A low-bias multi-scale regional brain age prediction method based on adversarial training

Abstract

The application discloses a low-bias multi-scale regional brain age prediction method based on adversarial training, and relates to the technical fields of medical image analysis, artificial intelligence and neurodegenerative disease diagnosis. The method firstly standardizes and preprocesses T1 weighted structural MRI data and uses AAL3 atlas to segment brain regions, then constructs an adaptive 3D convolutional neural network and introduces a gradient reversal layer to realize age group adversarial training to reduce systematic bias, then adopts multiple regression models for brain age prediction and further reduces residual bias through a linear correction method, and finally performs statistical analysis to identify disease-specific regional accelerated aging patterns. The experimental results show that the method effectively reduces the age-related bias of 67.6% ROI, successfully identifies significant accelerated aging of Parkinson's disease patients in the left middle temporal gyrus, the lateral hypothalamic nucleus and the brainstem monoamine nucleus group, and provides important technical support for early diagnosis and progress monitoring of neurodegenerative diseases.

Description

Technical Field

[0001] This invention relates to the fields of medical image analysis, artificial intelligence, and diagnosis of neurodegenerative diseases, specifically to a region-specific brain age prediction method based on three-dimensional convolutional neural networks (3D-CNN) and adversarial training, which is particularly suitable for the assessment of brain aging in Parkinson's disease (PD). Background Technology

[0002] Parkinson's disease is the second leading neurodegenerative disease after Alzheimer's disease, and its early diagnosis and disease progression monitoring urgently require sensitive structural aging biomarkers. Brain-age prediction can map brain imaging features into a single quantitative indicator—the predicted age difference (PAD). Existing brain-age prediction research has the following key problems: (1) Insufficient region-specific analysis. Brain aging is a heterogeneous process, with different brain regions having unique aging trajectories. However, existing methods mostly focus on the overall assessment of the whole brain, making it difficult to accurately capture local aging patterns; (2) Significant age-related bias exists. The predicted age difference (PAD) is highly correlated with the actual age, which can easily introduce systematic bias. Traditional linear post-correction methods may mask the true population differences; (3) Model selection is sensitive. Different regression models (such as linear regression, random forest, etc.) have large differences in predictive performance, resulting in insufficient stability and reproducibility of the results. Traditional post-correction methods cannot solve the above problems simultaneously. Summary of the Invention

[0003] This invention aims to address the problems of systematic bias, model selection sensitivity, and insufficient region-specific analysis in existing brain age prediction methods, and proposes a region-specific brain age prediction method based on adversarial training and multi-model evaluation.

[0004] The technical solution is a low-bias multi-scale regional brain age prediction method based on adversarial training, comprising the following steps:

[0005] Step 1: Data Acquisition and Preprocessing;

[0006] Step 2: 3D-CNN feature extraction based on adversarial learning;

[0007] Step 2.1: Construct a 3D convolutional neural network adapted to multi-scale ROIs, with the network structure dynamically adjusted according to the ROI size;

[0008] Large ROI with dimensions of 64×64×64: 4-layer convolutional block with output channels [8,16,32,64];

[0009] Medium-sized ROI with dimensions of 32×32×32: 3-layer convolutional block with output channels [8,16,32];

[0010] Small ROI with dimensions of 16×16×16: 2-layer convolutional block with output channels [4,8];

[0011] Each convolutional block contains 3D convolution operations, batch normalization, and a ReLU activation function. Large and medium-sized ROIs are followed by a 2×2×2 max pooling operation after each layer, while small ROIs are only pooled in the last layer. The input kernel size is 2 for small ROIs and 3 for the others.

[0012] Step 2.2: Introduce age-group adversarial training to suppress age-related bias;

[0013] Auxiliary task: Divide the sample into 4 age groups;

[0014] Youth (44-51 years old), Middle-aged (52-65 years old), Pre-elderly (66-75 years old), Elderly (76-84 years old) – Age groups are distinguished using an adversarial classifier; Gradient Reversal Layer (GRL): A GRL is inserted between the feature extractor and the adversarial classifier. Features are preserved during forward propagation, and gradients are reversed during backward propagation. The formula for the GRL is:

[0015]

[0016] Where GRL(x) represents the feature vector output by the gradient reversal layer during forward propagation; x is the original feature vector input to the gradient reversal layer; I is the identity matrix, which multiplies the gradient by -1 during the backpropagation stage to achieve a complete reversal of the gradient direction, thereby creating an adversarial effect on the feature extractor.

[0017] Total loss function:

[0018]

[0019] in, To regress the loss, To perform adversarial classification loss, θ, φ, ψ are the parameters of the feature extractor, regressor, and adversarial classifier, respectively, and λ is the adversarial weight, where y pred The predicted brain age is represented by the regressor output; y represents the subject's actual age; g pred represents the predicted age group label output by the adversarial classifier; g represents the true age group label to which the sample belongs;

[0020] Counterweight scheduling: A gradual strategy is adopted, the formula is as follows:

[0021]

[0022] Where, λ (t)Let λ0 represent the maximum weight in round t, λ0 be the initial maximum weight, t be the current round, and T be the total number of training rounds.

[0023] Step 2.3: Compress the convolutional features into a fixed-length feature vector using global average pooling, which will then be used as input to the subsequent prediction model.

[0024] Step 3: Brain age prediction and bias correction using a multi-regression model;

[0025] Step 4: Model evaluation and region-specific analysis;

[0026] Step 4.1: Evaluation indicators: The mean absolute error is used to quantify the prediction accuracy, and the Pearson correlation coefficient between the predicted age difference (PAD) and the actual age is used to evaluate the deviation.

[0027] Step 4.2: Select the optimal model: On the validation set, select the optimal regression model for each dataset based on the mean absolute error and correlation coefficient;

[0028] Step 4.3: Region-specific analysis: PAD based on the optimal model * The regional brain age differences between the case group and the control group were compared using independent samples t-test, Welch test or Mann-Whitney U test, and multiple comparison errors were controlled by FDR correction.

[0029] Furthermore, the specific method of step 1 is as follows:

[0030] Step 1.1: Acquire T1-weighted structural magnetic resonance imaging data of subjects, including healthy controls and patients with Parkinson's disease;

[0031] Step 1.2: Standardize and preprocess the image data:

[0032] The BET tool was used for skull dissection and visual field cropping; linear registration to the 12-DOF MNI152 template was achieved using FLIRT, followed by nonlinear registration using FNIRT; brain tissue segmentation was performed using the FAST tool to obtain probability maps of cerebrospinal fluid, gray matter, and white matter; FNIRT is a nonlinear image registration tool.

[0033] Step 1.3: Based on the AAL3 brain region template, the standardized image is segmented into 166 regions of interest (ROIs), and ROI-specific preprocessing is performed: non-gray matter voxels are removed based on the gray matter probability map, effective voxels are retained, and 3D bounding boxes are determined; standardization is performed according to the spatial scale of the ROI: ROIs with a 3D size greater than or equal to 64 are resampled to 64×64×64, those less than 32 are resampled to 16×16×16, and those between 32 and 64 are resampled to 32×32×32.

[0034] Furthermore, in step 2.2 Mean squared error loss is used. Cross-entropy loss is used.

[0035] Furthermore, the specific method for step 3 is as follows:

[0036] Step 3.1: Construct a set of multiple regression models, including Gaussian process regression, support vector regression, K nearest neighbor regression, kernel regression, linear regression, Lasso regression, random forest, XGBoost, and multilayer perceptron, and optimize hyperparameters through grid search;

[0037] Step 3.2: Calculate the predicted age difference (PAD):

[0038]

[0039] in, To predict brain age, y represents actual age;

[0040] Step 3.3: Correct the deviation using the Beheshti linear correction method:

[0041] Linear regression was performed on the validation set with actual age y as the independent variable and PAD as the dependent variable:

[0042] PAD = αy + β;

[0043] Where α is the slope obtained by linear regression of PAD and actual age y on the validation set, used to describe the linear dependence between the two; β is the corresponding intercept, used to compensate for the overall bias.

[0044] The corrected PAD is:

[0045]

[0046] Furthermore, step 3.1 employs Lasso regression and performs a grid search using {0.0001,0.001,0.01,0.1}.

[0047] This invention achieves age-group adversarial training by introducing a gradient inversion layer into the feature layer, suppressing PAD-age correlation at its source. Combined with Beheshti linear correction, it further eliminates residual bias and employs multi-model benchmark evaluation to select the optimal regressor, significantly reducing model selection sensitivity. Compared to traditional overall brain age or post-hoc single linear correction schemes, this technology simultaneously achieves lower systematic bias, higher predictive stability, and cross-dataset consistency. It can also accurately locate region-specific accelerated aging brain regions in diseases such as Parkinson's disease, significantly enhancing the application value of brain age indicators in early clinical diagnosis and disease monitoring. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the overall technical process.

[0049] Figure 2 This section compares model strategies and the best model across datasets.

[0050] Figure 3 This image shows key brain regions that significantly accelerate aging in Parkinson's disease. Detailed Implementation

[0051] Step 1: Data Sources and Preprocessing

[0052] 1. Dataset Setup

[0053] ·UKB dataset

[0054] The UK Biobank dataset used for training contained 1000 healthy participants (aged 46–84 years), excluding those with a history of neurological diseases (such as Parkinson's disease, Alzheimer's disease) or significant motion artifacts or structural abnormalities. The test set consisted of 132 PD patients from the UKB dataset who did not participate in training and 140 healthy controls, with the test group strictly matched to the PD sample in age (±3 years) and sex. PD diagnosis was based on the UKB clinical criteria.

[0055] ·PPMI dataset

[0056] As an independent validation set, 98 PD patients and 88 healthy controls were used from Parkinson's Progression Markers Initiative (PPMI) database. To ensure data consistency, only structural images acquired using a 3T Siemens TrioTim or Prisma scanner were included.

[0057] 2. MRI Image Preprocessing Workflow

[0058] Brain tissue extraction

[0059] Use the BET tool in FSL 6.0.5 for stripping, with the parameters set to -f 0.5-g 0, to preserve the edge gray structure.

[0060] Spatial registration

[0061] First, linear registration was performed using the FLIRT tool, followed by a 12-DOF affine transformation, referencing the standard MNI152 template (1mm). 3 Resolution), mutual information was used as the registration criterion; then, the FNIRT tool was used for nonlinear registration, cubic B-spline interpolation, and the smoothing parameter was set to 3 mm.

[0062] Organizational segmentation

[0063] The image was divided into three tissue components—gray matter, white matter, and cerebrospinal fluid—using FSL's FAST tool. A probability threshold of 0.5 was set, and the pure gray matter portion was extracted for subsequent analysis.

[0064] ROI Construction and Resampling

[0065] Based on the AAL3 template, 166 regions of interest (ROIs) were defined. Voxels were selected using gray matter probability maps (threshold > 0.5). After extracting the effective regions, they were classified according to the median axis length: ≥ 64 → resampled to 64. 3 ; 32–64 → resampling to 32 3 <32 → Resampling to 16 3 Resampling is implemented using trilinear interpolation from the Python nibabel library.

[0066] Step 2: Model Building

[0067] 1. Multi-scale 3D-CNN network structure

[0068] Each convolutional layer is followed by Batch Normalization (momentum 0.9, epsilon = 1 × 10⁻⁵) and a ReLU activation function. Different network depths are used for different ROI sizes: large ROIs (64... 3 ): 4 convolutional layers (kernel = 3×3×3, stride = 1), followed by pooling after each layer (kernel = 2×2×2, stride = 2); medium-sized ROI (32 3 ): 3 convolutional layers, pooling layer settings as above; small ROI (16 3 ): 2-layer convolution (kernel = 2×2×2), pooling only in the last layer. Feature output dimension

[0069] Each ROI is output as a 64-dimensional feature vector after global average pooling.

[0070] 2. Age-based competitive training setup

[0071] Based on age, the participants were divided into four groups: young adults (44–51 years), middle-aged adults (52–65 years), pre-old adults (66–75 years), and elderly adults (76–84 years). The adversarial classifier was a three-layer fully connected network (128→64→4) used to implement GRL (Gradient Reversal Layer) adversarial training.

[0072] • Loss function setting

[0073] The total loss function is defined as:

[0074]

[0075] in, For the MSE (mean squared error) of the regression task, Let λ be the cross-entropy loss for age group classification, and λ be the weights for the adversarial loss. The initial weights are set to λ0 = 0.1. With a total training epochs T = 100, the dynamic update formula is as follows:

[0076]

[0077] Optimizer and Training Strategy

[0078] The model was trained using the Adam optimizer (learning rate = 0.001), with a batch size of 8 and 100 training iterations.

[0079] Step 3: Deviation Correction and Statistical Analysis

[0080] 1. Linearity Correction (Beheshti Method)

[0081] Fitting a linear regression on a healthy control validation set:

[0082] PAD=αy+β

[0083] The corrected predicted age difference is:

[0084]

[0085] 2. Statistical Analysis Process

[0086] • Distribution and variance test

[0087] Distribution test: Shapiro-Wilk (p>0.05 indicates normality); Homogeneity of variance test: Levene (p>0.05 indicates homogeneity);

[0088] • Strategies for testing differences between groups

[0089] Normality & homogeneity → Independent samples t-test; Normality & heteroscedasticity → Welch t-test; Non-normal distribution → Mann-Whitney U test.

[0090] Multiple comparison correction

[0091] FDR correction was performed using the Benjamini-Hochberg method. The brain was divided into three groups: cortex, subcortex, and brainstem. The q-value was calculated for each group, with a significance criterion of q < 0.05.

[0092] Step 4: Performance Evaluation and Key Brain Region Analysis

[0093] To comprehensively evaluate the performance of the proposed region-specific brain age prediction model, quantitative evaluations were conducted on both internal test sets and external validation sets. The metrics used included: mean absolute error (MAE): measuring the degree of deviation between predicted and actual age; correlation coefficient (|r|) between predicted and actual age: assessing the presence of age bias in the model; and significance test for intergroup differences (p-value): assisting in validating the model's ability to separate data across different populations (e.g., PD vs. HC).

Claims

1. A low-bias multi-scale regional brain age prediction method based on adversarial training, comprising the following steps: Step 1: Data Acquisition and Preprocessing; Step 2: Feature extraction based on adversarial learning using 3D-CNN; Step 2.1: Construct a 3D convolutional neural network adapted to multi-scale ROIs, with the network structure dynamically adjusted according to the ROI size; Large ROI with dimensions of 64×64×64: 4 convolutional blocks with output channels of [8,16,32,64]; Medium-sized ROI with dimensions of 32×32×32: 3 convolutional blocks with output channels [8,16,32]; Small ROI with dimensions of 16×16×16: 2-layer convolutional block with output channels [4,8]; Each convolutional block contains 3D convolution operations, batch normalization, and ReLU activation functions. Large and medium-sized ROIs are appended with 2×2×2 max pooling after each layer, while small ROIs are appended with pooling only in the last layer. The input kernel size is 2 for small ROIs and 3 for the rest. Step 2.2: Introduce age-group adversarial training to suppress age-related bias; Auxiliary task: Divide the sample into 4 age groups; Youth (44-51 years old), Middle-aged (52-65 years old), Pre-elderly (66-75 years old), Elderly (76-84 years old) – Age groups are distinguished using an adversarial classifier; Gradient Reversal Layer (GRL): A Gradient Reversal Layer (GRL) is inserted between the feature extractor and the adversarial classifier. Features remain unchanged during forward propagation, and gradients are reversed during backward propagation. The formula for the GRL is: ； in, This represents the feature vector output by the gradient reversal layer during forward propagation; is the original feature vector input to the gradient reversal layer; I is the identity matrix, which multiplies the gradient by -1 during the backpropagation stage to completely reverse the gradient direction, thereby creating an adversarial effect on the feature extractor. Total loss function: ； in, To regress the loss, To combat classification loss, These are the parameters for the feature extractor, regressor, and adversarial classifier, respectively. To counteract the weighting, among which, This represents the predicted brain age output by the regressor; Indicates the subject's actual age; This represents the predicted age group label output by the adversarial classifier; This indicates the actual age group label to which the sample belongs; Counterweight scheduling: A gradual strategy is adopted, the formula is as follows: ； in, This represents the maximum weight in round t. The initial maximum weight, For the current round, Total number of training rounds; Step 2.3: Compress the convolutional features into a fixed-length feature vector using global average pooling, which will then be used as input to the subsequent prediction model; Step 3: Brain age prediction and bias correction using a multi-regression model; Step 3.1: Construct a set of multiple regression models, including Gaussian process regression, support vector regression, K nearest neighbor regression, kernel regression, linear regression, Lasso regression, random forest, XGBoost or multilayer perceptron, and optimize hyperparameters through grid search; Step 3.2: Calculate the predicted age difference (PAD): ； in, To predict brain age, Actual age; Step 3.3: Correct the deviation using the Beheshti linear correction method: Linear regression was performed on the validation set with actual age y as the independent variable and PAD as the dependent variable: ； Where α is the slope obtained by linear regression of PAD and actual age y on the validation set, used to describe the linear dependence between the two; β is the corresponding intercept, used to compensate for the overall bias. The corrected PAD is: ； Step 4: Model evaluation and region-specific analysis; Step 4.1: Evaluation indicators: The mean absolute error is used to quantify the prediction accuracy, and the Pearson correlation coefficient between the predicted age difference (PAD) and the actual age is used to evaluate the deviation; Step 4.2: Select the optimal model: On the validation set, select the optimal regression model for each dataset based on the mean absolute error and correlation coefficient; Step 4.3: Region-specific analysis: based on the optimal model The regional brain age differences between the case group and the control group were compared using independent samples t-test, Welch test or Mann-Whitney U test, and multiple comparison errors were controlled by FDR correction.

2. The method for predicting brain age in multiple scales with low bias based on adversarial training as described in claim 1, characterized in that, The specific method for step 1 is as follows: Step 1.1: Acquire T1-weighted structural magnetic resonance imaging data of subjects, including healthy controls and patients with Parkinson's disease; Step 1.2: Standardize and preprocess the image data: The BET tool was used for skull dissection and visual field cropping; linear registration to the 12-DOF MNI152 template was achieved using FLIRT, followed by nonlinear registration using FNIRT; brain tissue segmentation was performed using the FAST tool to obtain probability maps of cerebrospinal fluid, gray matter, and white matter; FNIRT is a nonlinear image registration tool. Step 1.3: Based on the AAL3 brain region template, the standardized image is segmented into 166 regions of interest (ROIs), and ROI-specific preprocessing is performed: non-gray matter voxels are removed based on the gray matter probability map, effective voxels are retained, and 3D bounding boxes are determined; standardization is performed according to the spatial scale of the ROI: ROIs with a 3D size greater than or equal to 64 are resampled to 64×64×64, those less than 32 are resampled to 16×16×16, and those between 32 and 64 are resampled to 32×32×32.

3. The method for predicting brain age in low-bias multi-scale regions based on adversarial training as described in claim 1, characterized in that, In step 2.2 Mean squared error loss is used. Cross-entropy loss is used.

4. The method for predicting brain age in multiple scale regions with low bias based on adversarial training as described in claim 1, characterized in that, Step 3.1 employs Lasso regression and performs a grid search using {0.0001, 0.001, 0.01, 0.1}.

Images