Softmax-Derived Brain Age Mapping: An Interpretable Visualization Framework for MRI-Based Brain Age Prediction

Abstract

Brain age has been widely recognized as an important biomarker for monitoring adolescent brain development and assessing dementia risk. However, existing model visualization methods primarily highlight brain regions associated with aging, making it difficult to comprehensively reveal broader brain changes. In this study, we developed a VGGNet-based brain age prediction model and proposed the Softmax-Derived Brain Age Mapping algorithm to simultaneously identify brain regions associated with both youthful and aging features. The resulting saliency maps provide explicit representations of developmental and degenerative processes across different brain regions. Brain Age Map analysis revealed that aging features in the healthy group were primarily confined to the frontal cortex, aligning with findings that the frontal lobe is the earliest region to undergo natural senescence. In contrast, the dementia group exhibited widespread aging across the frontal, temporal, parietal, and occipital lobes, as well as the ventricular regions. These results suggest that the spatial distribution of brain aging can serve as a critical biomarker for distinguishing normal aging trajectories from pathological degeneration. From an application perspective, we further explored the potential of the proposed framework in neurodegenerative diseases. The analysis reveals that dementia patients generally exhibit an advanced brain age, with cortical aging being markedly more pronounced than in age-matched healthy samples. Notably, although dementia cases were not included in the training set, the model was still able to localize abnormalities in relevant brain regions, underscoring its potential value as an assistive tool for early dementia diagnosis.

1. Introduction

With the intensification of declining birth rates and population aging, the demand for precision medicine has become increasingly urgent. Predicted Age Difference (PAD) has been shown to be highly associated with cognitive impairment and dementia, making brain age prediction an important tool for assessing neurodevelopment and identifying early cognitive decline. Ref. [1] demonstrated that deep learning models based on Magnetic Resonance Imaging (MRI) can accurately predict brain age, providing reliable and quantifiable indicators for adolescent brain development. Similarly, refs. [2,3,4] reported that PAD is strongly linked to cognitive decline, mild cognitive impairment (MCI), and Alzheimer’s disease (AD) in middle-aged and elderly populations. Consequently, brain age has increasingly been adopted as a formal clinical indicator in recent years.

Most existing models output brain age as a single scalar value, which limits their diagnostic applicability. If visualization techniques are integrated to elucidate how predictions are generated and to map the key brain regions involved, brain age prediction models would no longer serve merely as numerical indicators but rather as comprehensive frameworks that enable region-specific analyses of brain development and aging. Particularly for the early detection of dementia, a model capable of accurately tracking abnormal regional aging would provide substantial clinical value by enabling timely intervention, thereby helping to slow disease progression and improve patient quality of life. Motivated by these considerations, this study proposes a deep learning-based brain age prediction framework that integrates brain MRI with a key region identification algorithm. The framework is capable of assessing brain development in adolescents and assisting in the identification of cognitive decline risk in older adults, holding significant implications for monitoring child development and preventing dementia.

The visualization technique Grad-CAM (Gradient-weighted Class Activation Mapping), proposed by [5], has been widely applied for highlighting salient regions in medical imaging and has inspired multiple variants. In the context of brain age prediction, ref. [6] employed the Guided Backpropagation [7] algorithm and found that the model primarily focused on cerebrospinal fluid, particularly the lateral ventricles, underscoring their importance in brain age estimation. Similarly, ref. [8] utilized SmoothGrad, another gradient-based technique, and identified periventricular, temporal, and insular regions as being highly influential in brain age prediction. However, these approaches face inherent limitations when applied to brain age prediction. Since the output of brain age prediction is a single scalar value, gradient-based visualization techniques such as Grad-CAM and its variants typically highlight only regions strongly associated with aging features. To overcome this limitation, we propose a refined algorithm that can simultaneously identify regions associated with both youthful and aging features, thereby providing more comprehensive information in brain saliency maps.

2. Related Works

We reviewed recent literature on brain age prediction and summarized the best-performing models (

Table 1. Summary of Best-Performing Brain Age Prediction Models.

Articles	Dataset (n)	Training/Test (n)	Test Set Age Range (Years)	MAE (Years)
[2]	DeCODE (1469) a, IXI (440), UK biobank (12,395)	1909/12,395 a	45–80	3.63
[3]	LIFE (2354)	1177/1177	19–82	4.29
[4]	BAHC (2001) b	1601/200	18–90	4.16
[11]	CamCAN (483)	-	18–88	5.47
[12]	ABIDE I (228), ABIDE II (505), BNU (328), Berlin (142), Cleveland Clinic (150), IXI (544), ADNI (219), Train-39 (135)	1801/225	6–90	3.56
[13]	Healthy Group (1182) c	850/332	19–84	4.60
[14]	CamCAN (651)	521/130	18–89	7.46
[15]	IXI (459), CoRR (266), OASIS (264) ABIDE I (258), ABIDE II (217), Local Centers (118)	1464/118	18–94	3.08
[6]	Local Hospital (19,807)	15,146/4661	18–95	2.97
[16]	CamCAN (613)	-	18–88	4.88
[9]	Rotterdam Study (8768)	5865/550	46–96	4.45
[17]	BAHC (2003) d, CamCAN (648)	-	18–88	5.08

^a DeCODE and IXI are used as training datasets, while the UK Biobank is used as the test dataset. ^b BAHC includes 14 publicly available datasets: ABIDE (184), Beijing Normal University (179), Berlin School of Brain & Mind (49), CADDementia (12), Cleveland Clinic (31), ICBM (322), IXI (561), MCIC (93), MIRIAD (23), NEO2012 (39), NKI (160), OASIS (288), WUSL (24), TRAIN-39 (36). ^c The Capital Medical University Xuanwu Hospital, in collaboration with 15 universities and hospitals nationwide, created a self-built dataset of whole-brain sMRI images collected from healthy volunteers using 3.0T MRI scans. ^d BAHC includes 14 publicly available datasets: ABIDE (184), Beijing Normal University (181), Berlin School of Brain & Mind (49), CADDementia (12), Cleveland Clinic (31), ICBM (322), IXI (561), MCIC (93), MIRIAD (23), NEO2012 (39), NKI (160), OASIS (288), WUSL (24), TRAIN-39 (36).

and

Table 2. Image Processing Methods and Models of Top Brain Age Prediction Models.

Articles	Preprocessing Methods	Data Augmentations	Model Architecture
[2]	GM/WM segmentation, bias correction, skull stripping, rigid registration, Jacobian maps	Rotation of 0–40°, Translation of 10 voxels.	ResNet WB, Jacobian maps, GM and WM combination model (10-fold cross-validation).
[3]	Skull stripping, linear registration, filtering, nonlinear registration, mcflirt, GM/WM segmentation, bias correction, denoising		SVR + Random Forest Multimodal stacking model (connectivity matrix 197, connectivity matrix 444, cortical thickness, cortical surface area, subcortical volumes)
[4]	GM segmentation, motion artifact removal, nonlinear registration, resampling include modulation, spatial smoothing	Rotation of 0–40°, Translation of 10 pixels.	CNN GM normalized volume map model
[11]	T1-MRI: GM/WM/CSF segmentation, skull stripping, linear registration, spatial normalization, bias correction rs-fMRI: Removal of initial volumes, slice-timing correction, motion correction, rigid-body registration, spatial normalization, Gaussian smoothing, denoising, temporal filtering		BrainDCNw Dual-modality (weighted rs-fMRI functional connectivity + weighted T1 brain volume) model. (10-fold cross-validation).
[12]	Skull stripping, linear registration, GM segmentation, resample and align	Weighted MAE Loss (assign higher importance to samples from underrepresented age groups).	Lightweight CNN GM model.
[13]	Brain region segmentation, skull stripping, affine registration, volume normalization		RFBLSO Brain region normalized volume feature model.
[14]	Bias correction, skull stripping, affine registration, spatial smoothing, spatial normalization		Tri-UNet WB model.
[15]	GM/WM segmentation, skull stripping, linear registration	Rotation of 0–40°, Translation of 10 voxels.	VGG-13 WB, GM and WM combination model (10-fold cross-validation).
[6]	Cropping to the same size, intensity normalization e		DenseNet121 WB (non-skull-stripped) model.
[16]	MRI: GM/WM/CSF segmentation, cortical segmentation, skull stripping, nonlinear registration, Z-score normalization MEG: Filtering, resampling, denoising, leakage correction, source localization, brain region segmentation		CCA+GPR MRI-MEG feature stacking model (10-fold cross-validation)
[9]	GM/WM/CSF segmentation, nonlinear registration, spatial modulation	Single-subject longitudinal model (one subject with one or multiple MRI scans from different time).	CNN GM density map model.
[17]	GM/WM segmentation, smoothing, affine registration, nonlinear registration, spatial normalization, modulation, bias correction		SVR GMV and WMV combination model.

^e T2-w MRI is used as the input, and for other models, if not specified, T1-w MRI is used as the input.

). The results indicate that preprocessing, data augmentation, model architecture, and training strategies all play critical roles in determining model accuracy and generalizability. Common preprocessing techniques include the segmentation of gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF), skull stripping, and image registration, which allow models to focus on brain structures most associated with age-related changes. Bias correction and intensity normalization further help to reduce imaging heterogeneity and smooth the loss surface, thereby facilitating model convergence. Data augmentation methods, such as rotation and translation, are frequently employed to enhance data diversity and reduce overfitting. Model architecture is a critical factor influencing predictive performance. Although conventional convolutional neural networks (CNNs) have achieved certain success in brain age prediction, their performance has not been optimal. As shown in Table 2, CNN-based models such as those reported by [4,9] yielded higher MAE values, indicating that traditional CNNs may not be the best choice for this task. Recent studies suggest that lightweight models exhibit stronger generalizability and improved accuracy, while ensemble approaches can further enhance reliability and mitigate overfitting [10].

However, evidence from the experimental results of [18,19] suggests that data augmentation does not necessarily provide positive benefits. Ref. [18] reported that, under standardized VBM preprocessing, data augmentation offered no significant advantages, and excessive augmentation could even disrupt the model’s ability to learn brain age-related imaging features. Similarly, ref. [19] argued that the introduction of additional spatial variations during MRI acquisition (e.g., subject head motion) substantially reduces model accuracy, indicating that when imaging quality and spatial alignment are already highly standardized, further augmentations such as rotations or translations may act as sources of noise rather than improvements. The effectiveness of data augmentation in brain age prediction remains uncertain under standardized preprocessing.

3. Experiments

This study proposes an interpretable visualization-driven brain age prediction model inspired by the VGGNet architecture, enabling fast and accurate brain age estimation from T1-weighted MRI scans. The proposed Softmax-Derived Brain Age Mapping (SD-BAM) is further applied to generate Brain Age Maps, providing visual clinical support (Figure 1).

3.1. Data Sets

This experiment utilized T1-weighted 3D MRI scans from four publicly available neuroimaging datasets: ABIDE, ADNI, CamCAN, and IXI (

Table 3. Statistical Data of Healthy Samples from Each Dataset.

	Samples	Average Age ± Standard Deviation	Age Range	Female (%)
ABIDE	561	17.1 ± 7.7	6–56	17.6%
ADNI	562	71.5 ± 7.2	51–90	59.4%
CamCAN	653	54.8 ± 18.6	18–89	50.5%
IXI	563	48.7 ± 16.5	20–86	55.6%

). The CamCAN dataset was obtained from the CamCAN repository [20,21]. These datasets cover an age range of 6 to 90 years (Figure 2), enabling the model to learn structural brain changes across the entire human lifespan. In total, 2339 MRI scans were integrated, with a broad age span and balanced sample distributions across datasets, each containing more than 500 subjects, thereby ensuring sufficient training resources.

3.2. Data Preprocessing

Prior to model training, all T1-weighted 3D brain MRI scans were subjected to preprocessing. Since the proposed model performs six successive downsampling operations by a factor of two, the input images must have edge lengths divisible by 64 (i.e., $2^6$) to avoid boundary loss. Therefore, we first preprocessed the MNI template (189 × 197 × 233) by resampling it to 128 × 192 × 128 and subsequently used this template for image registration to ensure that all input images conformed to the required dimensions. The preprocessing pipeline for whole-brain (WB) features is as follows. First, non-brain tissues were removed (skull stripping) using the U-net-based brain extraction model provided by ANTsPyNet [22]. Next, each image was rigidly registered to the preprocessed MNI152 template using SimpleITK 2.5.0. Finally, image intensity values were scaled using the 5th and 95th percentiles and normalized to the range [0, 1] (Figure 3).

3.3. Model Architecture Design

Our model design is inspired by VGGNet [23] and further guided by the architectures proposed by [10,24], where fully connected layers are replaced with convolutional layers to achieve a lightweight structure. The model consists of three main stages: feature extraction, nonlinear enhancement, and output adjustment. In the feature extraction stage, five consecutive convolutional blocks (Conv Blocks) are employed. Within each Conv Block, the input passes sequentially through a convolutional layer (Conv Layer), a batch normalization layer (BN), a max pooling layer, and a ReLU activation function. The nonlinear enhancement stage operates similarly to the Conv Blocks, except that the convolutional kernel size is modified from 3 × 3 × 3 to 1 × 1 × 1. Finally, a global average pooling layer and a 1 × 1 × 1 convolutional layer are used in the output adjustment stage to generate continuous brain age predictions. Two different output adjustment strategies were adopted for the brain age prediction task, namely the regression strategy and the softmax strategy. Both strategies share identical feature extraction and nonlinear enhancement modules, with the key difference lying only in the design of the output adjustment block.

Table 4. Model Architecture with Regression Strategy.

Layer Configuration	Input → Output (Channels)
Conv → BN → Pool → ReLU	1 → 32
Conv → BN → Pool → ReLU	32 → 64
Conv → BN → Pool → ReLU	64 → 128
Conv → BN → Pool → ReLU	128 → 256
Conv → BN → Pool → ReLU	256 → 256
Conv (1 × 1 × 1) → BN → Pool → ReLU	256 → 64
Global Average Pooling	-
Dropout	-
Conv (1 × 1 × 1)	64 → 1

presents the model architecture using the regression strategy. In this configuration, the output adjustment block consists of a global average pooling layer followed by a 1 × 1 × 1 convolutional layer, which transforms the extracted features into a single-channel continuous brain age prediction value. Compared with conventional deep networks such as ResNet [25] and DenseNet [26] (implemented via the MONAI framework “https://github.com/Project-MONAI/MONAI (accessed on 18 December 2025)”as ResNet-18 and DenseNet-18, see

Table 5. Comparison of Parameter Counts between the Proposed Model and Other Architectures.

Article	Model Architecture	Parameter Counts
Ours	Ours	2,953,586
[25]	ResNet-18 (3D)	33,186,546
[26]	DenseNet-121 (3D)	11,293,874

), our architecture is significantly more lightweight and streamlined. By avoiding excessively deep layer stacking, the substantial reduction in parameters effectively mitigates the risk of overfitting.

Table 6. Model Architecture with Softmax Strategy.

Layer Configuration	Input → Output (Channels)
Conv → BN → Pool → ReLU	1 → 32
Conv → BN → Pool → ReLU	32 → 64
Conv → BN → Pool → ReLU	64 → 128
Conv → BN → Pool → ReLU	128 → 256
Conv → BN → Pool → ReLU	256 → 256
Conv (1 × 1 × 1) → BN → Pool → ReLU	256 → 64
Global Average Pooling	-
Dropout	-
Conv (1 × 1 × 1)	64 → 50
Softmax	-
Weighted Sum	50 → 1

presents the model architecture using the softmax strategy, which transforms the task from regression to classification. In this configuration, the output adjustment block also consists of a global average pooling layer followed by a 1 × 1 × 1 convolutional layer. Unlike the single-channel output of the regression strategy, the convolutional layer here produces a 50-dimensional vector. This vector is then passed through a softmax function to generate the probability distribution over age classes (

Table 7. Softmax Layer Output Illustration.

Output of the Softmax Layer
Channel $\left(c\right)$	1	2	3	…	48	49	50
Age Class ${(a}_c)$	2	4	6	…	96	98	100
Probability ${(p}_i)$	p1	p2	p3	…	p48	p49	p50

), and the final predicted brain age is computed using a weighted-sum approach.

The predicted brain age in the softmax strategy is computed using a weighted-sum approach as follows:

$$\widehat{y} =\sum _{c = 1}^{50}{a}_c\times p_c$$

(1)

where $\widehat{y}$ denotes the predicted brain age, $a_c$ represents the age corresponding to class $c$, and $p_c$ is the probability of class $c$ being the predicted age.

3.4. Softmax-Derived Brain Age Mapping

To identify the brain regions most influential in the model’s brain age estimation, we designed a visualization algorithm based on Grad-CAM++ (Gradient-weighted Class Activation Mapping++) [27] and Guided Backpropagation [7] tailored to the softmax strategy. This algorithm generates heatmaps of key regions for individual samples, referred to as Brain Age Maps (BAM). We denote this method as Softmax-Derived Brain Age Mapping (SD-BAM), and its architecture is illustrated in Figure 4.

Recalling the output structure of the softmax layer (Table 6), the proposed algorithm computes the activation maps for each age category’s probability by taking the element-wise product of Grad-CAM++ and Guided Backpropagation. We define this resultant map as $L^c$. The underlying principle is as follows: Grad-CAM++ localizes regions that contribute positively to brain age prediction, while Guided Backpropagation captures the fine-grained edge features within those regions. The formula for calculating $L^c$ is defined as follows:

$$L^c={\left.{\displaystyle \frac{\partial p^c}{\partial I_{\mathit{ijk}}}}\right|}_{\mathit{guided} \mathit{ReLU}}⨀ \mathit{UpSample} \left(\sum _d{w}_d^c\cdot A_{\mathit{ijk}}^d\right)$$

(2)

where $p^c$ is the model-predicted probability for class $c$, $I_{\mathit{ijk}}$ represents the input image, and $A_{\mathit{ijk}}^d$ represents the $d$-th channel’s feature map. The term $w_d^c$ is the weight coefficient for the $d$-th channel regarding class $c$ (derived by Grad-CAM++), which is calculated following the Grad-CAM++ formulation [27] as follows:

$$w_d^c=\sum _i\sum _j\sum _k{\displaystyle \frac{\frac{{\partial }^2{p}^c}{{\left(\partial A_{\mathit{ijk}}^d\right)}^2}}{2\cdot \frac{{\partial }^2{p}^c}{{\left(\partial A_{\mathit{ijk}}^d\right)}^2}+{\sum }_a{\sum }_b{\sum }_t{A}_{\mathit{abt}}^d\cdot \frac{{\partial }^3{p}^c}{{\left(\partial A_{\mathit{ijk}}^d\right)}^3}}}\cdot \mathit{ReLU}({\displaystyle \frac{\partial p^c}{\partial A_{\mathit{ijk}}^d}})$$

(3)

The Softmax-Derived Fusion Map (SD-Fusion Map, B) is defined as in (4). Through softmax-weighted fusion, this map reveals both aging and youthful features. The negative regions (blue) indicate features that are strongly associated with brains under the age of 50, whereas the positive regions (red) correspond to features that are strongly associated with brains over the age of 50.

$$B =\sum _{c=1}^{50}\left({\mathit{age}}^c - 50\right)\cdot p^c\cdot L^c$$

(4)

where ${\mathit{age}}^c$ denotes the age corresponding to class $c$ and $L^c$ represents the Guided Grad-CAM++ activation map of class $c$. Subsequently, the final Brain Age Map (BAM) is obtained through tanh normalization and threshold processing, thereby accentuating the most significant brain regions associated with youthful and aging features:

$$\mathit{BAM} = \raisebox{1ex}{$\mathit{Threshold}\left(\mathit{tanh}\left({\displaystyle \frac{B⨀\left|\nabla B\right|}{\lambda }}\right)\cdot \lambda \right)$}\!\left/ \!\raisebox{-1ex}{$\mathit{percentile} = 0.995$}\right.$$

(5)

where $B$ denotes the SD-Fusion Map derived from the input image, and $\lambda$ represents the scaling factor used to clip the pixel intensities within the interval $\pm \lambda$.

3.5. Training Method

The experiments were conducted using the widely adopted 5-fold cross-validation method, as in previous studies [10,28,29], to train the brain age prediction models. During inference, the final brain age estimate was obtained by averaging the outputs of the five fold-specific models (one trained on each fold). Following the approaches of [2,10,15,16], data augmentation was applied to 70% of the training samples. The augmentation procedures included gamma correction with a correction factor ranging from 0.5 to 2.0 [30], translations of 3–5 voxels, and random rotations of 3–5°. To further examine its effect, additional experiments were conducted without data augmentation for comparison. The training hyperparameters were set as follows: batch size = 4, learning rate = 0.001, and 300 epochs.

4. Experimental Results and Discussion

4.1. Models Evaluation Methods

To quantitatively assess the performance of the proposed brain age prediction models, two primary evaluation metrics were employed: mean absolute error (MAE) and predicted age difference (PAD). The MAE is defined as follows:

$$\mathit{MAE} = {\displaystyle \frac{1}{n}} \sum _{i=1}^n {\widehat{y}}_i-y_i$$

(6)

where ${\widehat{y}}_i$ represents the predicted brain age for the $i$-th subject, $y_i$ denotes the corresponding chronological age, and $n$ is the total number of samples.

In addition, PAD is calculated as the difference between the predicted and actual chronological age:

$$\mathit{PAD} = \widehat{y}-y$$

(7)

This metric is particularly useful for evaluating whether the model exhibits age-dependent bias, such as consistently overestimating or underestimating brain age in specific age groups.

4.2. SD-BAM Evaluation Method

To quantitatively evaluate the regional characteristics of the proposed SD-BAM, a region-of-interest (ROI) analysis was conducted using the Automated Anatomical Labeling (AAL) atlas. The AAL template provides a standardized anatomical parcellation for anatomically consistent evaluation across subjects. For each subject, voxel-wise SD-BAM attention maps were aggregated within each atlas-defined region. The ROI attribution value was computed as the mean attention over all voxels in the region:

$${\mathit{ROI}}_r = {\displaystyle \frac{1}{\left|V_r\right|}} \sum _{v\in V_r}\mathit{BAM}\left(v\right)$$

(8)

where $\mathit{BAM}\left(v\right)$ denotes the SD-BAM attention value at voxel $v$, and $V_r$ represents the set of voxels within ${\mathit{ROI}}_r$. Positive and negative attributions were analyzed separately to characterize regions that contribute toward or against the predicted brain age. Since rigid registration was employed in the preprocessing stage, inter-subject anatomical variability could not be fully compensated at a fine-grained regional level. As a result, ROI analysis was restricted to coarse anatomical subdivisions to ensure robustness and anatomical consistency.

Specifically, the brain was partitioned into four broad regions: the anterior region (including the frontal lobe), the posterior region (including the occipital lobe and posterior parietal areas), the lateral region (including the temporal lobe), and the central region (including the thalamus, hippocampus, basal ganglia, and periventricular areas). This coarse-level regional grouping mitigates potential misalignment effects introduced by rigid registration while preserving interpretable anatomical trends for SD-BAM evaluation.

4.3. Comparisons of Regression and Softmax Performance

To investigate the effect of dropout rate [31] on different output layer designs, we conducted comparative experiments on models with both regression and softmax strategies. Considering that most studies have employed the Adam optimizer [24,32], while a few have adopted SGD [10], both optimizers were included as variables in this experiment. Apart from these factors, all models shared the same network architecture and dataset settings, and their performance was evaluated under the 5-fold ensemble framework.

As shown in

Table 8. Comparison of Regression and Softmax Performance.

Model	Optimizer	Dropout Rate	Training Set MAE	Test Set MAE
Regression	SGD	0	1.19	4.66
	Adam	0	1.19	4.34
	Adam	0.2	1.42	4.38
Softmax	SGD	0	1.13	4.43
	Adam	0	1.12	4.38
	Adam	0.2	1.06	4.33
	Adam	0.5	1.26	4.35

, both strategies achieved better training performance when using the Adam optimizer compared with SGD. We attribute this to the fact that the learning rate of standard SGD does not decay automatically during the training iterations. For the regression strategy, the best results were obtained with a dropout rate of 0. This can be explained by its single-class output design, in which the absence of a softmax layer after dropout causes the perturbation to directly and more significantly affect the final regression output, thereby degrading performance. In contrast, the softmax strategy performed best with a dropout rate of 0.2, whereas its performance began to decline at a dropout rate of 0.5. These findings suggest that the softmax layer is more tolerant to moderate dropout rates and, compared with the regression strategy, is better able to exploit dropout for alleviating overfitting. Overall, the softmax strategy achieved performance comparable to that of the regression strategy under optimal hyperparameter settings. Given that the regression-by-classification approach of the softmax strategy enhances model interpretability through weighted averaging, we adopted the softmax strategy for subsequent experiments.

4.4. Evaluation of Data Augmentation

This section presents evaluation plots of the model performance on the test set, with a focus on examining the impact of data augmentation on brain age prediction. Figure 5 shows scatter plots of predicted versus chronological ages, where the gray line represents the reference line (x = y) and the red line indicates the regression line of the model predictions. Figure 5a,b correspond to the results with and without data augmentation, respectively. It can be clearly observed that the predictions of the model without augmentation align more closely with the reference line, indicating higher overall accuracy. In contrast, the predictions from the augmented model display greater divergence, suggesting that data augmentation may reduce predictive accuracy under the conditions of this study.

Figure 6 presents the MAE distribution plots, where Figure 6a,b correspond to the prediction results with and without data augmentation, respectively. It can be observed that in the model without augmentation, outliers with MAE greater than 15 occurred less frequently, indicating reduced instances of extreme prediction errors. Moreover, the denser distribution of points within the MAE < 5 range suggests higher prediction accuracy for the majority of samples.

Table 9. Comparison of Model Performance with and Without Data Augmentation.

	Test Set with Data Augmentation	Test Set Without Data Augmentation
ABIDE	2.85	2.77
ADNI	4.33	4.14
CamCAN	5.53	4.90
IXI	4.34	4.33
Whole Test Set	4.33	4.08

compares the performance of models trained with and without data augmentation on the test sets of ABIDE, ADNI, CamCAN, and IXI, as well as on the overall test set. In all cases, the models without augmentation demonstrated superior predictive performance.

For the overall test set, the model with data augmentation yielded an MAE of 4.33, whereas the model without augmentation achieved a reduced MAE of 4.08. The most notable improvement was observed on the CamCAN test set, where the MAE decreased from 5.53 to 4.90. Improvements were also seen in the ADNI and ABIDE datasets, where the MAE decreased from 4.33 to 4.14 and from 2.85 to 2.77, respectively, indicating that data augmentation failed to provide performance gains under the current parameter settings.

Table 10. Comparison of Top Models in Brain Age Prediction Based on MAE.

Articles	Training/Test (Images)	Test Set Age Range (Years)	Test Set MAE (Years)
Ours	1871/468	6–90	4.08
Ours	1871/110 a	6–56	2.77
[12]	1801/225	6–90	3.56
[13]	850/332	19–84	4.60
[15]	1464/118	18–94	3.08
[6]	15,146/4661	18–95	2.97
[9]	5865/550	46–96	4.45
[2]	1909/12,395	45–80	3.63
[3]	1177/1177	19–82	4.29
[4]	1601/200	18–90	4.16

^a ABIDE only.

presents a comparison between this study and related works. On the ABIDE test set, our model achieved an MAE of 2.77 years, which is comparable to large-scale studies such as ref. [2] (3.63 years), ref. [15] (3.08 years), and ref. [6] (2.97 years). Overall, this study demonstrates strong data efficiency: using only 1871 MRI scans, our model achieved accuracy close to that of large-scale studies. It should be noted that performance comparisons must be conducted with caution due to variations in datasets and preprocessing methodologies. Furthermore, this study opted to analyze the ABIDE dataset independently to highlight the model’s superior predictive performance within that specific age range.

We summarized studies that exclusively used CamCAN as the test set in

Table 11. Comparison of Related Studies Using CamCAN as Test Set.

Model	Training/Test (n Images)	Test Set MAE (Years)
[11]	-	5.47
[14]	521/130	7.46
[16] (MRI + MEG)	-	4.88
[16] (MRI only)	-	5.33
[17]	2003/648	5.08
Ours	1871/135	4.90

. Because CamCAN covers a wide age range, brain age prediction on this dataset is particularly challenging. In this study, by incorporating multiple datasets to expand the training set, we achieved an MAE of 4.90 on the CamCAN test set, outperforming most existing studies. Notably, ref. [16] combined brain MRI with MEG to reduce the MAE to 4.88 years; however, this approach required additional MEG acquisition, which increases both participant burden and cost. When relying on MRI alone, their model yielded an MAE of 5.33 years. By contrast, our model achieved a comparable level of accuracy using MRI only. These results demonstrate that augmenting training data with multiple datasets can overcome the limitations of insufficient samples and enhance predictive performance on specific datasets.

4.5. Clinical Applications and Efficacy Assessment

We further explored the application of the model in neurodegenerative diseases. Figure 7 compares the prediction results between Alzheimer’s disease (AD) samples and cognitively normal (CN) controls, where orange and gray denote AD and CN samples, respectively. For subjects aged 50–70, many AD samples exhibited “advanced brain age,” with predicted ages substantially exceeding chronological ages and falling outside the variation range of the CN group. In contrast, for subjects aged 80–90, a trend of “delayed brain age” was observed. According to the Limited Lifespan Hypothesis [33], the average human lifespan is approximately 85 years in the absence of biological aging interventions. Therefore, we speculate that individuals surviving to advanced ages tend to have healthier brain conditions, which may explain the manifestation of delayed brain age in this group.

To explain why some AD samples still exhibited delayed brain age, we hypothesize that these individuals were already in a delayed state prior to disease onset. This was particularly evident among subjects over 80 years old, who likely had relatively healthier brain conditions before onset. Even after the accelerated aging associated with AD, it would take time before their brain age predictions shifted to an advanced state. Consequently, as shown in Figure 7, a small number of samples still appeared delayed; however, their predicted brain ages never fell below the minimum values observed in age-matched CN subjects, indicating that they nevertheless exhibited an abnormal aging trajectory. To compare brain age differences between AD and CN groups, we plotted PAD values as shown in Figure 8, where orange and gray represent AD and CN samples, respectively. The results show that, for subjects aged 50–80, most AD samples exhibited “advanced brain age.” Notably, this phenomenon was observed even though the model was not trained on AD data, thereby confirming the strong association between PAD and dementia. For subjects aged 80–90, AD samples did not show clear advancement; however, their trend of “delayed brain age” was weaker compared to CN subjects. We attribute this to the accelerated brain aging caused by AD [34,35,36].

4.6. Softmax-Derived Brain Age Mapping Analysis

We employed the SD-BAM algorithm to analyze regional aging patterns across various ages and physiological conditions. Figure 9 illustrates the BAMs of healthy subjects aged 20, 55, and 70 years. Observations reveal that the model primarily focuses on the ventricular regions. In the 20-year-old sample, the markers appear in deep blue, signifying prominent youthful features. In contrast, the 55-year-old brain exhibits partial orange-yellow markers, indicating early signs of senescence. By age 70, the markers have almost entirely transitioned to red. This progression from blue to red precisely reflects the normal developmental trajectory of the human brain with increasing age. We attribute these findings to the trend of ventricular enlargement as described in [14,37].

Next, we analyzed the Alzheimer’s Disease (AD) samples aged 55 and 70, as illustrated in Figure 10. In contrast to the healthy controls, the model not only identifies ventricular enlargement but also demonstrates significant attention toward cortical contours and the hippocampal regions. While normal aging is accompanied by gradual cortical and hippocampal atrophy, AD induces abnormally accelerated brain aging, leading to more severe atrophy of the cortex and hippocampus, as well as pronounced ventricular expansion. Consequently, the AD samples exhibit almost entirely red aging markers within the BAM [38].

Notably, the training dataset used in this study did not include any AD samples. These findings confirm that AD indeed leads to accelerated brain aging and demonstrate that brain age can serve as a reliable quantitative biomarker. Furthermore, through BAM, researchers can examine neurodegenerative progression across different brain regions in detail, providing an objective spatial basis for diagnosis and clinical assessment.

Figure 11, Figure 12 and Figure 13 present the ROI significance analysis histograms across different age groups (15–20, 55–60, and 70–75 years) and health states (CN, AD), evaluated under three pivot ages (40, 50, and 60) and various threshold settings (None, P50, and P99.5). The left subfigures (a,c,e) display positive weights, representing pixels that significantly contribute to aging features. Conversely, the right subfigures (b,d,f) show negative weights, corresponding to pixels associated with youthful features. For this analysis, the original brain regions were integrated into four major spatial zones: Anterior, Posterior, Lateral, and Central.

All CN subjects included in the ROI analysis were drawn exclusively from the test set. The training set was not used for ROI analysis to avoid potential data leakage and to ensure a fair and unbiased comparison across groups. In contrast, AD subjects were not included in the model training process and were therefore all incorporated into the ROI analysis. The corresponding sample sizes for each group are reported in

Table 12. Sample distribution for ROI analysis.

Age Group	CN (Images)	AD (Images)
15–20	34	-
55–65	27	15
70–75	56	116

.

In this study, P99.5 was selected as the fusion threshold primarily because the responses in the final saliency maps tend to be obscured by noise generated during the fusion process described in Equation (4). Observation of subfigures (a) through (d) reveals that synthesizing multiple saliency maps can lead to a loss of features; consequently, aging-related (positive) pixels are preserved only in the AD group, while youthful (negative) pixels show almost no response. In contrast, by applying a high-intensity mask at the 99.5th percentile, subfigure (e) successfully captures the physiological aging features of the CN group, and subfigure (f) clearly delineates brain regions associated with youthful characteristics. Without this thresholding, it would be impossible to effectively define the underlying physiological significance.

The setting of the pivot age can be adjusted according to the specific analytical goals. Given the wide age span of the dataset (6–90 years, as detailed in Table 3), we established 50 years as the primary pivot point to ensure a balanced distribution. To validate the impact of this parameter, we further tested the ROI performance under three different pivot ages: 40, 50, and 60. Experimental results demonstrate that the responses across all brain regions remained stable across these different pivot age settings, showing no significant fluctuations.

Figure 12e,f present the ROI analysis results using the parameters proposed in this study. From the analysis of aging-related pixels (Positive) in Figure 12e, it is evident that the AD group exhibits extremely high significance across all four regions (Anterior, Posterior, Lateral, and Central). In contrast, the responses in the CN group are primarily concentrated in the Anterior region. This phenomenon reveals the natural pattern of human brain senescence, where evolutionarily later-developing regions, such as the frontal lobe, tend to manifest aging characteristics first [39]. Conversely, in the youthful features (Negative) shown in Figure 12f, the significant response in the Anterior region serves as a metric for assessing brain youthfulness. This quantitative analysis confirms that SD-BAM can effectively differentiate between physiological aging and pathological degeneration, demonstrating its potential as an index for early dementia screening.

4.7. Limitations and Discussions

This study proposed a lightweight model based on VGGNet [23] and evaluated the effects of preprocessing procedures, model design, and data augmentation on predictive performance. Finally, we integrated our newly developed algorithm, Softmax-Derived Brain Age Mapping (SD-BAM), to enable the visualization and identification of key brain regions. The experimental results on the output adjustment layer showed that, compared with the regression strategy, the softmax strategy achieved better MAE performance. However, since the two strategies differ only in a small number of neurons, the overall difference in predictive accuracy was not substantial. From the perspective of model interpretability, the softmax strategy provides a key advantage, as its fixed weighting parameters elegantly assign physical meaning to the feature vectors. Regarding the output structure, this study employed 50 age classes with 2-year intervals. While this configuration yielded competitive results, we recognize that the granularity of age partitioning may influence model sensitivity and error distribution. Future investigations could explore different granularities, such as 1-year or 5-year intervals, to systematically evaluate their impact on predictive performance and clinical utility.

This study observed a clear improvement in predictive performance after data augmentation was removed. We speculate that this may be related to our preprocessing pipeline: MNI registration and skull stripping had already substantially reduced inter-image variations in position and intensity, such that additional augmentation instead introduced unwanted noise. While these results were consistent across our tests, we acknowledge that this conclusion is contingent upon our current methodology. Specifically, the augmentation parameters (3–5 voxel shifts and 3–5° rotations) might be too conservative to offer positive regularization. Moreover, the detrimental effect of Gamma correction may be attributed to its interaction with the intensity normalization process rather than the augmentation itself. PAD (Predicted Age Difference) analysis revealed a tendency of overestimation in younger subjects and underestimation in older subjects, indicating that sample size and distribution influence prediction accuracy. In particular, for individuals over 80 years of age, most samples exhibited delayed brain age. We attribute this to the fact that this age range approaches the human lifespan limit [33], such that those who survive beyond 80 years are more likely to maintain relatively healthy brain conditions. However, this observation should be interpreted with caution, as it may also reflect the uneven sample distribution of the elderly population and the regression-toward-the-mean effect commonly observed in brain age estimation. Consequently, the extent to which this collective delayed brain age phenomenon arises from insufficient training samples versus genuinely healthier brain states remains to be further investigated. Importantly, our analysis shows that even when AD samples appear to exhibit delayed brain age, they still demonstrate a trend of accelerated aging. These results suggest that PAD remains an effective indicator of abnormal brain aging and holds potential as an early screening tool for dementia risk.

The BAM analysis revealed an important finding: when estimating brain age, the model focused particularly on ventricle, cortical and hippocampal regions, which are closely associated with cognitive function. This not only aligns with conclusions from existing neuroscience literature [1,38,40] but also demonstrates the model’s ability to automatically capture core brain regions related to aging and cognitive decline. The SD-BAM algorithm provides visualization of both youthful and aging feature dimensions in brain age prediction models, thereby laying a solid foundation for clinical applications. Despite its effectiveness, the current implementation of SD-BAM is inherently coupled with a classification-based regression framework. This architectural dependency arises from the need to leverage discrete age-class probabilities to weight the feature maps. While this approach excels at capturing stage-specific anatomical markers, it may constrain the direct integration of SD-BAM into pure regression models or emerging backbone architectures that lack a softmax-based output structure. Recognizing this limitation, future research should focus on decoupling the saliency generation mechanism from the classification layer. Developing more generalized SD-BAM variants that can be seamlessly applied to continuous regression tasks and diverse neural network designs remains a priority for enhancing the framework’s versatility. However, the algorithm remains constrained by the classification-weighted output structure, limiting its applicability across all model architectures. Beyond architectural dependencies, as SD-BAM is built upon the Grad-CAM++ framework, it inherits a degree of sensitivity to gradients. This dependency implies that the stability of the generated saliency maps may be susceptible to issues such as gradient vanishing and noise amplification. Therefore, the interpretation of these results should be approached with caution, taking into account these inherent algorithmic limitations. Regarding spatial precision, the use of rigid-body registration in our preprocessing may limit the precision required for investigating subtle brain regions. This study also faces the challenge of high MRI heterogeneity, which could affect the robustness of the saliency regions. In light of these challenges, recent studies [41,42] have demonstrated that strategies for robust feature representation and geometric perturbation modeling can significantly enhance cross-scene consistency. Specifically, the domain transform strategies (e.g., PHFM/QPHFMs) proposed in these works provide stable mappings for complex structural features, which is particularly enlightening for enhancing the stability of SD-BAM under multi-center MRI conditions. Integrating such deformation-resistant feature preservation could mitigate the impact of data heterogeneity and ensure more reliable brain-age representations. Therefore, we believe that future research could incorporate non-linear registration techniques and advanced feature-stabilizing frameworks to enable more granular ROI analysis and mitigate the impact of data heterogeneity. An additional point for consideration is the quantitative evaluation of spatial stability. While our results show consistent neuroanatomical patterns, the reproducibility of these saliency maps across different training folds and random initializations was not formally quantified. Future research should incorporate metrics such as the Dice coefficient to rigorously assess the consistency of these identified biomarkers. Consequently, we believe that future investigations could benefit from non-linear registration techniques and advanced feature-stabilizing frameworks to enable more granular ROI analysis.

Fortunately, our comparison of regression and softmax output strategies showed that, under appropriate dropout rates, both approaches achieved comparable performance. This suggests that models originally designed with regression outputs can be safely converted to the softmax framework without compromising accuracy. In the CamCAN dataset, achieving significant improvements in brain age prediction accuracy has long been challenging. Ref. [16] demonstrated that combining MRI and MEG modalities could enhance model performance, achieving an MAE of 4.88 years. However, our experiments showed that by incorporating external datasets into training, comparable performance (MAE = 4.90 years) could be achieved using MRI alone. This finding indicates that expanding the training dataset is an effective strategy for improving prediction accuracy on specific datasets. Moreover, this approach eliminates the need for additional MEG acquisition, thereby reducing clinical costs while maintaining competitive predictive performance.

We believe that SD-BAM holds significant potential across various advanced applications. It can serve as a critical reference for generative AI in brain synthesis tasks; specifically, these saliency maps can guide generative models to accurately reproduce aging or rejuvenating features in key regions when synthesizing brain images of specific ages or pathological states. Furthermore, SD-BAM has the potential to function as an early biomarker for neurodegenerative diseases, providing essential markers for early lesion detection and dementia risk assessment.

5. Conclusions

This study conducted a comprehensive investigation of brain age prediction models. Experimental results regarding the output adjustment layer indicate that under a regression-based strategy, the absence of Softmax modulation caused Dropout to degrade performance, with the optimal Mean Absolute Error (MAE) achieved at a Dropout rate of 0. Conversely, the Softmax-based strategy benefited positively from Dropout, reaching peak performance at a rate of 0.2. Under their respective optimal hyperparameter configurations, both strategies demonstrated comparable performance, yielding MAEs of 4.34 and 4.33 years, respectively, with the Softmax strategy showing a marginal advantage.

Regarding data augmentation, we observed that omitting augmentation operations actually improved the performance of most models. Notably, the performance of the whole-brain model advanced from an MAE of 4.33 to 4.08 years. This suggests that spatial augmentations, such as translation and rotation, may have compromised the spatial alignment achieved through MNI registration, leading to a degradation of the preprocessing effects. After removing these augmentations, our model achieved an MAE of 4.90 years on the CamCAN test set, outperforming most previous studies that relied solely on CamCAN MRI data.

Clinical application analysis reveals a prevalent ‘advanced brain age’ phenomenon in patients with dementia aged 50 to 80, validating the potential of the Predicted Age Difference (PAD) as a quantitative screening biomarker. Our proposed visualization technique, SD-BAM, effectively identifies both youthful and aging features simultaneously. Experimental evidence shows that when evaluating older brains, the model exhibits high attention toward the ventricles and cortex, accurately reflecting the trends of ventricular enlargement and cortical atrophy associated with senescence. In healthy subjects, the model precisely captures physiological transitions: younger brains are characterized by features in anterior regions (such as the frontal lobe), which are also the earliest areas to manifest signs of aging.

In conclusion, ROI quantitative analysis validates the fundamental patterns of human brain senescence, with ventricular changes persisting throughout the entire aging process. In patients with Alzheimer’s Disease, pathological degeneration causes cortical aging features to be significantly more pronounced than in age-matched healthy controls, effectively overshadowing localized youthful characteristics. SD-BAM not only enhances the depth of brain age assessment but also provides an objective tool for diagnosing neurodegenerative diseases based on spatial distribution evidence.