OpenMAP-BrainAge: generalizable and interpretable brain age predictor from MRI Authors: Pengyu Kan, Craig Jones, Kenichi Oishi, Alzheimer’s Disease Neuroimaging Initiative and the Australian Imaging Biomarkers and Lifestyle Flagship Study of Aging Published: 2026 ABSTRACT AbstractBackgroundAccurately estimating brain age can help identify deviations linked to neurodegenerative diseases, underscoring the need for robust models that accurately perform across heterogenous cohorts.PurposeTo develop an age prediction model that is interpretable and robust to demographic and technological variations in brain MRI.Materials and MethodsWe propose a transformer-based brain age model that analyzes 3D T1-weighted MRI. Model performance was assessed using mean absolute error (MAE). Associations between brain age gap (BAG, ie, predicted minus chronological age) and chronological age were evaluated in cognitive normal (CN) participants. Clinical relevance was assessed by examining BAG differences across cognitive groups and correlations with Mini-Mental State Examination (MMSE) and Montreal Cognitive Assessment (MoCA).ResultsWe achieved an MAE 3.65 years on ADNI2 & 3 and OASIS3 test sets, and a high generalizability of MAE of 3.54 years on AIBL. In dementia, a notable increase in brain age gap (BAG) along with cognitive decline, with a mean of 0.15 years (95% CI: [−0.22, 0.51]) in CN, 2.55 years ([2.40, 2.70]) in mild cognitive impairment (MCI), and 6.12 years ([5.82, 6.43]) is noted. Negative correlation between BAG and cognitive scores was observed after adjustment for covariates, with r = −0.397 (P < 0.001) for MMSE and −0.393 (P < 0.001) for MoCA, where declining scores generally signify worsening cognitive performance. The saliency map highlighted white and deep gray matter structures as key regions influenced by brain aging.ConclusionOur model effectively integrated multiview and volumetric information to achieve state-of-the-art brain age prediction, with improved generalizability, interpretability, and association with cognitive function.Keywords: brain age prediction, deep learning, transformer, neurodegeneration, dementia Key Results OpenMAP-BrainAge is a transformer-based model, which can effectively fuse information from different views and volumetric information to achieve state-of-the-art brain age prediction accuracy with improved generalizability in linear computational complexity. OpenMAP-BrainAge is a transformer-based model, which can effectively fuse information from different views and volumetric information to achieve state-of-the-art brain age prediction accuracy with improved generalizability in linear computational complexity. We validated the relationship between brain age gap (BAG) and cognitive function, including dementia and mild cognitive impairment (MCI), and confirmed its association with cognitive decline. We validated the relationship between brain age gap (BAG) and cognitive function, including dementia and mild cognitive impairment (MCI), and confirmed its association with cognitive decline. The interpretability of our model verified the connection between white and deep gray matter structures with brain aging. The interpretability of our model verified the connection between white and deep gray matter structures with brain aging. Introduction The human brain undergoes complex structural changes with aging, including atrophy and ventricular enlargement.1,2 Because biological aging varies across individuals, brain aging may deviate from chronological age. The brain age gap (BAG) quantifies this deviation3 (Figure S1) and provides a clinically meaningful summary of cumulative neuroprotective and neurotoxic processes.4 Accurate BAG estimation potentially enables early risk stratification for neurodegenerative diseases such as mild cognitive impairment (MCI) and dementia,5–7 prognostication and monitoring of disease progression through longitudinal changes in brain aging, quantitative evaluation of treatment effects by detecting slowed or reversed brain aging, and enrichment and outcome assessment in clinical trials by identifying individuals with accelerated brain aging.4,8 BAG estimators have commonly utilized convolutional neural networks (CNNs) to extract spatial features from brain MRIs.5,9,10 While effective at capturing global structural patterns, CNNs may miss subtle, localized changes relevant to brain aging. Transformer-based models have therefore been introduced to better capture relationships across brain regions.3,6,7 However, limited generalizability across demographics and imaging protocols remains a challenge. In traditional transformers, computational cost increases with the square of the input dimensionality.11 For high-resolution 3D MRI volumes, this results in substantial memory and processing demands, which limit its scalability. To mitigate these challenges, we introduce OpenMAP-BrainAge. Leveraging the improved generalization of pretrained foundation models,12,13 we adopt a transformer-based foundation model,13 pretrained on 52 robotic proprioception and vision datasets spanning diverse environments and hardware configurations, which helps mitigate bias arising from variability in MRI acquisition. To balance local detail and global context while controlling model complexity, we employ pseudo-3D inputs sampled from sagittal, coronal, and axial views.3,7 We further incorporate volumetric features from 280 brain structures,14 whose changes are associated with aging.7 These multiview and regional features are integrated through a modified self-attention mechanism inspired by the stem-and-trunk design.13 The model was trained and tested on the Alzheimer’s Disease Neuroimaging Initiative 2 & 3 (ADNI 2&3)15 and Open Access Series of Imaging Studies 3 (OASIS 3)16 datasets, and externally tested for cross-domain robustness using the Australian Imaging Biomarkers and Lifestyle Study of Ageing (AIBL) dataset.17 Materials and methods OpenMAP-BrainAge OpenMAP-BrainAge is a transformer-based framework for estimating brain age from MRI-derived inputs (code is available at https://github.com/OishiLab/OpenMAP-BrainAge). The model takes as input (1) pseudo 3D T1-weighted brain MRI representations constructed from 30 central sagittal, coronal, and axial slices, and (2) quantitative 280 volumetric features extracted from whole-brain parcellation using OpenMAP-T1.14 It outputs a single continuous brain age prediction per subject (Figure 1A). The architecture integrates multiview image and volumetric information using attention mechanisms (Appendix A) and consists of encoders, stems, a trunk, and task-specific heads. MRI inputs are processed with a 3D ResNet1818 pretrained on the Kinetics-700,19 while volumetric features are embedded using a 2-layer multilayer perceptron (MLP). The stem module13 (Figure 1B) compresses tokenized image features into fixed-size representations using learnable queries, reducing computational cost (Appendix B). A transformer trunk pretrained on 52 robotic datasets13 integrates multiview and regional features via self-attention,11 and a regression MLP head predicts biological age. This design allows flexible adaptation to multiple data types while maintaining efficiency through shared modules and pretrained initialization. Architecture design of OpenMAP-BrainAge. (A) OpenMAP-BrainAge. The model takes pseudo 3D T1-weighted brain MRI representations constructed from central 30 slices in each of 3 anatomical views (image data) and quantitative 280 volumetric features extracted from whole-brain parcellation using OpenMAP-T114 (numerical data). A shared image encoder (*) extracts spatial features from each view, which are compressed by view-specific stem modules. A transformer trunk integrates embeddings across views and volumetric information, and the head predicts biological age. FC denotes a fully connected layer; MLP denotes multilayer perceptron. (B) Stem architecture. Stem module introduces a fixed number m of learnable queries (much fewer than the number n of tokens from the image input), as following Wang et al.13 Then, we capture the relationship between these learnable queries and the keys from the input through cross-attention to generate embeddings for the input. This design reduces the image input to a fixed dimensional representation and lowers computational complexity from quadratic to linear with respect to the size of input. Dataset We trained and validated our model using cognitively normal (CN) participants from the ADNI 2& 315 and the OASIS 3,16 aged 42 to 95. The ADNI was launched in 2003 as a public-private partnership, led by Principal Investigator Michael W. Weiner, MD. The primary goal of ADNI has been to test whether serial MRI, positron emission tomography (PET), other biological markers, and clinical and neuropsychological assessment can be combined to measure the progression of MCI and early Alzheimer's disease (AD). CN participants were assumed to have aligned biological and chronological ages and were randomly divided into 8: 1: 1 for train, validation, and test purposes (4631 scans from 1173 subjects, 584 scans from 145 subjects, and 543 scans from 146 subjects in each), with scans from a subject confined to one split. Scans acquired from a subject at different time points were treated as independent samples. Subjects who exhibited changes in cognitive status across time were excluded. For clinical correlation analyses, we additionally included 1831 scans from 512 subjects for dementia and 3934 scans from 761 subjects for MCI, defined by Clinical Dementia Rating (CDR; CN = CDR 0, MCI = CDR 0.5, and dementia CDR 1 and above. Overt and severe dementia (CDR 2∼3) was also included.15,16) All data were collected with institutional review board approval and informed consent; the de-identified dataset obtained from ADNI and OASIS were considered exempt from further review. Dataset details are in Table S1. 3D-acquired T1-weighted MRI scans were processed using OpenMAP-T1,14 including skull-stripping, rigid alignment to Montreal Neurological Institute (MNI) space, and downsampling to 2 × 2 × 2 mm resolution. Pseudo-3D inputs were generated by cropping 128 × 128 × 30 voxel volumes from each anatomical view, followed by min-max intensity normalization. Data augmentation during training included random rotations (±20°) and random transformations (±20 pixels), each applied with 50% probability. Model robustness and generalizability were evaluated using the CN group from the AIBL dataset,17 an independent Australian cohort containing 648 scans from 275 subjects (60 to 92 years old) with similar scanning protocols and preprocessing steps as the ADNI dataset. Data was collected by the AIBL study group. AIBL study methodology has been reported previously.17 Model training and evaluation Brain age prediction was formulated as a regression task. The model was trained to minimize the mean squared error (MSE) between predicted brain age and chronological age. Additional training details are provided in Appendix C. For clinical analyses, the brain age gap (BAG) was defined as the difference between predicted brain age and chronological age. Model performance was evaluated on the held-out test set using mean absolute error (MAE). To compare our model with previous published work, we retrained the following state-of-the-art neural networks under our setup: (i) MLP: a 2-layer fully connected network with ReLU activation to run the regression task. (ii) 3D ResNet1818,20: pretrained on the Kinetic-700 video dataset19 for classification. We changed the final fully connected layer to fit the brain age prediction task and fine-tuned all layers with our training data. (iii) Vision Transformer (ViT)21: we selected the ViT-Base model (12 layers, 768 hidden size, and 12 attention heads) pretrained through BERT self-supervised image pretraining (BEiT)12 on the ImageNet-1K dataset.22 The 2D convolutional patch embedding layer was replaced with a 3D version to handle MRI volumes. We fine-tuned all parameters over our training setup. (iv) Simple Fully Convolutional Network (SFCN)10: contained 7 convolutional layers, batch normalization, ReLU activation, and max pooling. We re-initialized the model with Kaiming uniform initialization23 and trained it over our training dataset. (v) Global–Local Transformer3: contained VGG8 as the image feature backbone to extract information from both the global and local pathways and an attention layer to merge the information. The Global–Local Transformer took 2D axial slices as input and averaged across these axial slices for the 3D scenario. To interpret model predictions, we computed gradient-based attribution maps for each test sample (Appendix D). The gradients were averaged across the test set to obtain a population-level importance map. To enable regional comparison, gradient values were normalized by the volume of each brain region. Regions were then ranked based on their normalized contributions, and the top contributing regions were identified for further analysis. Statistical analysis All statistical analyses were conducted in Python (SciPy and statsmodels). Unless otherwise specified, tests were 2‑sided with a significance threshold of P < 0.05, and continuous variables are reported as mean ± SD. Brain age prediction accuracy was assessed in CN participants by quantifying the association between predicted age and chronological age using Pearson’s correlation. Robustness was evaluated by comparing the performance between internal and external datasets under an unpaired 2-sample t-test. Computational cost is reported in GFlops, as a standardized measure of the number of floating‑point operations required by each model when evaluated on the same input, independent of hardware or implementation-specific optimizations. Model design choices were evaluated via ablation studies that compared single‑view versus 3‑view inputs and the inclusion of volumetric inputs, evaluating their impact on accuracy and robustness across internal and external sets. Differences in BAG across brain disorder groups (CN, MCI, and dementia) were summarized using group‑wise descriptive statistics (distribution center and 95% confidence intervals) for the top‑performing models. To quantify separability, we performed pairwise receiver operating characteristic (ROC) analyses and reported the area under the curve (AUC) with 95% confidence intervals. We also reported standardized effect sizes for BAG between groups to complement ROC metrics with Bonferroni adjustment. Associations between BAG and cognitive performance were evaluated using Pearson correlations with Mini‑Mental State Examination (MMSE) and Montreal Cognitive Assessment (MoCA) scores. We additionally computed partial correlations controlling for chronological age, sex, and years of education. For cognitive analyses, we used the assessment closest in time to the MRI within a prespecified window of 2 months. Results Performance on CN test set Table 1 compares the performance against state-of-the-art neural networks (Sec. 2.6) using 4 input types: (1) Volumetrics: the volumetric vector of 280-region parcellated by OpenMAP-T114; (2) Pseudo 3D—Coronal: central chunks from coronal view only; (3) Pseudo 3D—3-view + Volumetrics: combination of central chunks from all 3 anatomical views and the volumetric vector; and (4) Full 3D: the whole 3D MRI scans. Table 1Model comparison on brain age prediction in cognitive normal (CN) group of the test split (N = 543 scans from 146 subjects) of ADNI 2&315 and OASIS 3.16ModelsInputTest MAE+ SD (year) ↓Pearson r ↑(95% CI)Multilayer Perceptron (MLP)Volumetrics11.43± 8.970.33 (0.25, 0.40)3D ResNet18Pseudo 3D—Coronal3.78 ± 2.600.81 (0.79, 0.84)3D ResNet18Pseudo 3D—3-view + Volumetrics4.05± 2.690.78 (0.75, 0.81)3D ResNet18Full 3D 3.61 ± 2.700.82 (0.79, 0.85)3D ViTPseudo 3D—Coronal3.63 ± 2.790.81 (0.79, 0.84)3D ViTFull 3DNot applicableNot applicableSFCNFull 3D5.94 ±4.790.58 (0.52, 0.64)Global-Local TransformerFull 3D (average through 2D Axial slices)3.87 ± 2.820.74 (0.71, 0.77)OpenMAP-BrainAgePseudo 3D—3-view + Volumetrics3.65 ± 2.430.82 (0.79, 0.85)Mean ± SD is reported. Volumetrics-only features yield poor accuracy, whereas image-based models substantially improve performance. The proposed OpenMAP-BrainAge attains competitive mean absolute error (MAE) and the highest correlation, while using efficient pseudo-3D multiview inputs with volumetric information, demonstrating a favorable balance between accuracy and model complexity. Model comparison on brain age prediction in cognitive normal (CN) group of the test split (N = 543 scans from 146 subjects) of ADNI 2&315 and OASIS 3.16 Mean ± SD is reported. Volumetrics-only features yield poor accuracy, whereas image-based models substantially improve performance. The proposed OpenMAP-BrainAge attains competitive mean absolute error (MAE) and the highest correlation, while using efficient pseudo-3D multiview inputs with volumetric information, demonstrating a favorable balance between accuracy and model complexity. Our model achieved competitive accuracy (MAE = 3.65 years, r = 0.82) while requiring substantially lower computation (1.33 GFlops) than transformer-based models (65.43 GFlops for 3D ViT and 1793.48 GFlops for Global-Local Transformer). Figure 2 compares the 3 best-performing models. Both 3D ViT and 3D ResNet18 underestimated the older age group (> 85 years),10 whereas our model showed reduced bias (mean BAG = −2.21 years) compared to −3.70 years for 3D ViT and −3.60 years for 3D ResNet18 (Table 2). Although our model slightly overestimated ages for younger individuals (55-65 years) at 3.39 years compared to 2.86 years for 3D ViT and 2.44 years for 3D ResNet18, its overall stability was better with the lowest prediction variability (SD of 2.43 years) compared to 2.79 years for 3D ViT and 2.70 years for 3D ResNet18 (Figure 2), indicating more consistent and reliable age estimates. Stratified calibration curves are provided in Figure S2. Scatter of predictions versus actual ages in the cognitive normal (CN) group of the test split of ADNI 2&315 and OASIS 3.16 Values are reported as mean ± SD. (A) OpenMAP-BrainAge with “Pseudo 3D—three view + Volumetrics” input. (B) 3D ViT21 with “Pseudo 3D—Coronal” input. (C) 3D ResNet1818,20 with “Full 3D” input. OpenMAP-BrainAge achieves a mostly even scatter of predictions (dashed-circle region) on the elder age group (age over 85 years old), meanwhile other two models tend to underestimate the elder group though reaching a similar mean absolute error (MAE) in general. Further, OpenMAP-BrainAge has the lowest SD of error on predictions, which implies a higher confidence in the estimation. (A) MAE: 3.65 (±2.43) years, r = 0.82 (P < 0.001), slope = 0.75 (±0.02), intercept = 18.57 (±1.63) years. Pseudo 3D—3-view + Volumetrics. (B) MAE: 3.63 (±2.79) years, r = 0.81 (P < 0.001), slope = 0.73 (±0.02), intercept = 19.29 (±1.67) years. Pseudo 3D—Coronal. (C) MAE: 3.61 (±2.70) years, r = 0.82 (P < 0.001), slope = 0.78 (±0.02), intercept = 15.29 (±1.69) years. Full 3D. Table 2Mean ± SD of the brain age gap (BAG, years) for cognitive normal (CN) participants in the test split of ADNI2&3 and OASIS3 across age bins.ModelsInput55-65 (N = 45 scans from 22 subjects)65-75 (N = 260 scans from 86 subjects)75-85 (N = 188 scans from 58 subjects)85-95 (N = 41 scans from 11 subjects)3D ResNet18Full 3D2.44 ± 3.62−0.20 ± 4.20−2.19 ± 4.41−3.60 ± 2.413D ViTPseudo 3D—Coronal2.86 ± 2.991.06 ± 4.77−1.63 ± 3.47−3.70 ± 2.85OpenMAP-BrainAgePseudo 3D—3-view + Volumetrics3.39 ± 3.531.11 ± 4.13−1.52 ± 3.70−2.21 ± 5.19Values closer to zero indicate higher prediction accuracy. Our proposed model shows improved accuracy in older subjects (75-95 years) while maintaining comparable performance in younger age groups (55-75 years). Mean ± SD of the brain age gap (BAG, years) for cognitive normal (CN) participants in the test split of ADNI2&3 and OASIS3 across age bins. Values closer to zero indicate higher prediction accuracy. Our proposed model shows improved accuracy in older subjects (75-95 years) while maintaining comparable performance in younger age groups (55-75 years). Performance on external cohort On the AIBL CN cohort,17 our model maintained a performance with MAE = 3.54 years compared to the original 3.65 years (unpaired 2-sample t-test P = 0.442), with only a modest reduction in correlation (r from 0.82 to 0.76), shown in Table 3. In contrast, other models showed larger MAE increases (by 0.14 to 0.61 years) and correlation drops, demonstrating the superior robustness of our approach to domain shift. Table 3Model comparison on brain age prediction in cognitive normal (CN) group of the AIBL dataset (N = 648 scans from 275 subjects).17ModelsInputTest MAE+ SD (year) ↓Pearson r ↑(95% CI)3D ResNet18Pseudo 3D—Coronal4.01±3.040.71 (0.68, 0.74)3D ResNet18Pseudo 3D—3-view + Volumetrics3.98± 2.980.70 (0.66, 0.74)3D ResNet18Full 3D3.78 ± 2.950.74 (0.71, 0.77)3D ViTPseudo 3D—Coronal3.77 ±2.940.72 (0.69, 0.76)3D ViTFull 3DNot applicableNot applicableGlobal-Local TransformerFull 3D (average through 2D Axial slices)4.48± 3.540.64 (0.59, 0.69)OpenMAP-BrainAgePseudo 3D—3-view + Volumetrics 3.54 ±2.630.76 (0.73, 0.79)AIBL dataset is demographically different from the datasets during training. Our proposed model has a consistent performance under domain shifting. MAE denotes mean absolute error. Model comparison on brain age prediction in cognitive normal (CN) group of the AIBL dataset (N = 648 scans from 275 subjects).17 AIBL dataset is demographically different from the datasets during training. Our proposed model has a consistent performance under domain shifting. MAE denotes mean absolute error. Ablation analysis Ablation results (Table S2) showed that multiview inputs reduced MAE from 3.72 to 3.51 years. Adding volumetric information resulted in a slight increase in MAE on the ADNI 2&3 and OASIS 3 test set (from 3.51 to 3.65 years). However, volumetric features notably improved robustness under domain shift. On the external dataset (AIBL), the “three-view” model exhibited a performance drop from an MAE of 3.51 years to 3.70 years, whereas the “three-view + volumetrics” model maintained consistent performance (MAE = 3.54 years externally vs. 3.65 years in-house). Analysis of brain disorders BAG increased monotonically with diagnostic severity (Figure 3). In CN participants, BAG was centered near zero (mean = 0.15 years), whereas MCI and dementia groups showed progressively higher BAG values (means = 2.55 and 6.12 years, respectively). A few case images are included in Figure 4. Analysis of brain aging across different subject groups (cognitive normal (CN): N = 543 scans from 146 subjects; mild cognitive impairment (MCI): N = 3934 scans from 761 subjects; and dementia: N = 1831 scans from 512 subjects) of ADNI 2&315 and OASIS 3.16 In the first row, we plotted the prediction versus the actual age of each scan. Values are reported as mean ± SD. The solid diagonal line is the identity line representing perfect prediction, and the dashed line represents the linear regression trend. In the second row, we plotted the histogram of the brain age gap (BAG) for each group. The solid line is BAG=0 and the dashed line represents the mean of each distribution. The predictions show increasing overestimation and the mean BAG values are increasing as the subject groups suffer from more severe disorders. (A) MAE: 3.65 (±2.43) years, r = 0.82 (P < 0.001), slope = 0.75 (±0.02), intercept = 18.57 (±1.63) years. (B) MAE: 4.33 (±3.20) years, r = 0.79 (P < 0.001), slope = 0.71 (±0.01), intercept = 24.12 (±0.64) years. (C) MAE: 7.05 (±5.16) years, r = 0.62 (P < 0.001), slope = 0.42 (±0.01), intercept = 49.75 (±1.02) years. (D) Mean: 0.15 years. 95% CI: (−0.22, 0.51) years. (E) Mean: 2.55 years. 95% CI: (2.40, 2.70) years. (F) Mean: 6.12 years. 95% CI: (5.82, 6.43) years. Examples of brain MRI inputs (T1-weighted coronal and axial images respectively left to right) and corresponding predicted brain ages; chronological ages are reported in parentheses. Compared with 3D ResNet18 and 3D ViT (Figure 5), our model demonstrated greater separation between CN and MCI groups, with mean BAG values increasing from 0.15 to 2.55 years, compared with smaller shifts observed for the 3D ResNet18 (−0.89 to 1.46 years) and the 3D ViT (−0.08 to 2.06 years). For the comparison between CN and dementia, our model showed a mean BAG increase from 0.15 to 6.12 years. This magnitude of separation was comparable to that observed with the 3D ResNet18 (−0.89 to 5.83 years) and the 3D ViT (−0.08 to 5.94 years). The ROC and Cohen’s d standardized measure of effective size is reported in Table S3 and the associated Bonferroni-adjusted DeLong P-value is reported in Table S4. In ROC analyses, our model achieved the highest performance for CN vs MCI (AUC = 0.642, Cohen’s d = 0.510), slightly exceeding both 3D ResNet18 (AUC = 0.639, d = 0.441) and 3D ViT (AUC = 0.628, d = 0.405). However, for CN vs dementia and MCI vs dementia, performance was comparable to 3D ViT but slightly lower than 3D ResNet18. Statistically significant differences were observed only between our model and 3D ResNet18 in the dementia-related comparisons, while differences with 3D ViT were not significant. Comparison of brain aging across different subject groups in state-of-the-art models (cognitive normal (CN): N = 543 scans from 146 subjects; mild cognitive impairment (MCI): N = 3934 scans from 761 subjects; and dementia: N = 1831 scans from 512 subjects on ADNI 2&315 and OASIS 316). In the first row, we plotted the histogram of the brain age gap (BAG) for 3D ResNet18 model with “Full 3D” as input and the second row for 3D ViT model with “Pseudo 3D—Coronal” as input. In both rows, the solid line is BAG=0 and the dashed line represents the mean of each distribution. The larger the predicted BAG value increases across different disorder groups, the better the model performs in terms of differentiating brain disorders. Specifically, our proposed method showed a better separation between CN and MCI groups, and a comparable separation between the CN and dementia group as the 3D ResNet18 and 3D ViT models. (A) Mean: −0.89 years. 95% CI: (−1.26, −0.51) years. (B) Mean: 1.46 years. 95% CI: (1.29, 1.63) years. (C) Mean: 5.83 years. 95% CI: (5.51, 6.14) years. (D) Mean: −0.08 years. 95% CI: (−0.46, 0.30) years. (E) Mean: 2.06 years. 95% CI: (1.89, 2.23) years. (F) Mean: 5.94 years. 95% CI: (5.62, 6.26) years. We further examined the relationship between brain aging and cognitive function using MMSE (Figure S3a and b) and MoCA scores (Figure S3c and d). Increased predicted age showed a significant negative correlation with lower cognitive scores. Our model achieved the strongest correlation with r = −0.397 (P < 0.001) for MMSE and r = −0.393 (P < 0.001) for MoCA in comparison with top-performing models (Table 4), conditioning on demographic factors. Table 4Comparison of correlation between predicted BAG and cognitive scores (Mini-Mental State Examination (MMSE) and Montreal Cognitive Assessment (MoCA); N = 5811 scans from 971 subjects, with cognitive normal (CN) test split, mild cognitive impairment (MCI) and dementia group from ADNI2&3 and OASIS3) across different state-of-the-art models.ModelsInputMMSEMoCA3D ResNet18Full 3D r = −0.34895% CI: (−0.37, −0.32)(P < 0.001) r = −0.34695% CI: (−0.37, −0.32)(P < 0.001) 3D ViTPseudo 3D—Coronal r = −0.36695% CI: (−0.39, −0.34)(P < 0.001) r = −0.37095% CI: (−0.39, −0.35)(P < 0.001) OpenMAP-BrainAgePseudo 3D—3-view + Volumetrics r = −0.39795% CI: (−0.42, −0.37)(P < 0.001) r = −0.39395% CI: (−0.42, −0.37)(P < 0.001) The calculated correlation is conditioned on demographic factors including sex, age, and education level. Our proposed model showed the strongest negative correlation between the predicted BAG as compared to other models. Comparison of correlation between predicted BAG and cognitive scores (Mini-Mental State Examination (MMSE) and Montreal Cognitive Assessment (MoCA); N = 5811 scans from 971 subjects, with cognitive normal (CN) test split, mild cognitive impairment (MCI) and dementia group from ADNI2&3 and OASIS3) across different state-of-the-art models. r = −0.348 r = −0.348 95% CI: (−0.37, −0.32) 95% CI: (−0.37, −0.32) (P < 0.001) (P < 0.001) r = −0.346 r = −0.346 95% CI: (−0.37, −0.32) 95% CI: (−0.37, −0.32) (P < 0.001) (P < 0.001) r = −0.366 r = −0.366 95% CI: (−0.39, −0.34) 95% CI: (−0.39, −0.34) (P < 0.001) (P < 0.001) r = −0.370 r = −0.370 95% CI: (−0.39, −0.35) 95% CI: (−0.39, −0.35) (P < 0.001) (P < 0.001) r = −0.397 r = −0.397 95% CI: (−0.42, −0.37) 95% CI: (−0.42, −0.37) (P < 0.001) (P < 0.001) r = −0.393 r = −0.393 95% CI: (−0.42, −0.37) 95% CI: (−0.42, −0.37) (P < 0.001) (P < 0.001) The calculated correlation is conditioned on demographic factors including sex, age, and education level. Our proposed model showed the strongest negative correlation between the predicted BAG as compared to other models. Gradient-based feature attribution Gradient-based analysis (Appendix D) identified white matter and deep gray matter structures as dominant contributors to age prediction, consistent with prior studies.1,2 The anterior and posterior limbs of the internal capsule (ALIC and PLIC) were among the most influential regions (Figure S4) and showed significant age-related volume decline (Figure S5) with ALIC exhibiting r = −0.21 (P < 0.001) and PLIC with r = −0.27 (P < 0.001). Additional contributions from the caudate nucleus and thalamus were observed, aligning with known aging effects.24,25 Figure 6 visualizes these regions. Average gradient mapping of input with respect to training loss across the test split (N = 543 scans from 146 subjects). The anterior limb of the internal capsule (ALIC; orange arrowhead) and posterior limb of the internal capsule (PLIC; red arrow) are the top critical regions, determining our proposed model’s estimation (A–D). Caudate nuclei (pink dashed arrow) and thalami (blue arrow) are also highlighted as top 15 critical regions (E–H). (A) PLIC-R and ALIC-R. (B) AILC-R and ALIC-L. (C) PLIC-R and PLIC-L. (D) ALIC and PLIC. (E) Caudate-L and -R. (F) Caudate-L and -R. (G) Thalamus-L and -R. (H) Thalamus-L and -R. Discussion Accuracy and generalizability Our stem-and-trunk transformer architecture effectively integrates multiview MRI and volumetric inputs, enabling more efficient cross-view information synthesis than traditional CNN-based feature concatenation. Unlike CNNs, which restrict cross-view interactions through shared kernels, self-attention allows explicit modeling of relationships across anatomical views. Although some 3D transformer models achieve slightly lower MAE under specific conditions, they are less reliable in older age groups and substantially more computationally expensive. In contrast, heterogeneous pre-training on robotic proprioception and vision datasets improved robustness to acquisition variability, resulting in stable performance across cohorts, including the external AIBL dataset. Ablation analyses indicate that multiview inputs are sufficient for strong in-house accuracy, whereas adding volumetric features improves cross-cohort generalizability. This balance between accuracy and stability is particularly advantageous for real-world clinical deployment, where demographic and protocol heterogeneity are unavoidable. Categorical interpretability We observed a monotonic increase in BAG with diagnostic severity, progressing from cognitively normal participants to MCI and dementia. This pattern is consistent with prior studies3,5,7 and supports the clinical relevance of BAG as a marker of neurodegenerative burden. Importantly, previous work has emphasized that accuracy in cognitively normal populations alone is insufficient, as models must also differentiate disease-related deviations in brain aging.5 Our model demonstrated competitive separation between CN, MCI, and dementia groups, comparable to or exceeding prior reports,7,9 indicating sensitivity to pathological aging beyond normative variation. However, despite this, the model showed slightly lower AUC than 3D ResNet18 in dementia-related classifications. While BAG captures population-level shifts, AUC evaluates individual rankings. As such, our model appears more sensitive to continuous disease-related trends rather than optimized for binary classification. Functional interpretability BAG also showed significant negative correlations with MMSE and MoCA scores, reflecting worsening cognitive performance with accelerated brain aging. These associations have been infrequently examined in prior work,26 and the stronger correlations observed with our model suggest improved sensitivity to clinically meaningful cognitive differences. This improved sensitivity may result from the model’s ability to integrate multiview structural information with regional volumetric features, enabling it to capture aging-related changes closely linked to cognitive function. Biological interpretability The gradient map from our model predominantly highlighted the white matter, caudate, and thalami, consistent with prior cross-sectional and longitudinal studies.2,24,25,27 In particular, the ALIC and PLIC were listed among the most critical regions for age prediction, congruent with a prior study,1 which reported a strong linear decline in ALIC and PLIC volumes with age. The ALIC contains the frontopontine fibers involved in motor planning and coordination,28 the thalamocortical fibers that contribute to cognitive and executive functions,29,30 and corticostriatal and caudatoputaminal fibers involved in the reward system and motor modulation.31 The PLIC contains the corticospinal and corticobulbar tracts essential for motor functions,28 and the superior thalamic, optic, and auditory radiations responsible for transmitting sensory information.32 Research shows a significant linear negative correlation between age and the volume of these fiber tracts.1,33 The decline in the volume and integrity of the ALIC and PLIC is primarily linked to degeneration in myelinated fibers due to aging.34 Our model captured these biological changes, offering potential utility for early pathological assessments. Limitations and future work Multiple scans per subject were treated as independent samples during evaluation, which may slightly overestimate model performance. Although some prior studies3,7 report lower MSE when trained on large multi-cohort datasets, differences in cohort composition, age distribution, and evaluation protocols limit direct comparison. Within specific age ranges, performance across studies is often comparable, suggesting overall metrics may obscure age-dependent biases. Our current work is a cross-sectional study. Future longitudinal studies tracking individual brain age trajectories will establish the mechanistic relationship between BAG and dementia. We observed a decelerated trend in BAG among older individuals in the dementia group, potentially due to a survival effect or to the under-representation of subjects aged 85+ in the training dataset. Since the model learns age-related patterns only within the training age range (≤ 95 years), extrapolation beyond this range is limited. Consequently, BAG values may be artificially constrained at higher ages, reducing sensitivity to age-related deviations in older cohorts. Our focus on older adults and Alzheimer’s disease-related phenotypes enhances relevance to age-related neurodegeneration but limits generalizability to other neurological or psychiatric conditions. While BAG correlates with cognitive performance, it may not optimally discriminate between adjacent stages such as CN and early MCI. Future studies should evaluate broader populations and age ranges to assess external validity and age-by-disease interactions. Although we demonstrated generalization across independent cohorts, ADNI and AIBL use relatively harmonized MRI protocols. Robustness to more heterogeneous scanners, acquisition parameters, and real-world clinical workflows remains to be established. Additional validation across diverse imaging environments, potentially incorporating harmonization and domain adaptation strategies, will be important for clinical applications. BAG may also be influenced by lifestyle, vascular, and socioeconomic factors.4,26 While analyses adjusted for age, sex, and education, broader covariates were not comprehensively modeled, which may limit causal interpretation. Future studies incorporating further covariate measurements, including vascular risk, physical activity, genetic factors, and longitudinal outcomes, may clarify the biological pathways reflected by BAG. Importantly, BAG is not intended as a diagnostic biomarker for Alzheimer’s disease or other specific neurodegenerative etiologies. Established approaches, including volumetric MRI analysis, cerebrospinal fluid biomarkers, and amyloid or tau PET imaging,35 remain superior for molecular diagnosis. Instead, BAG should be interpreted as a quantitative marker of accelerated brain aging and neurobiological burden. Its clinical value lies in characterizing the downstream structural impact of dementia and potentially serving as a non-invasive adjunct for risk stratification, disease monitoring, and evaluation of physiological aging changes caused by Alzheimer’s disease. Conclusion We introduced the interpretable OpenMAP-BrainAge, a model that effectively combines different views and volumetric data using a stem-and-trunk architecture. Our model achieved comparable state-of-the-art accuracy and improved generalizability. It also reduced computational burden for better scalability with 3D MRI data. We found a positive association between BAG and cognitive decline including MCI and dementia. The BAG was also related to cognitive performance measures. Future work with longitudinal data and larger cohorts may enhance insights into brain aging and aid in quantification of physiological changes associated with Alzheimer’s dementia. Supplementary Material Acknowledgments Data collection and sharing for this project were funded by the Alzheimer’s Disease Neuroimaging Initiative (ADNI) (National Institutes of Health Grant U01 AG024904) and DOD ADNI (Department of Defense award number W81XWH-12-2-0012). ADNI is funded by the National Institute on Aging, the National Institute of Biomedical Imaging and Bioengineering, and through generous contributions from the following: AbbVie, Alzheimer’s Association; Alzheimer’s Drug Discovery Foundation; Araclon Biotech; BioClinica, Inc.; Biogen; Bristol-Myers Squibb Company; CereSpir, Inc.; Cogstate; Eisai Inc.; Elan Pharmaceuticals, Inc.; Eli Lilly and Company; EuroImmun; F. Hoffmann-La Roche Ltd and its affiliated company Genentech, Inc.; Fujirebio; GE Healthcare; IX-ICO Ltd.; Janssen Alzheimer Immunotherapy Research & Development, LLC.; Johnson & Johnson Pharmaceutical Research & Development LLC.; Lumosity; Lundbeck; Merck & Co., Inc.; Meso Scale Diagnostics, LLC.; NeuroRx Research; Neurotrack Technologies; Novartis Pharmaceuticals Corporation; Pfizer Inc.; Piramal Imaging; Servier; Takeda Pharmaceutical Company; and Transition Therapeutics. The Canadian Institutes of Health Research is providing funds to support ADNI clinical sites in Canada. Private sector contributions are facilitated by the Foundation for the National Institutes of Health (www.fnih.org). The grantee organization is the Northern California Institute for Research and Education, and the study is coordinated by the Alzheimer’s Therapeutic Research Institute at the University of Southern California. ADNI data are disseminated by the Laboratory for Neuro Imaging at the University of Southern California. Data collection in Open Access Series of Imaging Studies 3 is provided through OASIS-3: Longitudinal Multimodal Neuroimaging: Principal Investigators: T. Benzinger, D. Marcus, J. Morris; NIH P30 AG066444, P50 AG00561, P30 NS09857781, P01 AG026276, P01 AG003991, R01 AG043434, UL1 TR000448, R01 EB009352. AV-45 doses were provided by Avid Radiopharmaceuticals, a wholly owned subsidiary of Eli Lilly. Contributor Information Pengyu Kan, Department of Computer Science, Johns Hopkins University, Baltimore, MD, United States. Craig Jones, Department of Computer Science, Johns Hopkins University, Baltimore, MD, United States. Kenichi Oishi, The Russell H. Morgan Department of Radiology and Radiological Science, The Johns Hopkins University School of Medicine, Baltimore, MD 21205, United States; Department of Neurology, The Johns Hopkins University School of Medicine, Baltimore, MD, United States. Author contributions Pengyu Kan (Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Validation, Visualization, Writing—original draft), Craig Jones (Conceptualization, Methodology, Resources, Software, Supervision, Writing—review & editing), Kenichi Oishi (Conceptualization, Funding acquisition, Methodology, Project administration, Resources, Software, Supervision, Writing—review & editing) Supplementary material Supplementary material is available at Radiology Advances online. Funding This study was supported by the Richman Family Precision Medicine Center of Excellence in Alzheimer’s Disease, including significant contributions from the Richman Family Foundation, the Rick Sharp Alzheimer’s Foundation, the Sharp Family Foundation, and others. Conflicts of interest Kenichi Oishi is a consultant for “AnatomyWorks” and “Corporate-M.” This arrangement is being managed by the Johns Hopkins University in accordance with its conflict-of-interest policies. The remaining authors declare that they have no conflicts of interest to disclose. Data availability Data used in preparation of this article were obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu). As such, the investigators within the ADNI contributed to the design and implementation of ADNI and/or provided data but did not participate in analysis or writing of this report. A complete listing of ADNI investigators can be found at: http://adni.loni.usc.edu/wp-content/uploads/how_to_apply/ADNI_Acknowledgement_List.pdf Data used in the preparation of this article was obtained from the Australian Imaging Biomarkers and Lifestyle flagship study of ageing (AIBL) funded by the Commonwealth Scientific and Industrial Research Organisation (CSIRO) which was made available at the ADNI database (www.loni.usc.edu/ADNI). The AIBL researchers contributed data but did not participate in analysis or writing of this report. AIBL researchers are listed at data.aibl.org.au/adni/. References Associated Data Supplementary Materials Data Availability Statement Data used in preparation of this article were obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu). As such, the investigators within the ADNI contributed to the design and implementation of ADNI and/or provided data but did not participate in analysis or writing of this report. A complete listing of ADNI investigators can be found at: http://adni.loni.usc.edu/wp-content/uploads/how_to_apply/ADNI_Acknowledgement_List.pdf Data used in the preparation of this article was obtained from the Australian Imaging Biomarkers and Lifestyle flagship study of ageing (AIBL) funded by the Commonwealth Scientific and Industrial Research Organisation (CSIRO) which was made available at the ADNI database (www.loni.usc.edu/ADNI). The AIBL researchers contributed data but did not participate in analysis or writing of this report. AIBL researchers are listed at data.aibl.org.au/adni/.