Brain age prediction model construction method combined with auxiliary classification and program product
By combining a gated sparse hybrid expert model for auxiliary classification with a multi-task learning mechanism, the accuracy and relevance issues of brain age prediction models in small to medium-sized disease populations were resolved, achieving higher prediction accuracy and robustness.
Patent Information
- Application Number
- CN202511440459.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-10
- Publication Date
- 2026-01-02
AI Technical Summary
Existing brain age prediction models show decreased accuracy and relevance in small to medium-sized disease populations, exhibiting age bias and heteroscedasticity. Disease information is not provided as a training aid, and auxiliary classification may interfere with the main regression performance.
We employ a gated sparse hybrid expert model and a multi-task joint learning mechanism, combined with auxiliary classification of Alzheimer's disease/mild cognitive impairment/healthy controls, and perform end-to-end optimization through standardized age supervision, dual ranking regularization, and category inverse frequency weighting.
It significantly reduced the mean absolute error of brain age prediction, improved age-rank consistency and clinical interpretability, and enhanced the robustness of the model in real clinical distributions.
Smart Images
Figure CN121260418A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of medical image processing, and in particular to a brain age prediction model construction method combined with auxiliary classification and a program product. BACKGROUND
[0002] Brain age is an individual "biological age" predicted based on neuroimaging. The difference between brain age and actual age (brain age gap, BAG) has potential value in the screening and prognosis evaluation of neurodegenerative diseases. In recent years, three-dimensional convolutional neural networks and their variants have been widely used in brain age regression of T1-weighted MRI, and have achieved lower mean absolute error (MAE) and higher correlation coefficients on multiple cohorts. However, a systematic review of published literature and a replication experiment on ADNI (Alzheimer's Disease Neuroimaging Initiative) shows that the existing mainstream methods still have the following shortcomings:
[0003] (1) In ADNI and other relatively small-scale / disease-populated cohorts, the precision and correlation decrease significantly:
[0004] For example, the performance on ADNI is generally lower than that on the UK Biobank with a large sample: in the replication experiment, MAE ≈ 3.0-3.5 years; while in the UK Biobank, MAE ≈ 2.1-2.7 years. For example, Global-Local Transforme has MAE of 3.32-3.44 years on ADNI; SFCN (Peng 2021) has MAE of 3.51-3.52 years on ADNI; 3D-DenseNet (Lee 2022) has MAE of 3.02-3.48 years on ADNI; DGA-DMIL (Yan 2024) has the best MAE of 2.79 years on ADNI. This scheme selects a small number of experts from heterogeneous subjects according to the sample, and the gradient of the experts not selected on the sample is zero, thereby providing "conditional" model capacity without increasing over-coupling, reducing overfitting and mutual interference in small samples. At the same time, a disease classification head is introduced in parallel on the shared representation to train with inverse frequency weighted cross-entropy, so that the disease-related structural differences participate in representation shaping during the training period, improving the transferability and clinical relevance in disease-rich domains (ADNI).
[0005] (2) There is age bias and heteroscedasticity: BAG is significantly correlated with age, and uncorrected will amplify error variance:
[0006] Previous studies have shown that the predicted BAG is negatively correlated with the actual age (regression to the mean), resulting in overestimation of the young and underestimation of the old; for example, in 3D-DenseNet, the Spearman correlation coefficient (BAG vs age) is equal to -0.464 without bias correction, and the MAE decreases from 4.206 to 3.487 after linear correction, proving that uncorrected will systematically amplify errors. This age-related bias is repeatedly observed in different models / datasets. Further studies have shown that even after common sample-level linear correction, residual correlation and error variance may still occur at the age level (different age groups have different residual distribution widths), which requires more granular age-level correction to suppress the coupling of error variance and age. This scheme introduces (i) sample-to-age difference constraints and (ii) differentiable ranking approximation Spearman rank correlation in addition to the main regression loss, which forces the global order of predicted age to be consistent with the true age order, and from a mechanism to suppress the systematic drift of BAG with age, reducing the correlation and variance amplification effect of age-error.
[0007] (3) Disease information is mainly used for post-analysis rather than training period auxiliary supervision, limiting the transferability and clinical relevance of the representation:
[0008] Lee et al. (Nature Aging 2022) verified the correlation between BAG and multiple clinical scales / molecular images, and observed significantly higher BAG in the Alzheimer's disease / healthy control group of ADNI, suggesting that disease status can affect brain age phenotype. However, most brain age work does not explicitly introduce Alzheimer's disease / mild cognitive impairment / healthy control labels as auxiliary tasks during training, and still mainly uses single-task regression, which does not systematically constrain shared representations to align with disease-related physiological characteristics. This scheme uses a parallel age regression head and disease classification head on shared representations during multi-task joint learning, and the gradients of the two heads update the backbone and mixed expert model together, so that structural differences related to Alzheimer's pathology are encoded during feature learning, improving the transferability of representations to the disease domain. At the same time, the weighted cross-entropy of Alzheimer's disease / mild cognitive impairment / healthy control is inversely proportional to the class frequency to ensure that the supervision signal of the minority class is not drowned, further improving the robustness in real clinical distribution.
[0009] (4) A technical bias in the field: concern that auxiliary classification will interfere with main regression (negative transfer):
[0010] Multi-task learning literature has long suffered from negative transfer and task seesaw risk: when task relevance is insufficient or weight is improper, the introduction of auxiliary tasks may lower the performance of the main task. Therefore, a large number of representative works in brain age research (such as TSAN (Cheng 2021), SFCN (Peng 2021), Global-Local Transformer r (He 2022) and the like) still follow the single-task regression paradigm, and Alzheimer's disease / mild cognitive impairment / healthy controls are not included in the training period supervision. The present scheme solves the problem of negative transfer through structural decoupling and sparse routing. The mixed expert Top-k sparse routing activates only a small number of experts for each sample, naturally reducing the global sharing of classification and regression in the parameter space, thereby reducing the invalid interference of cross-task gradients and reducing the negative impact of auxiliary tasks on the main task. Disease classification supervision is only introduced in the training stage to shape the shared representation, and the classification head is closed in the inference stage so as to not affect the age estimation output, further avoiding negative coupling in deployment. Finally, although the total loss is composed of the main regression (MAE / MSE), resulting in possible interference of auxiliary tasks, but the ranking / differential ranking and weighted cross-entropy are composed of coefficients, and the MAE of the validation set is always used as the only model selection standard, ensuring that the optimization process always takes regression performance as the main target, which can further suppress the adverse shift that may be brought by auxiliary tasks. SUMMARY
[0011] The present application uses a gated sparse mixed expert model and a multi-task joint learning mechanism on the basis of unified preprocessing of three-dimensional MRI to explicitly constrain the auxiliary classification of Alzheimer's disease / mild cognitive impairment / healthy controls on age regression; through standardized age supervision+double ranking regularization+category inverse frequency weighting and gender embedding, a platform descent learning strategy and an early stopping strategy are used for end-to-end optimization to minimize the MAE of the validation set.
[0012] The present application provides the following technical solutions:
[0013] The present application provides a brain age prediction model construction method combined with auxiliary classification, characterized by the following steps:
[0014] S1: Obtain sample data, and divide the sample data into a training set and a validation set, wherein the sample data includes brain MRI data, actual age, gender and neurodegenerative disease label of each subject;
[0015] S2: using a feature extraction network, extracting shared representation of each subject from brain MRI data and gender in the training set , wherein the shared representation of the i-th subject is represented as ;
[0016] S3: input the shared representation of each subject into an age regression head and a disease classification head respectively; the age regression head is a regression model outputting expected brain age, and the disease classification head is a classification model outputting classification of neurodegenerative disease;
[0017] The actual age corresponding to each subject is used as a supervision signal of the age regression head, and the neurodegenerative disease label corresponding to each subject is used as a supervision signal of the disease classification head, and the parameters in the age regression head, the disease classification head and the feature extraction network are jointly trained using a total loss; the total loss includes a regression loss of the age regression head and a classification loss of the disease classification head.
[0018] During the joint training process, the regression loss of the age regression head in the validation set is used as a monitoring index, and the feature extraction network and the age regression head at the moment when the regression loss of the age regression head in the validation set is the smallest during the joint training process are saved as the final brain age prediction model.
[0019] Preferably, the label elements in the neurodegenerative disease label include at least one of Alzheimer's disease, mild cognitive impairment and healthy control.
[0020] Preferably, the training set The form of the data in the training set is:
[0021]
[0022] The preprocessed brain MRI data of the i-th subject in the training set
[0023] The actual age value of the i-th subject in the training set
[0024] The gender one-hot encoding of the i-th subject in the training set The neurodegenerative disease label of the i-th subject in the training set
[0025] The number of subjects in the training set
[0026] The feature extraction network includes the following contents: 1) a gender embedding layer for mapping the gender of each subject to an embedding vector
[0027] , and the specific formula is:
[0028]
[0029]
[0030] where, is the training set is the gender embedding vector of the i-th subject in
[0031] , is the parameter of the gender embedding layer;
[0032] is the dimension of ;
[0033] When gender is missing or disabled, directly let = 0;
[0034] 2) Cross-scale dense connection backbone and cross-scale fusion layer, used to extract three-dimensional feature sequence from and get image representation ;
[0035] where the output of the j-th layer of is defined as:
[0036]
[0037] where, represents resampling the j-th layer feature to the j+1-th layer scale;
[0038] represents channel concatenation;
[0039] is a convolution-attention hybrid module;
[0040] is the number of layers of the three-dimensional feature sequence;
[0041] The calculation formula of image representation is:
[0042]
[0043] where is the global average pooling function;
[0044] 3) Gating sparse hybrid expert layer, used to output ;
[0045] The calculation method of is:
[0046] 3.1) Calculate the input vector of the i-th subject :
[0047]
[0048] wherein, represents the concatenation in the channel dimension;
[0049] 3.2) Calculate the gating score vector:
[0050]
[0051] wherein, is the number of expert subnetworks;
[0052] is a gating scoring network, outputting a vector of dimension , each element of which represents the score of the corresponding expert subnetwork;
[0053] is a gating score vector, containing elements, wherein the mth element is ; represents the score of the mth expert subnetwork for the ith subject;
[0054] 3.3) Select the index set of k expert subnetworks by a gating algorithm:
[0055]
[0056] wherein, represents the gating algorithm for selecting the index set of k expert subnetworks;
[0057] represents the index set of k expert subnetworks selected by the gating algorithm for the ith subject;
[0058] 3.4) Calculate the normalized gating probability of the expert subnetworks with index in the range of :
[0059]
[0060] wherein, represents the normalized gating probability of the expert subnetworks with index in the range of ;
[0061] 3.4) Calculate :
[0062]
[0063] wherein, is the mth expert subnetwork; Preferably,
[0064] A convolution-attention hybrid module composed of 3D convolution, batch normalization, exponential linear activation, channel attention and pooling.
[0065] Preferably, Specifically Gating algorithm.
[0066] Preferably, , .
[0067] Preferably, the age regression head is represented as , the output of the age regression head is , and the formula is:
[0068]
[0069] The disease classification head is represented as , the output of the disease classification head is , and the formula is:
[0070]
[0071] wherein, is a normalized vector of C dimensions, satisfying ; wherein, represents the cth element in , which means the probability of the cth disease;
[0072] The actual age in the form of a supervised signal is the standardized actual age, and the calculation formula is:
[0073]
[0074] wherein, is the standardized actual age;
[0075] is the mean of the actual age in the training set , and the formula is:
[0076]
[0077] is the standard deviation of the actual age in the training set , and the formula is:
[0078]
[0079] The total loss is calculated according to the formula:
[0080]
[0081] wherein, The regression loss of MAE angle is as follows:
[0082]
[0083] The regression loss of MSE angle is as follows:
[0084]
[0085] The regression loss of difference ranking angle is as follows:
[0086]
[0087] wherein, ;
[0088] The regression loss of differentiable rank consistency angle is as follows:
[0089]
[0090] wherein is a differentiable ranking operator with temperature ;
[0091] The classification loss of cross-entropy angle is as follows:
[0092]
[0093] wherein is a category weight obtained after normalization, and the formula is as follows:
[0094]
[0095] wherein, is the number of samples equal to 1 in the training set ;
[0096] represents the cth element in , and the meaning is whether the cth label element in of the ith subject is true, if true, the value is 1, if false, the value is 0;
[0097] , , , , are , , , , The weighting coefficients.
[0098] Preferably, the regression loss used as the monitoring indicator is .
[0099] The present invention also provides a brain age prediction method, which inputs the subject's brain MRI data and gender into a brain age prediction model to obtain the subject's predicted brain age; wherein the brain age prediction model is obtained according to the above-described construction method.
[0100] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.
[0101] Beneficial effects:
[0102] Compared to single-task networks that only perform regression, this method uses Alzheimer's disease / mild cognitive impairment / healthy controls auxiliary classification gradients and Top-k expert selection from a sparse hybrid expert model to jointly shape shared representations. It also incorporates dual ranking regularization and gender embedding. Under a standardized-anti-standardized evaluation system, this significantly reduces age-rank consistency (MAE) and improves age-rank consistency and clinical interpretability. The preprocessing employs domain-standard methods (FSL medical image processing → Montreal neural space → brain extraction → intensity normalization), making it easily transferable and reproducible. Attached Figure Description
[0103] Figure 1 This is a schematic diagram of the overall structure of the brain age prediction method of the present invention. Detailed Implementation
[0104] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.
[0105] like Figure 1 As shown, a method for constructing a brain age prediction model that combines auxiliary classification includes the following steps:
[0106] S1: Obtain sample data and divide the sample data into training sets. The sample data includes each subject's brain MRI data, actual age, sex, and neurodegenerative disease label; and a validation set.
[0107] In this embodiment, the sample data is in the following form:
[0108]
[0109] in, For the sample data, the first Preprocessed brain MRI data of one subject;
[0110] Specifically, for the original T1 MRI (other modal MRIs, such as T2, FLAIR, DTI, etc., do not affect the processing method), the following steps are performed sequentially: nonlinear registration to Montreal neural space, brain tissue extraction, voxel intensity z-score normalization within the brain region, and resampling to a uniform voxel size, to obtain... The above preprocessing method is a common preprocessing method in this field, and will not be elaborated further. Using this form of MRI data facilitates cross-center transfer and reproduction. In this embodiment, for The voxel grid.
[0111] For the sample data, the first The actual age value of each subject;
[0112] For the sample data, the first The gender of each subject is uniquely encoded; in this embodiment, the gender of each subject is uniquely encoded. ;
[0113] For the sample data, the first The neurodegenerative disease label of each subject; in this embodiment, , =3, representing the number of label elements for neurodegenerative diseases, corresponding to Alzheimer's disease, mild cognitive impairment, and healthy control. Each person can only have one state: Alzheimer's disease is [1,0,0], mild cognitive impairment is [0,1,0], and healthy is [0,0,1]. Currently, Alzheimer's disease and mild cognitive impairment are mutually exclusive. If other non-mutually exclusive neurodegenerative diseases are added, additional dimensions need to be added to express multiple diseases simultaneously. For example, after adding Parkinson's disease, three dimensions need to be added: Parkinson's-Alzheimer's, Parkinson's-mild cognitive impairment, and Parkinson's disease.
[0114] The number of subjects in the sample data;
[0115] The training set The data is in the following format:
[0116]
[0117] For training set pre-processed brain MRI data of the i-th subject in the training set;
[0118] for the training set actual age value of the i-th subject in the training set;
[0119] for the training set gender one-hot encoding of the i-th subject in the training set;
[0120] for the training set neurodegenerative disease label of the i-th subject in the training set;
[0121] for the training set number of subjects in the training set;
[0122] S2: using the feature extraction network, extracting shared representation of each subject from brain MRI data, gender in the training set, wherein the shared representation of the i-th subject is represented as .
[0123] In this embodiment, the feature extraction network includes the following:
[0124] 1) gender embedding layer, used to map to generate embedding vector , the specific formula is:
[0125]
[0126] wherein, gender embedding vector of the i-th subject in the training set;
[0127] , parameters of the gender embedding layer;
[0128] dimension of ; When the gender is missing or disabled, directly let
[0129] = 0; Here, the design of "one-hot encoding → linear mapping" makes the network equivalent to learning two interpretable gender prototype vectors. The linear layer does not construct a complex code, but corresponds to a small vector prototype of each gender class, which will automatically move to a more beneficial position for age regression and classification assistance under the constraint of the total loss during the training process.
[0130] , The network parameters are randomly initialized along with other components (linear layers are initialized using Xavier, convolutional layers using Kaiming He), and updated end-to-end using Adam; during backpropagation, gradients from regression, ranking, and classification branches work together. , This makes the "baseline shift" of both methods more closely match the data distribution for different population groups. If gender is missing or disabled, it will directly... (instead of substitution) This is to avoid artificially introducing bias. After multiple experimental verifications, it was finally found that... It achieves a good balance between expressive power and the risk of overfitting. If only one-hot encoding is used for direct concatenation, the numerical dimension is too small and discrete, making it difficult to have a continuously adjustable impact on the gating score and regression baseline. Furthermore, gender itself is a low-entropy label, and introducing pre-embedding from a large external model yields limited benefits and poor reproducibility; while this approach uses a very small number of parameters (only...). This allows the system to adapt to the optimal "conditional bias" within the training domain, while also being more friendly to missing values and privacy control.
[0131] 2) A cross-scale densely connected backbone and a cross-scale fusion layer are used to... Extracting three-dimensional feature sequences And obtain image representation ;
[0132] in The The output of the layer is defined as:
[0133]
[0134] in, Indicates the first Layer feature resampling to the first Layer scale (maximum pooling in this embodiment);
[0135] Indicates channel splicing;
[0136] It is a hybrid convolution-attention module;
[0137] The number of layers in the three-dimensional feature sequence;
[0138] In this embodiment, It is a hybrid convolution-attention module consisting of 3D convolution, batch normalization, exponential linear activation, channel attention, and pooling. In some embodiments, For 3D-ResNet / ResNeXt, 3D-ConvNeXt, UNet / UNet++, HRNet, Swin-Transformer-3D / ViT-3D or other convolution-attention hybrid modules; different forms of convolution-attention hybrid modules only affect the implementation of feature extraction, and do not change the original formula and symbol of the subsequent hybrid expert model, fusion and loss.
[0139] The calculation formula of the image representation is:
[0140]
[0141] wherein is a global average pooling function;
[0142] 4) Gating sparse hybrid expert layer, used to output ;
[0143] The calculation method of is:
[0144] 3.1) Calculate the input vector of the ith subject :
[0145]
[0146] wherein represents concatenation in the channel dimension; that is, by concatenating the image representation and the gender embedding vector in the channel dimension to obtain ;
[0147] 3.2) Calculate the gate value scoring vector:
[0148]
[0149] wherein is the number of expert subnets;
[0150] is a gating scoring network, which outputs a dimensional vector, and each element represents the score of the corresponding expert subnet;
[0151] is a gate value scoring vector, containing elements, wherein the mth element is ; represents the score of the mth expert subnet for the ith subject;
[0152] 3.3) Select the index set of k expert subnets through the gating algorithm:
[0153]
[0154] wherein, denotes the gating algorithm that selects the index set of k expert subnets; in this embodiment, Specifically the gating algorithm that selects the index of the k expert subnets with the highest value; in some embodiments, can be Noisy Top-k, Top-p, Gumbel-Softmax, Hash or switch gating algorithm;
[0155] denotes the index set of k expert subnets selected by the gating algorithm for the i-th subject;
[0156] 3.4) Calculate the normalized threshold probability of the expert subnet whose index is in
[0157]
[0158] wherein denotes the normalized threshold probability of the expert subnet whose index is in
[0159] 3.4) Calculate
[0160]
[0161] wherein is the i-th expert subnet; The gradient of the unselected expert subnet on the sample is shielded to achieve sparse routing. The preferred configuration is
[0162] . Expert gate implementation can use Noisy Top-k / Top-p / Gumbel-Softmax / Hash or switch gating. The shunt routing granularity can be expanded from channel to patch / token / ROI grouping and hierarchical expert model, while adding load balancing constraints without changing the implementation. S3: input the shared representation of each subject into the age regression head and the disease classification head respectively; the age regression head is a regression model outputting the expected brain age, and the disease classification head is a classification model outputting the classification of neurodegenerative diseases;
[0163] Use the actual age corresponding to each subject as the supervision signal of the age regression head, use the neurodegenerative disease label corresponding to each subject as the supervision signal of the disease classification head, and use the total loss to jointly train the parameters in the age regression head, the disease classification head and the feature extraction network; the total loss includes the regression loss of the age regression head and the classification loss of the disease classification head;
[0164]
[0165] In the joint training process, the regression loss of the age regression head in the validation set is taken as a monitoring index. The feature extraction network and the age regression head in the joint training process when the regression loss of the age regression head in the validation set is the smallest are saved as the final brain age prediction model.
[0166] In the present embodiment, the age regression head is represented as The output of the age regression head is , and the formula is:
[0167]
[0168] The disease classification head is represented as The output of the disease classification head is , and the formula is:
[0169]
[0170] wherein, is a normalized vector of C dimensions, satisfying ; wherein, represents the cth element in , which means the probability of the cth disease class;
[0171] In the present embodiment, the actual age in the form of a supervision signal is a standardized actual age, and the calculation formula is:
[0172]
[0173] wherein, is the standardized actual age;
[0174] is the mean of the actual age in the training set , and the formula is:
[0175]
[0176] is the standard deviation of the actual age in the training set , and the formula is:
[0177]
[0178] Therefore, the final output of the brain age prediction model is also a standardized expected brain age, and only needs to be converted to obtain the actual expected brain age:
[0179]
[0180] wherein, is the actual expected brain age;
[0181] In this embodiment, the total loss is calculated as follows:
[0182]
[0183] wherein, is the regression loss of MAE angle, and the formula is:
[0184]
[0185] is the regression loss of MSE angle, and the formula is:
[0186]
[0187] is the regression loss of differential ranking angle, and the formula is:
[0188]
[0189] wherein,
[0190] is the regression loss of differentiable rank consistency angle, and the formula is:
[0191]
[0192] wherein is the differentiable ranking operator with temperature ;
[0193] is the classification loss of cross-entropy angle, and the formula is:
[0194]
[0195] wherein is the category weight obtained after normalization, and the formula is:
[0196]
[0197] wherein, is the number of samples in the training set equal to 1;
[0198] represents the cth element in , which means whether the cth label element in of the ith subject is true, and if true, the value is 1, and if false, the value is 0;
[0199] ,
[0200]
[0201] The training process uses platform learning rate scheduling (decay learning rate by a set factor when the indicator stagnates) and early stopping strategy (stop training if the validation set MAE does not improve for several rounds). The minimum value of the validation set MAE is used as the criterion for the best model and the corresponding parameters are saved. During training, disease classification supervision and age regression update the shared backbone and gate / specialist modules together; during inference, only the regression head output age can be enabled. Replace the optimizer with AdamW / RMSProp / LAMB / Lion or SAM+ base optimizer; the learning rate strategy can be cosine annealing / One-Cycle / linear warm-up; all are training strategy replacements that do not affect algorithm design.
[0202] During the inference process, the disease classification head can output Alzheimer's disease / mild cognitive impairment / healthy control probability according to the specific task needs for risk prompt, and closing does not affect brain age estimation.
[0203] Synergistic technical effects and quantitative verification
[0204] Data and evaluation criteria: except for special instructions, all are evaluated on ADNI; brain age takes MAE (years) as the main indicator; AD auxiliary classification takes Acc (%) as the main indicator. The public method uses its representative paper implementation and unified evaluation criteria for comparison; "our method (Ours)" is implemented by the applicant. See Tables 1-3 for details.
[0205] Table 1 is a comparison experiment of brain age prediction (ADNI; indicator: MAE, unit: years) under the unified evaluation criteria to obtain the current optimal performance, which quantifies the significant lead of the current method in ADNI disease-enriched cohort.
[0206] Table 2 is an Alzheimer's disease auxiliary classification comparison experiment (ADNI; index: Acc, %) relative to EfficientNet-B0 +1.65 percentage points, and relative to group self-calibrated +1.82 percentage points. This shows that when the shared representation is subject to multi-expert model multi-routing and ordering consistency constraints, disease discrimination also benefits. Table 3 is an ablation experiment. Adding AD auxiliary classification under the joint action of weight balancing and ordering regularization, relative to only MoE, further reduces MAE by about 6.0% ((1.33-1.25) / 1.33). This result refutes the technical bias that "auxiliary classification interferes with regression" and shows that under the loss design and sparse routing mechanism of the method, the auxiliary task forms an effective regularization and prior for the main task. At the same time, it shows that the lean yield of the present scheme cannot be directly derived from "linear superposition of common effects of each single module": at the level of architecture, Top-k sparse routing allows only a small number of experts to be activated for samples, significantly reducing cross-task gradient interference; at the prior level, the few-class weight shaping of Alzheimer's disease / mild cognitive impairment / healthy controls shapes the discriminant direction of the shared representation; at the rank level, differential ordering + differentiable rank consistency suppresses age-error coupling and rank disorder; the three work together to make the main task error decrease and the auxiliary task accuracy increase at the same time, reflecting the synergistic technical effect rather than simple superposition.
[0207] Table 1
[0208] Method (abbreviation, author, year) ADNI (MAE) TSAN (Cheng et al., 2021) 1.48 Global-Local Transformer (He et al., 2022) 3.32 SFCN / Lightweight DNN (Peng et al., 2021) 3.52 Deep learning-based brain age (Lee et al., 2022) 3.02 Dual Graph Attention-based Disentanglement MIL (Yan et al., 2024) 2.79 VNN / coVariance NN (Sihag et al., 2023) 3.56 Clinical MRI brain age framework (Kim et al., 2025) 2.05 Our method (Ours) 1.25
[0209] Table 2
[0210] Method (abbreviation, author, year) ADNI (Acc, %) AlexNet 89.95 EfficientNet-B0 92.98 LeNet 87.72 MobileNet 83.40 MobileNetV2 80.19 ResNet-50 88.56 ResNet-152 87.77 VGG-19 89.77 Xception 77.40 Group self-calibrated (Yu et al., 2024) 92.81 Our method (Ours) 94.63
[0211] Table 3
[0212] Component configuration ADNI (MAE) MoE only (without AD assistance) 1.33 MoE + AD assistance classification 1.25
[0213] Table 3
[0214] The above only describes the preferred embodiments of the present application, and it should be noted that for ordinary skilled persons in the art, without departing from the principles of the present application, several improvements and refinements can be made, and these improvements and refinements should be considered as the protection scope of the present application.
Claims
1. A method for constructing a brain age prediction model combined with auxiliary classification, characterized in that, Includes the following steps: S1: Obtain sample data and divide the sample data into training sets. The sample data includes each subject's brain MRI data, actual age, sex, and neurodegenerative disease label; and a validation set. S2: Using a feature extraction network, extract features from the training set... The shared representation of each subject was extracted from the brain MRI data and gender, where the shared representation of the i-th subject is represented as... ; S3: Input the shared representation of each subject into the age regression head and the disease classification head respectively; the age regression head is a regression model that outputs the expected brain age, and the disease classification head is a classification model that outputs the classification of neurodegenerative diseases; The actual age of each subject is used as the supervision signal for the age regression head, and the neurodegenerative disease label of each subject is used as the supervision signal for the disease classification head. The parameters in the age regression head, disease classification head, and feature extraction network are jointly trained using the total loss. The total loss includes the regression loss of the age regression head and the classification loss of the disease classification head. During joint training, the regression loss of the age regression head in the validation set is used as a monitoring indicator. The feature extraction network and age regression head that minimize the regression loss of the age regression head in the validation set during joint training are saved as the final brain age prediction model.
2. The method for constructing a brain age prediction model combining auxiliary classification according to claim 1, characterized in that, The label elements in the neurodegenerative disease label include Alzheimer's disease, mild cognitive impairment, and healthy controls.
3. The method for constructing a brain age prediction model combining auxiliary classification according to claim 1, characterized in that, The training set The data is in the following format: For training set Preprocessed brain MRI data of the i-th subject; For training set The actual age value of the i-th subject; For training set The unique-hot encoding of the gender of the i-th subject; For training set The neurodegenerative disease label of the i-th subject; For training set The number of subjects in the study; Feature extraction networks include the following: 1) Gender embedding layer, used to embed gender into a single layer. Mapping to generate embedding vectors The specific formula is as follows: in, For training set The gender embedding vector of the i-th subject; , For parameters of the gender embedding layer; for dimensionality; When gender is missing or disabled, directly order =0; 2) A cross-scale densely connected backbone and a cross-scale fusion layer are used to... Extracting three-dimensional feature sequences And obtain image representation ; in The The output of the layer is defined as: in, Indicates the first Layer feature resampling to the first Layer scale; Indicates channel splicing; It is a hybrid convolution-attention module; The number of layers in the three-dimensional feature sequence; The formula for calculating image representation is: in This is the global average pooling function; 3) Gated sparse hybrid expert layer, used for output ; The calculation method is as follows: 3.1) Calculate the input vector of the i-th subject. : in, This indicates splicing in the channel dimension; 3.2) Calculate the threshold scoring vector: in, The number of expert subnets; For the gating scoring network, output a A dimensional vector, where each element represents the score of the corresponding expert subnet; The threshold scoring vector contains There are n elements, where the m-th element is... ; represents the score given by the m-th expert subnet for the i-th subject; 3.3) Select the index set of k expert subnets using a gating algorithm: in, This represents a gating algorithm for selecting the index set of k expert subnets; This represents the set of indices of the k expert subnets selected by the gating algorithm for the i-th subject; 3.4) Calculate the index in Normalized gate probability of the expert subnet within the range: in, Indicates that the index is in Normalized gate probability of the expert subnet within the range; 3.4) Calculation : in, For the first A sub-network of experts.
4. The method for constructing a brain age prediction model combining auxiliary classification according to claim 3, characterized in that, It is a convolution-attention hybrid module consisting of 3D convolution, batch normalization, exponential linear activation, channel attention and pooling.
5. The method for constructing a brain age prediction model combining auxiliary classification according to claim 3, characterized in that, Specifically Gating algorithm.
6. The method for constructing a brain age prediction model combining auxiliary classification according to claim 3, characterized in that, 、 。 7. The method for constructing a brain age prediction model combining auxiliary classification according to claim 3, characterized in that, The age regression head is represented as The output of the age regression head is The formula is: The disease classification header is represented as follows: The output of the disease classification head is The formula is: in, Let be a C-dimensional normalized vector that satisfies ;in, express The c-th element in the equation represents the probability of the c-th type of disease. The actual age used as a monitoring signal is in the form of a standardized actual age, calculated using the following formula: in, Standardized actual age; For training set The formula for the average actual age in the data is: For training set The standard deviation of actual age in the formula is: Total loss The calculation formula is: in, The regression loss from the MAE perspective is calculated using the following formula: The regression loss from the perspective of MSE is given by the following formula: The regression loss from the difference ranking perspective is given by the following formula: in, ; The regression loss from the perspective of differentiable rank consistency is given by the following formula: in For temperature Differentiable sorting operators; The classification loss from the perspective of cross-entropy is given by the following formula: in To obtain the class weights after normalization, the formula is: in, For training set middle The number of samples equal to 1; express The c-th element in the equation represents the i-th subject's... Check if the c-th tag element is true; if true, the value is 1; if false, the value is 0. , , , , They are respectively , , , , The weighting coefficients.
8. The method for constructing a brain age prediction model combining auxiliary classification according to claim 7, characterized in that, The regression loss used as the monitoring indicator is .
9. A brain age prediction method, characterized in that, The subject's brain MRI data and gender are input into the brain age prediction model to obtain the subject's predicted brain age; wherein the brain age prediction model is obtained by the construction method according to any one of claims 1 to 8.
10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the method described in any one of claims 1 to 9.