attention maps are upsampled to 91×109×91 and summed to obtain an aggregated multi-scale attention map. The age prediction module sequentially performs adaptive average pooling, 3D convolution alignment, dimension permutation serialization, cross-level feature concatenation, Transformer global dependency modeling, spatial topology restoration, and linear regression on the five-stage features of the decoder, outputting a predicted brain age. The overall model flow and data direction are as follows: Figure 4 As shown.

[0047] S3. Model Training: A five-fold cross-validation strategy is employed to strictly isolate the training and validation sets at the subject level, ensuring that the same subject's image does not appear in both subsets simultaneously. Mean squared error is used as the loss function. The AdamW optimizer has an initial learning rate of 0.0007 and a weight decay of 0.01. The first 5 epochs are used for learning rate warm-up, after which the ReduceLROnPlateau scheduler monitors the validation set loss. If the loss stagnates for 5 consecutive epochs, it decays in a stepwise manner at a rate of 0.95, with a lower limit of 0.0000001. Multi-dimensional data augmentation is introduced during the training phase, including random flipping, rotation, affine transformation, noise injection, and intensity shifting. A single training run is capped at 200 epochs and is executed in parallel on a high-performance computing platform equipped with 8 NVIDIA RTX 3090 GPUs.

[0048] S4. Verification and Application: (1) Benchmark test of high-resolution research data: ATUN-O achieved MAE=2.50±0.02 years on 1mm isotropic high-resolution T1-weighted images, with a Pearson correlation coefficient of r=0.988 between the predicted value and the actual age, and a coefficient of determination R 2 =0.975, significantly better than benchmark models such as SFCN (MAE=2.83 years) and SQET (MAE=2.54 years).

[0049] (2) Clinical style simulation training: A 4:1 slice averaging operation was performed on high-resolution T1 images to simulate the volume of thick-slice MRI in clinical settings, constructing the ATUN-S variant model. On paired images in the LEMON dataset, ATUN-O showed a MAE of 3.36 years on high-resolution T1 images, while ATUN-S showed a MAE of 5.04 years on clinical FLAIR images. The predicted values ​​were highly consistent (r=0.950), and the predicted value on FLAIR was consistent with the actual age (r=0.948). The differences in brain age distribution between models, the scatter plot of predicted age versus chronological age, and the correlation of predicted age are shown below. Figure 5 As shown.

[0050] (3) Test-Retest reliability analysis: In 676 scans of 11 datasets from DS1 to DS11, the ICC was >0.90 for 9 datasets and >0.80 for all cohorts. The ATUN-S two-timepoint test-retest ratio was r=0.955 (p<<0.001), and the pairwise correlations at the three time points remained between 0.912 and 0.940, which was significantly higher than the 0.841-0.882 of ATUN-O.

[0051] (4) Validation of the alignment between attention maps and pathology: In the early stages of training (Epoch 0), attention was diffuse and nonspecific. As optimization progressed to Epoch 40, it significantly converged to a specific anatomical region in the brain characterized by high amyloid β load and low glucose metabolism. The prediction error MAE showed a strong negative correlation with the Aβ-PET modality Dice similarity coefficient (r = -0.80) and a high negative correlation with the FDG-PET modality Dice score (r = -0.74). The attention maps of selected healthy subjects during training, the changes in Dice scores and Jaccard coefficients of attention maps at different training periods with Aβ-rSUV, the changes in Dice scores and Jaccard coefficients with FDG-rSUV, the changes in loss and MAE are shown in the figure. Figure 6 As shown.

[0052] (5) Cross-diagnostic assessment and modal adaptability validation: ATUN-S achieved an AUC of 0.904 in distinguishing between Alzheimer's disease and healthy controls on FLAIR images, an AUC of 0.735 in distinguishing between healthy controls and mild cognitive impairment, and an AUC of 0.819 in distinguishing between Alzheimer's disease and mild cognitive impairment; the AUC in distinguishing between diagnosed major depressive disorder and depressive state was 0.741. The average performance of the ATUN model on FLAIR and T1 imaging modalities in different classification tasks is shown in Table 3.

[0053] The above embodiments demonstrate that the method of the present invention possesses high-precision prediction, strong cross-modal robustness, high test-retest reliability, and good biological interpretability in real clinical settings, and can effectively assist in the early screening of Alzheimer's disease and the differential diagnosis of major depressive disorder.

[0054] In this application, unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. In case of any inconsistency, the meaning set forth in this specification or derived from the content described herein shall prevail. Furthermore, the terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit the scope of this application.

[0055] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for predicting brain age and assessing mental illness based on multimodal constraints, characterized in that, Includes the following steps: S1. Collect multicenter brain MRI research datasets, and sequentially perform bias field correction, skull dissection, nonlinear spatial standardization, unified resampling, and intensity normalization preprocessing to obtain standardized brain image data. S2. Based on standardized brain imaging data, an attention-gated and Transformer U-Net model is constructed, including a U-Net encoder-decoder network with an attention-gated structure and an age prediction module. The U-Net encoder-decoder network with an attention-gated structure fuses encoder features and upsampled decoder features through skip connections via an attention-gated mechanism to generate a multi-scale attention map. The age prediction module performs pooling alignment, Transformer global dependency modeling, and linear regression on the multi-stage features of the decoder to output the predicted brain age. S3. The attention gate and TransformerU-Net model are trained using a five-fold cross-validation strategy with mean squared error as the loss function, and the AdamW optimizer is used to configure the learning rate strategy. Multidimensional data augmentation is introduced, and the model is trained based on standardized brain imaging data to obtain the trained brain age prediction model. S4. Based on the trained brain age prediction model, perform benchmark tests, simulation training, reliability analysis, attention map and pathological image alignment verification, and cross-diagnostic assessment, and output brain age prediction values ​​and mental illness risk assessment results.

2. The method for brain age prediction and mental illness assessment based on multimodal constraints according to claim 1, characterized in that, The dataset in step S1 includes: The dataset integrates a multicenter publicly available research database and a single-site dataset, along with a model training dataset containing no fewer than 10,000 healthy control subjects; a reliability assessment dataset including 676 scans from 11 independent test-retest datasets; a cross-diagnostic validation dataset integrating the LEMON, ADNI, and Anding datasets; and a biointerpretation cross-modal dataset containing 3D T1-weighted images, amyloid β-PET, and fluorodeoxyglucose PET images.

3. The method for brain age prediction and mental illness assessment based on multimodal constraints according to claim 2, characterized in that, The preprocessing in step S1 specifically includes: The N4ITK algorithm was used for bias field correction; the ANTS tool was used for skull dissection; nonlinear spatial normalization was performed to the ICBM2009c asymmetric MNI space using OASIS30ANTS as the target template; uniform resampling was performed to a 91×109×91 voxel matrix; and the image gray values ​​were normalized to the [0,1] interval.

4. The method for brain age prediction and mental illness assessment based on multimodal constraints according to claim 3, characterized in that, The workflow of the attention gate mechanism in step S2 is as follows: The encoded features and upsampled decoder features are respectively subjected to 3D convolution, followed by element-wise summation, ReLU activation, 3D convolutional compression, batch normalization, and Sigmoid activation to obtain attention coefficients. These attention coefficients are then multiplied element-wise with the encoder features to obtain a refined feature representation. Finally, the attention maps from all levels are upsampled to a uniform spatial resolution and summed to obtain an attention map aggregating multi-scale regions of interest. It is represented as: ； in, This represents the attention weight tensor of the k-th layer. This represents the trilinear interpolation operation, and N represents the total number of skip connection levels.

5. The method for brain age prediction and mental illness assessment based on multimodal constraints according to claim 4, characterized in that, In step S2, the age prediction module first achieves alignment and reconstruction of heterogeneous feature maps in terms of spatial resolution and channel dimension through adaptive average pooling and 3D convolution; then, it uses dimension permutation operation to map the spatial tensor into a serialized representation and performs cross-level feature concatenation along the channel dimension. The fusion representation uses the Transformer architecture to model global dependencies and achieves dynamic feature selection and long-range context enhancement through a self-attention mechanism. The enhanced feature vectors are restored to the original spatial topology configuration and mapped to the linear regression space to complete physiological age estimation.

6. The method for brain age prediction and mental illness assessment based on multimodal constraints according to claim 5, characterized in that, In step S3, the five-fold cross-validation strictly isolates the training set from the validation set at the subject level; the AdamW optimizer has an initial learning rate of 0.0007 and a weight decay coefficient of 0.01; the learning rate is warmed up for the first 5 epochs, and then the ReduceLROnPlateau scheduler monitors the validation set loss. If the performance stagnates for 5 consecutive epochs, the learning rate decays in a stepwise manner at a rate of 0.95, and the lower limit of the learning rate is set to 0.0000001.

7. The method for brain age prediction and mental illness assessment based on multimodal constraints according to claim 6, characterized in that, In step S3, the multidimensional data augmentation scheme encompasses spatial dimension transformation and intensity dimension perturbation. Spatial dimension transformation includes random flipping, rotation, and affine transformation, while intensity dimension perturbation includes noise injection and intensity shift. Model performance is evaluated using mean squared error to measure prediction accuracy. , represented as: ； in, and Let represent the actual age and predicted brain age of the i-th subject, respectively, and N represent the number of subjects. Pearson correlation coefficient was used to assess linear association, coefficient of determination was used to quantify the proportion of variance explained by the regression model, and within-group correlation coefficient was used to assess Test-Retest reliability.

8. The method for predicting brain age and assessing mental illness based on multimodal constraints according to claim 7, characterized in that, In step S4, clinical style simulation training simulates the volume of clinical thick-slice MRI by averaging slices on high-resolution T1 images, artificially reducing transillumination resolution, constructing an ATUN-S variant model optimized for low-quality images, and performing end-to-end learning on paired simulation datasets. Reliability analysis uses the intraclass correlation coefficient to quantify the consistency of brain age estimates for the same subject across different imaging sessions. Represented as: ； in, Indicates the mean square among subjects. The mean square between scans is represented by , and k represents the number of repeated scans. Spatial alignment validation of attention maps and molecular pathological images was performed using a subset of ADNI images. Spatial correlation analysis was conducted between the attention maps generated by the model and paired amyloid β-PET and fluorodeoxyglucose PET images. The Dice similarity coefficient and Jaccard index were used to quantify the spatial overlap between the model's region of interest and the neuropathological lesion. The Dice coefficient is expressed as: ； The Jaccard index is expressed as: ； Where A represents attention mapping and B represents the region of interest derived from PET.

9. The method for predicting brain age and assessing mental illness based on multimodal constraints according to claim 8, characterized in that, In step S4, cross-diagnostic assessment and modal adaptability validation are performed on two heterogeneous sequences, T1-weighted images and FLAIR. Cross-site transfer is achieved by finite parameter fine-tuning through unfreezing the bottom convolutional layers. Weighted binary cross-entropy loss is used to address class imbalance. Continuous age regression outputs are mapped to disease risk probabilities to validate the model's ability to assist in decision-making in the diagnosis of Alzheimer's disease and major depressive disorder.

10. A brain age prediction and mental illness assessment system based on multimodal constraints, characterized in that, include: The data acquisition module is used to collect multicenter brain MRI research datasets and sequentially performs bias field correction, skull dissection, nonlinear spatial standardization, unified resampling, and intensity normalization preprocessing to obtain standardized brain image data. The model building module is used to construct attention-gated and Transformer U-Net models based on standardized brain imaging data. It includes a U-Net encoder-decoder network with attention-gated structure and an age prediction module. The U-Net encoder-decoder network with attention-gated structure fuses encoder features and upsampled decoder features through skip connections via the attention-gated mechanism to generate a multi-scale attention map. The age prediction module performs pooling alignment, Transformer global dependency modeling, and linear regression on the multi-stage features of the decoder to output the predicted brain age. The model training module is used to train the attention gate and TransformerU-Net model using a five-fold cross-validation strategy, with mean squared error as the loss function, and the AdamW optimizer to configure the learning rate strategy. Multidimensional data augmentation is introduced, and the model is trained based on standardized brain imaging data to obtain the trained brain age prediction model. The prediction and evaluation module is used to perform benchmark tests, simulation training, reliability analysis, attention map and pathological image alignment verification, and cross-diagnostic assessment based on the trained brain age prediction model, and output brain age prediction values ​​and mental illness risk assessment results.

Images