Deep learning models identify brain changes during the progression of Alzheimer’s disease

Abstract

Alzheimer’s disease (AD) is an irreversible neurodegenerative disorder whose progression is closely associated with time. However, most diagnostic models are based on single time-point data, overlooking longitudinal disease characteristics. Structural magnetic resonance imaging (sMRI) has been widely utilized in the study of AD. To address the need for multi-time series analysis in longitudinal AD research and the integration of features from different brain tissues, we propose a Multi-Branch Fusion Channel Attention Network (MBFCA-Net) for disease diagnosis. This network leverages the temporal correlations across longitudinal scans for effective AD detection. We further conduct retrospective interpretability analysis to quantify the contributions of brain regions across disease stages. This enables a detailed investigation of dynamic changes in brain regions associated with AD and normal aging. The results indicate that the importance of regions such as the amygdala, parahippocampal gyrus, and temporal lobe undergoes dynamic changes throughout the progression of AD. Furthermore, AD-related voxel clusters exhibit a developmental trend, shifting from the hippocampus to the temporal lobe and transitioning from a dispersed to a more aggregated distribution. Our study provides novel insights into the longitudinal patterns of AD-related changes, offering valuable contributions to early diagnosis and pathological understanding of the disease.

Introduction

Alzheimer’s disease (AD) is the most common form of dementia. It is a covert, progressive neurodegenerative disorder primarily affecting cognitive and memory functions of the brain, and may ultimately lead to death¹. Mild cognitive impairment (MCI) is one of the most significant conditions impacting the health of elderly individuals. It progresses slowly and irreversibly and represents an early stage of AD. Without timely intervention and systematic treatment, MCI patients face a high risk of progression to AD, which significantly impairs their quality of life and imposes a substantial burden on society². Despite extensive efforts, the pathogenesis of AD remains unclear due to the complex interrelations among its etiological factors³. AD not only places enormous financial pressure on healthcare systems but also causes considerable psychological and emotional distress to patients and their families. Currently, there is no treatment capable of reversing AD progression; therefore, early diagnosis and intervention are crucial for slowing disease advancement⁴.

Today, several established biomarkers have enhanced understanding of AD progression⁵. With the aid of brain imaging devices and techniques, brain abnormalities in AD and MCI patients, such as gray matter (GM) atrophy, can be clearly visualized and observed via neuroimaging^6,7. Investigating the pathogenesis of AD facilitates early diagnosis and aids in timely detection and treatment^8,9. MRI provides high-resolution images of brain structures, effectively capturing anatomical and functional neurological changes associated with AD¹⁰. Over the past few decades, sMRI has been extensively applied in AD detection¹¹. Numerous studies utilizing deep learning techniques have been conducted with MRI as the foundational imaging modality. For instance, Sathish Kumar et al.¹² conducted experiments with 2D slice data using classic models, such as AlexNet, VGG16, GoogLeNet, and ResNet18 to compare their performance. Salami et al.¹³ adapted models, such as ResNet, Inception-v3, and DenseNet into 3D versions to fully utilize the spatial information in 3D MRI and compared these methods with their proposed 3D convolutional model. In addition to commonly used CNN architectures, other models, such as RNN, Transformer, and graph convolution networks have also been applied in current research. For example, Xin et al.¹⁴ proposed an efficient model combining CNN and Swin-Transformer, employing lightweight techniques to minimize model parameters and computational costs. Li et al.¹⁵ combined 3D CNN with LSTM to extract four-dimensional feature from fMRI, encompassing both spatial and temporal dimensions. Zhang et al.¹⁶ utilized a graph convolutional network to develop a novel Multi-Relational Reasoning Network (MRN), which learns multi-relational perceptual representations of brain regions from sMRI data for AD diagnosis, integrating spatial correlations and topological information.

Nevertheless, the majority of current studies rely solely on single time-point data for model construction, disregarding the temporal sequence characteristics inherent to AD as a progressive disease and ignoring the temporal dependencies present in longitudinal data¹⁷. Although some studies have employed multiple time-point data using traditional classification methods, such as support vector machines, least absolute shrinkage and selection operator (LASSO), and random forests, these approaches typically concatenate features from multiple time points for unified classification or classify each time point separately before integrating the results¹⁸. For example, Alinsaif et al.¹⁹ concatenated 3D spatial-temporal and convolutional neural network (CNN) feature vectors for AD classification, while Kang et al.²⁰ combined three classifiers to build an ensemble model. These methods do not consider intrinsic correlations across different time points; consequently, their performance in diagnosing brain diseases remains suboptimal²¹.

To overcome these limitations and effectively integrate multi-time-point data, various data fusion approaches have recently been developed in cognitive and clinical research. For instance, multi-task learning not only explores latent relationships among different time points but also improves model generalizability²². Jie et al.²³ proposed a group sparse learning method with temporal constraints for longitudinal data analysis, which captures smooth changes between adjacent time points. Cui et al.²⁴ conducted longitudinal AD diagnostic analysis based on recurrent neural networks. However, these methods only consider latent relationships within longitudinal data, overlooking the unique characteristics of brain images themselves—unlike natural images, normal brain function depends on the coordinated interaction of various brain regions, and there often exist latent correlations between different voxels. Based on the topological correlations arising from latent brain network patterns among voxels in brain images, a multi-branch fusion attention module based on voxel correlations was proposed, alongside a longitudinal 3D convolutional diagnostic model incorporating a three-dimensional spatial coordinate attention mechanism to enhance feature extraction capabilities. Through visual analysis, we quantified the contributions of brain regions at different stages to the model’s predictions and identified regions strongly associated with the longitudinal progression of AD and normal aging. In addition, we further investigated the associations among these regions, cognitive assessment scores, and temporal dynamics.

Results

Performance Evaluation and Comparative Analysis

As shown in Table 1, this study trained several AD classification models using different types of input data. The models trained on the three tissue types (GM, WM, and CSF) at a single time point exhibited consistent performance across different stages (m0, m12, and m24). This consistency can be attributed to the use of transfer learning during training, which enabled the model to share feature representations across different time points and maintain stable performance. Notably, the model trained on longitudinal GM data achieved the best performance.

Table 1.

Comparison of classification results of different task models

Model	Data	Accuracy	Sensitivity	Specificity
Model_m0	GM、WM、CSF (m0)	89.66%	81.82%	94.44%
Model_m12	GM、WM、CSF (m12)	89.66%	81.82%	94.44%
Model_m24	GM、WM、CSF (m24)	89.66%	81.82%	94.44%
Model_wm	WM (m0 + m12 + m24)	82.76%	63.64%	94.44%
Model_gm	GM (m0 + m12 + m24)	93.1%	81.82%	100%
Model_csf	CSF (m0 + m12 + m24)	86.21%	72.73%	94.44%

To evaluate the effectiveness of the proposed method, a comparative analysis was conducted between the proposed model and several state-of-the-art approaches from related studies. As illustrated in Table 2, the proposed model outperformed the others across all evaluation metrics. Furthermore, to evaluate the generalization ability of the model, we applied the trained model to the AIBL dataset for external validation, achieving an accuracy of 92%. The results are presented in Supplementary Table S1.

Table 2.

Comparison of classification performance with existing research results

References	Data	Method	Accuracy	Sensitivity	Specificity
Gonuguntla et al.49.	65AD + 65CN	ROIs	85.3%	-	-
Kang et al.20.	187AD + 229CN	DCGAN	84.23%	-	-
		Vgg16	83.57%	-	-
		ResNet50	74.7%	-	-
Kushol et al.50.	159AD + 229CN	Transformer	88.2%	95.6%	77.4%
Cai et al.51.	49AD + 43CN	Graph Transformer	85.9%	79.8%	92.8%
Lu et al.52.	193AD + 345CN	Vision Transformer	91.77%	84.38%	-
Qiu et al.53.	188AD + 229CN	FCN	83.4%	76.7%	88.9%
Shahamat et al.54.	70AD + 70CN	3D CNN	85%	-	-
Shojaei et al.55.	74AD + 71CN	3D CNN	87%	-	-
Tinauer et al.56.	128AD + 290CN	3D CNN	86.19%	79.73%	92.66%
ours	110AD + 176CN	3D CNN	93.1%	81.82%	100%

To assess the stability of the model, ten-fold cross-validation was performed using the ADNI dataset, and the results are presented in Table 3. The experiments demonstrated that the model trained on longitudinal GM data exhibited stable and superior classification performance. Moreover, in these cross-validation experiments, models trained on the three tissue types (GM, WM, and CSF) at a single time point showed gradually improving performance over time. This trend indicates that as the disease progresses, pathological differences become more pronounced, enabling the model to extract increasingly effective classification features. This trend is consistent with the natural progression of the disease and further confirms that the proposed model effectively captures longitudinal temporal features, validating its structural stability and applicability.

Table 3.

Comparison of Ten fold Cross Validation Training Results

Model	Data	Accuracy	Sensitivity	Specificity
KF_m0	GM、WM、CSF (m0)	91.27%	85.46%	94.87%
KF_m12	GM、WM、CSF (m12)	92.99%	88.18%	95.98%
KF_m24	GM、WM、CSF (m24)	94.04%	91.82%	95.46%
KF_gm	GM (m0 + m12 + m24)	93.35%	90.91%	94.91%

In addition, ablation experiments were conducted on each module of the model in this study. The results showed that the performance improvement of the model was not significant when using any single module alone, but it significantly improved after integrating all modules. Detailed results are shown in Supplementary Table S2.

Visualization Analysis and Pathological Interpretation

To trace the decision-making process of the model and to analyze the key brain regions it focuses on at different time points, this study employed SHAP for visualization analysis of the model trained on GM data. First, the contribution of each brain region was computed at three separate time points, followed by the calculation of the average contribution across these time points for each region. Based on the contribution scores, the top 15 brain regions were identified for each individual time point, as well as the top 15 regions based on average contribution, and subsequently visualized.

As illustrated in Fig. 1, the visualization reveals the model’s focus on GM regions over time. Overall, the parahippocampal gyrus, amygdala, hippocampus, temporal lobe, and cerebellum emerged as the primary regions of interest. These findings align closely with previous studies on the pathological mechanisms of AD^6,25–31, further validating the model’s effectiveness in feature extraction and reinforcing its biological interpretability.

Fig. 1. Global feature maps from different periods.

a–c Show the global feature maps for periods m0, m12, and m24, respectively. d Shows the feature map representing the average contribution over the three periods.

To elucidate the progression of brain structural changes in AD over time, this study further analyzes the variation in brain region contributions across different time points. At time point m0 (baseline), the amygdala exhibited a significantly higher contribution compared to other regions, indicating that it serves as a key distinguishing feature between AD patients and cognitively normal (CN) aging individuals during the early stages (relative to m12 and m24). Additionally, the hippocampus, temporal lobe, and parahippocampal gyrus also ranked among the top contributors, suggesting these areas were focal points for the model at this stage.

At time point m12 (12 months later), the amygdala remained the most influential region in the model’s decision-making process; however, its dominance in contribution decreased relative to other brain regions, such as the cerebellum and posterior temporal lobe. This suggests that as the disease progresses, pathological changes in additional brain regions begin to emerge, increasing their significance in the model’s predictions. In particular, the posterior temporal lobe showed a rapid increase in influence, reflecting intensified structural alterations in this region, and marking it as a critical feature for distinguishing AD from normal aging at this stage.

By time point m24 (24 months later), contribution values across all brain regions had increased, indicating that structural differences between AD patients and CN individuals became more pronounced over time. The parahippocampal gyrus, hippocampus, amygdala, and posterior temporal lobe were among the most significantly affected regions. Compared to m0 and m12, the amygdala’s contribution further declined, while the posterior temporal lobe continued to increase in importance. The parahippocampal gyrus emerged as the most contributive region at this stage, highlighting its increasingly critical role in AD progression.

To further explore the temporal dynamics of brain region contributions, this study computed the absolute differences in contribution scores between each pair of the three time points for all brain regions. Based on these differences, regions were ranked, and the top 15 brain areas showing the most substantial variation from m0 to m24 were selected. As illustrated in Fig. 2c, these regions and their contribution trends over time are visualized. The results indicate that the parahippocampal gyrus exhibited the greatest change in contribution, underscoring it as one of the most affected regions in the progression of AD. The structural divergence between patients and controls in this area increased over time, reinforcing its essential role in disease progression and supporting its prioritization as a key monitoring region. Moreover, the hippocampus, posterior temporal lobe, cerebellum, and amygdala also showed significant changes in their contributions, further validating their pathological relevance in the development of AD.

Fig. 2. Visual analysis of the importance of brain regions in classification tasks at different stages.

a Proportion of brain tissue contribution at different times. b The Venn diagram of brain regions ranked high in terms of average contribution and contribution change amplitude. c The brain region with the largest change in contribution over time.

To gain a deeper understanding of how the influence of different brain regions on the model’s decisions evolves across time, this study calculated the relative contribution proportions of all regions at the three time points and visualized the results (Fig. 2a). The analysis revealed that at m0 and m12, the top contributing brain regions each accounted for approximately 20% of the model’s focus. In contrast, by m24, the overall contribution from these regions had increased significantly, marking this as a critical period influencing model decisions. This trend suggests that as AD progresses, the structural differences between patients and CN individuals become increasingly prominent, aligning with known biological patterns of disease progression. Additionally, the trend may not follow a linear trajectory; instead, it may exhibit abrupt changes at specific stages, indicating a potential acceleration in the disease’s deterioration process.

Finally, based on the rankings of both average contribution across the three time points and the magnitude of contribution change, 18 key brain regions were identified as priority areas for further analysis. These regions, as shown in Fig. 2b, offer valuable insights for early detection and disease monitoring in AD.

Based on the contribution rankings shown in Fig. 1, the study found that at each time point, the key brain regions in the left cerebral hemisphere consistently exhibited higher contributions than those in the right hemisphere—a trend that became particularly pronounced at time point m24. This result suggests that AD-related pathological changes observed in sMRI may develop in an asymmetric manner, with more prominent degeneration occurring in the left hemisphere. Similar conclusions have been supported by several studies^32–36.

A more detailed examination of the contribution rankings at different time points in Fig. 1 further confirmed that the key brain regions in the left hemisphere demonstrated greater contributions than those in the right hemisphere across all stages, with the asymmetry most apparent at m24. This result supports the notion that AD-related structural changes in sMRI may follow an asymmetric pattern, in which the left hemisphere is more severely affected than the right. This observation aligns with previous research findings and reinforces the hypothesis that AD involves asymmetric patterns of neurodegeneration. Such asymmetry may be linked to the left hemisphere’s dominant role in higher cognitive functions, such as language and memory. Future studies could further investigate the underlying mechanisms of hemisphere-specific pathological changes in AD.

Statistical Analysis Results

To validate the reliability of the visualization results presented above, a systematic analysis of GM volume in key brain regions at three time points was conducted. First, intergroup comparisons were performed to evaluate significant differences in GM volume between the AD and CN groups at each time point. Raincloud plots were used to visualize the distributions (as shown in Fig. S1). Overall, significant group differences were observed at all time points, and these differences progressively increased over time. In particular, regions, such as the hippocampus, amygdala, parahippocampal gyrus, and temporal lobe exhibited markedly expanding intergroup differences. Although the cerebellum did not reach statistical significance at certain time points, a longitudinal trend of increasing divergence in GM volume between the AD and CN groups was still evident. These findings indicate that the brain regions identified by the model play an important role in the progression of AD, further validating their utility in distinguishing AD patients from healthy controls.

To further investigate the roles of these regions in AD pathology, GM volume data from all samples with available MMSE or CDR scores at all three time points were analyzed. Correlations between GM volume and age, as well as cognitive assessment scores, were calculated. The results of the correlation between volume and MMSE scores are presented in Fig. S2. All brain regions showed significant correlations with MMSE scores, indicating that structural changes in these regions are closely related to cognitive function and the progression of AD. Specifically, correlation coefficients were greater than 0.5 for the bilateral hippocampus, bilateral amygdala, bilateral parahippocampal gyrus, and bilateral posterior temporal lobes. Correlation coefficients ranged from 0.3 to 0.5 for the left anterior medial temporal lobe, left posterior insular gyrus, bilateral posterior cingulate gyrus, left middle frontal gyrus, and bilateral superior parietal lobules. The bilateral cerebellum showed the weakest correlations, with coefficients below 0.2. Given that MMSE scores reflect cognitive ability and are strongly associated with AD pathology, these results provide an important assessment of the relevance of these regions in AD development. Regions with correlation coefficients above 0.5 should be prioritized in AD diagnosis and are consistent with findings from earlier sections of this study, thereby enhancing the reliability of the results and supporting the model’s validity. Moreover, regions with coefficients between 0.3 and 0.5 include several that have not been highlighted in prior studies. Although their correlation with cognitive ability is not as strong as regions like the hippocampus, they still demonstrate substantial associations. These findings illustrate the model’s capacity to identify regions strongly linked to cognitive decline through longitudinal changes. As for the cerebellum, despite its weak correlation, its potential role in higher cognitive functions warrants further attention in disease monitoring and diagnostic research.

Further analysis was conducted to examine the correlation between GM volume and CDR scores, as shown in Fig. S3. Overall, GM volume in all regions showed significant negative correlations with CDR scores, consistent with prior clinical findings indicating that as cognitive impairment worsens (i.e., as CDR scores increase), GM volume tends to decline. The patterns of correlation across brain regions largely mirrored those observed with MMSE scores, reinforcing the robustness of the study’s results. These findings indicate that the proposed model not only accurately identifies brain regions associated with AD progression but also produces consistent and reproducible results, providing a solid foundation for future imaging-based AD diagnostic research.

To explore the longitudinal relationship between age and GM volume in these regions, data were preprocessed by using values at time point m0 as the baseline. Time increments and GM volume changes at m12 and m24 relative to m0 were calculated, and correlations between these variables were analyzed. The results are shown in Fig. S4.

Overall, both the AD and CN groups exhibited negative correlations between GM volume and time across the key regions, indicating that GM volume tends to decrease over time. Importantly, the correlation coefficients in the AD group were consistently higher than those in the CN group across all regions, suggesting that AD is associated with a more pronounced reduction in GM volume. This supports the notion that, compared to normal aging, AD accelerates the atrophy of GM in these regions due to pathological changes. For the CN group, all correlation coefficients—except for those of the bilateral hippocampus—were below 0.3, indicating relatively weak associations and suggesting a slower, more gradual decline in GM volume during normal aging. In contrast, for the AD group, correlation coefficients were above 0.5 for the bilateral hippocampus, bilateral amygdala, left anterior medial temporal lobe, bilateral parahippocampal gyrus, right posterior cingulate gyrus, and bilateral posterior temporal lobes. These findings demonstrate a strong association between these regions and AD-related pathological progression. Notably, the difference in correlation for the left parahippocampal gyrus between AD and CN groups was particularly striking. This result highlights the significant impact of AD on this region and aligns with the trend observed in Fig. 2c regarding its changing contribution over time, further validating the critical role of the parahippocampal gyrus in AD development.

VBM Analysis Results

Voxel-based morphometry (VBM) intergroup analyses were conducted on sMRI data from three time points: m0, m12, and m24. The results are summarized in Tables 4–6.

Table 5.

VBM analysis results during the m12 period

Index	Voxels	Peak coordinates(MNI)			P-value	Regions(harmmersimth atlas)
		x	y	z
1	170167	-27	-10	-16	1.3e-18	TL posterior temporal lobe R、 TL posterior temporal lobe L
2	3677	-30	20	48	3.7e-08	FL middle frontal gyrus L、 FL superior frontal gyrus L
3	1202	-44	10	32	1.3e-07	FL middle frontal gyrus L、 FL precentral gyrus L
4	326	30	-6	54	1e-06	FL superior frontal gyrus R、 FL middle frontal gyrus R、 FL precentral gyrus R

Table 4.

VBM analysis results during the m0 period

Index	Voxels	Peak coordinates(MNI)			P-value	Regions(harmmersimth atlas)
		x	y	z
1	89573	-27	-9	-16	5e-17	TL posterior temporal lobe R、 TL posterior temporal lobe L
2	6353	-2	-38	39	1.1e-09	CG posterior cingulate gyrus L、 CG posterior cingulate gyrus R、 PL superior parietal gyrus L、 PL superior parietal gyrus R
3	277	-58	-3	3	2.6e-07	TL superior temporal gyrus middle part R、FL precentral gyrus L、TL superior temporal gyrus anterior part L、PL postcentral gyrus L
4	507	27	58	3	6.4e-07	FL middle frontal gyrus R
5	606	-28	28	42	6.8e-07	FL middle frontal gyrus L、 FL superior frontal gyrus L
6	238	34	40	14	9.7e-07	FL anterior orbital gyrus R、 FL lateral orbital gyrus R、 FL posterior orbital gyrus R
7	176	-39	15	20	1.1e-06	FL middle frontal gyrus L、 FL inferior frontal gyrus L
8	124	12	-66	32	1.6e-06	PL superior parietal gyrus R
9	207	-32	54	8	2.1e-06	FL middle frontal gyrus L
10	115	-9	46	16	4.8e-06	FL superior frontal gyrus L、 CG anterior cingulate gyrus L

Table 6.

VBM analysis results during the m24 period

Index	Voxels	Peak coordinates(MNI)			P-value	Regions(harmmersimth atlas)
		x	y	z
1	171509	-28	-10	-18	1.4e-18	TL posterior temporal lobe R、 TL posterior temporal lobe L
2	6765	-30	20	48	4.3e-09	FL middle frontal gyrus L、 FL superior frontal gyrus L
3	362	30	-6	54	1e-06	FL superior frontal gyrus R、 FL middle frontal gyrus R、 FL precentral gyrus R

At time point m0, a larger number of regions showed significant differences between AD patients and CN controls. However, the voxel count per cluster was relatively small, with affected regions primarily located in the temporal, parietal, and frontal lobes. These findings suggest that even in the early stages of AD (relative to m12 and m24), structural alterations have already emerged in these brain areas. Nonetheless, the pathological changes at this stage remain in their initial phase—although numerous and spatially dispersed, the clusters are relatively small in volume and have not yet developed into extensive structural damage.

At time point m12, the number of significantly different regions decreased, but the volume of the clusters increased markedly, and the distribution of differences became more concentrated. The most affected regions were the posterior temporal lobe, superior frontal gyrus, and prefrontal cortex. This pattern indicates that, as the disease progresses, AD-related structural abnormalities begin to coalesce, leading to more extensive damage within specific brain areas.

By time point m24, the peak coordinates of intergroup differences remained relatively stable, but the significant clusters became even more concentrated, with signs of spatial merging and diffusion of pathological regions. The most prominently affected areas continued to be the posterior temporal lobe, superior frontal gyrus, and prefrontal cortex. These findings suggest that, in the later stages of disease progression, AD-related structural changes become more localized, with pathological damage increasingly confined to specific brain regions and exhibiting a trend toward higher spatial concentration.

As shown in Fig. 3, this study conducted a visual analysis of the intergroup differences between AD patients and normally aging individuals at three time points: m0, m12, and m24. Figure 3d–f clearly illustrate the temporal evolution of these differences. At the m0 stage, the intergroup differences are primarily concentrated in the hippocampal region, while differences in the temporal lobe are relatively scattered (Fig. 3d). As time progresses, the differences in the temporal lobe become increasingly pronounced (Fig. 3e), with a notable intensification observed in the posterior temporal region at the m24 stage (Fig. 3f). The posterior temporal lobe includes Wernicke’s area, a key center responsible for visual language processing in the brain. These findings suggest that language impairment in AD patients often emerges in the middle to late stages of the disease. Additionally, the distinctiveness of this region between groups indicates that language function does not decline significantly during normal aging, thereby highlighting language impairment as a potentially important indicator for distinguishing AD from normal aging.

Fig. 3. Visualization of VBM analysis.

a–c Visual comparison of axial slices between VBM groups. d–f Visual comparison of differences in brain regions between VBM groups. g–i Comparison of T-value mapping between VBM groups.

The aforementioned trend is also reflected in the slice-level visualization results in Fig. 3a–c and the analytical findings presented in Fig. 3g–i. Moreover, Fig. 3g–i reveals a distinct difference between AD patients and normally aging individuals in a specific region of the cerebellum. Although there remains debate regarding whether AD affects the cerebellum, the findings of this study suggest that any cerebellar involvement in AD may be confined to localized subregions. Therefore, future research should consider more fine-grained parcellation analyses of the cerebellum to further explore its potential role in the progression of AD.

Discussion

This study proposes a 3D convolutional neural network model for AD diagnosis, built on longitudinal sMRI data collected at multiple time points. Considering the spatial characteristics of brain imaging data, we introduce an attention mechanism based on voxel-wise correlation, which integrates spatial and channel information to enhance the model’s multi-branch feature fusion capability. Furthermore, a Coord-3D module was incorporated to improve the model’s representation of the fused feature maps. Experimental results demonstrate that the proposed model achieves superior classification performance for AD diagnosis. And through interpretability analysis, revealed key brain regions whose contributions evolved dynamically over time, providing a new quantitative perspective for understanding the structural progression of AD.

The interpretability analysis identified 18 brain regions that consistently contributed to AD diagnosis and revealed a critical dynamic pattern: the amygdala played prominent roles during the early stages of AD but showed decreasing importance as the disease progressed, whereas the parahippocampal gyrus and posterior temporal cortex exhibited steadily increasing contributions. This temporal pattern detected by the deep learning model was strongly supported by traditional neuroimaging analyses. VBM results showed that the structural differences between AD and CN groups evolved from small and scattered clusters at baseline (m0) to significantly converged regions in the posterior temporal and association cortices at m24 period. Furthermore, statistical validation confirmed that the GM volumes of these key regions were not only highly correlated with clinical cognitive scores (MMSE/CDR) but also exhibited a significantly faster atrophy rate in the AD group compared to normal aging. The convergence of these three lines of evidence indicates that the temporal features captured by the model reflect clinically relevant and accelerated neurodegenerative changes along the AD trajectory.

The high consistency among these findings enables a deeper mechanistic interpretation within the established neuropathological framework of AD. The observed progression pattern—from medial temporal to posterior cortical regions—aligns well with the Braak staging hypothesis describing the spatiotemporal spread of neurofibrillary tangles (NFTs) along specific neural pathways^37,38. The early dominance and subsequent decline of amygdalar contribution may correspond to the vulnerability of the amygdala and entorhinal cortex to tau pathology during Braak stages I-II, leading to functional impairment and serving as a structural substrate for early memory deficits. In contrast, the increasing influence of the parahippocampal and posterior temporal regions likely reflects disease propagation into the posterior associative cortex at later Braak stages (III-IV and beyond). Notably, VBM revealed significant involvement of the posterior temporal cortex, including Wernicke’s area, providing a neuroanatomical explanation for the language decline frequently observed in middle-to-late stages of AD^39–41.

Another important observation is the left-hemispheric asymmetry consistently identified by the model, which adds biological plausibility to the findings. Although AD has traditionally been considered a symmetric disease, accumulating neuroimaging evidence suggests that the left hemisphere—particularly the language-dominant network—may be more susceptible to neurodegeneration^32,34,35,42. Our results are consistent with this view, suggesting that the asymmetric progression of AD may reflect the higher metabolic demand or distinct connectivity patterns of the left hemisphere.

In summary, this study integrates longitudinal imaging analysis with deep learning techniques to characterize the dynamic structural changes in the AD brain. The deep learning model, VBM morphological analysis, and statistical validation collectively delineate a coherent spatiotemporal trajectory of AD-related structural degeneration: beginning in the medial temporal lobe, extending to posterior associative cortices, and exhibiting a left-hemispheric bias. These findings deepen the understanding of the relationship between AD and core brain regions, uncover the temporal evolution patterns of these regions, and extend previous research by highlighting the involvement of additional key areas, such as the bilateral posterior cingulate gyrus. This integrative multi-evidence framework provides deeper insights into the mechanisms underlying AD and normal aging, offering novel implications for early diagnosis and intervention.

Despite successfully identifying AD-related regions through longitudinal sMRI analysis and verifying them using deep learning and statistical methods, this study has some limitations. First, interpretability analysis was only conducted for the longitudinal GM model, which demonstrated the best performance. The evolution of other tissue types (e.g., white matter and cerebrospinal fluid) across time points was not comprehensively explored. Second, due to the difficulty of acquiring multi-time-point data, only sMRI data were utilized, without incorporating other imaging modalities, such as fMRI or DTI. This limits the scope of interpretation regarding AD from a functional or connectivity perspective. Finally, due to the limited pathological information in the current dataset, it remains challenging to elucidate the underlying disease mechanisms or to validate the observed regional brain alterations from a pathological perspective. The next step of work will involve more comprehensive datasets to enable deeper investigation and validation.

Future research may explore multi-modal imaging fusion approaches to comprehensively uncover longitudinal changes in AD at structural, functional, and network levels. Moreover, more advanced explainable AI techniques—such as Grad-CAM or SHAP-based feature importance analyses—could be employed to further elucidate the model’s decision-making mechanisms at different stages, enhancing its clinical interpretability and utility. In addition, integrating more pathological data and interpreting the model from a pathological perspective will enable a more comprehensive and in-depth exploration of the mechanisms underlying disease development and progression.

Methods

Data collection and preprocessing

In this study, the sMRI dataset used for training and internal validation was obtained from the AD Neuroimaging Initiative⁴³ (ADNI) database. ADNI is publicly accessible, developed by researchers, and ensures that informed consent was obtained from all participants during data collection. All procedures strictly adhered to relevant ethical guidelines and regulations.

To ensure comparability of results, inclusion and exclusion criteria were adapted based on the baseline recruitment protocol of the ADNI study⁴⁴. Specifically, participants were included if they were aged 55 years or older and had undergone 1.5 T T1-weighted MRI scanning within ±6 months of a clinical diagnosis of either AD or CN status. Exclusion criteria included individuals with mixed dementia involving AD, non-AD dementias, a history of severe traumatic brain injury, major depressive disorder, stroke, brain tumors, or significant systemic diseases. A total of 286 samples were selected from the ADNI dataset, comprising 110 AD patients and 176 CN individuals. Each subject contributed data from three time points, with approximately 12-month intervals between consecutive scans. No subjects had missing sMRI data at any time point. A few samples (≤5) lacked MMSE or CDR scores at m12 or m24, and these were excluded from the corresponding analyses. Detailed sample distribution information is provided in Supplementary Table S3. To avoid significant class imbalance during random dataset partitioning, both AD and CN samples were randomly divided into training, validation, and test sets in an 8:1:1 ratio, ensuring that class distributions in each subset were consistent with those of the original dataset. To prevent data leakage and ensure the reliability of the results, all dataset splits were performed at the subject level.

All preprocessing steps in this study were performed using the FSL⁴⁵ software suite. First, image registration was conducted using the FLIRT command, aligning each image to the MNI152 template with a resolution of 1 mm, resulting in images with dimensions of 182 × 218 × 182 mm. Subsequently, skull stripping was carried out using the BET command on the registered images. The skull-stripped images were then segmented using the FAST command to generate probability maps and segmentation results for GM, white matter (WM), and cerebrospinal fluid (CSF). During this process, all registered images were carefully reviewed, and the outcomes of automatic registration were confirmed to meet expectations.

In addition, subjects with imaging data available at all three time points were selected from the Australian Imaging and Lifestyle flagship (AIBL) database as an external validation dataset. Detailed sample distribution information is provided in Supplementary Table S4.

Pipeline of the designed method

This study focuses on the longitudinal analysis of AD and normal aging, aiming to develop a diagnostic model using sMRI data collected at multiple time points. By integrating visualization and data analysis techniques, the study retrospectively interprets the model outputs to identify key regions of structural brain changes associated with disease progression and normal aging. This approach seeks to provide both theoretical foundations and technical support for early diagnosis and prevention of AD.

As illustrated in Fig. 4a, the overall experimental workflow involves preprocessing the sMRI data to extract brain tissue information—specifically GM, WM, and CSF—across different time points. Two data input strategies are employed to train the diagnostic model: one based on the same brain tissue across multiple time points, and the other based on different brain tissues at a single time point. Following model training, a series of visualization analyses are conducted to reveal the temporal dynamics of brain structural changes and their associations with either pathological conditions or normal aging.

Fig. 4. The experimental process and the structure of the model.

a Pipeline of the designed method. b Model Structure.

In addition, the study applies VBM and correlation analysis to further explore the extracted brain structure data. These analytical results are then compared with the model’s visualization outputs to validate the identified key brain regions and assess their potential value in disease diagnosis.

Through this comprehensive analytical framework, the study not only distinguishes the structural differences between AD and normal aging but also provides a robust scientific and technical foundation for future clinical applications.

Model structure design

This study proposes an intelligent diagnostic model for AD based on a hierarchical deep learning architecture, designed to effectively represent brain imaging data through a multi-stage feature processing mechanism. The overall architecture comprises a feature downsampling module (Down_sample), a multi-scale convolution module (Fusion_group), a cross-modal fusion module (Fusion_channel), a 3D coordinate attention module (Coord-3D), and a classification module. These components work synergistically to progressively extract and optimize pathological features. As showed in Fig. 4b.

The feature downsampling module employs cascaded 3D pooling and strided convolution operations to reduce the spatial dimensionality of the input data. The convolution kernel sizes (1×1, 3×3, and 5×5) were chosen based on the principle of multi-scale feature extraction, inspired by the Inception architecture⁴⁶. The effectiveness of such multi-scale convolution combinations has been validated in several previous neuroimaging studies. This not only decreases computational complexity but also preserves key anatomical structures. The multi-scale convolution module captures both fine-grained local features and larger-scale morphological changes by applying convolutional kernels of varying sizes in parallel, effectively addressing the scale sensitivity of pathological changes, such as brain atrophy.

During the feature fusion stage, the multi-branch fusion module integrates multi-source data features through a dynamic weighting mechanism. This module innovatively introduces a voxel-wise correlation weight matrix in combination with a channel attention mechanism to nonlinearly couple features from different brain tissues—such as GM and cerebrospinal fluid—within a unified spatial framework. This significantly enhances the model’s ability to exploit cross-modal complementary information. In terms of utilizing longitudinal data, unlike traditional feature concatenation approaches, the proposed longitudinal feature fusion introduces a two-level dynamic weighting mechanism. (1) Intra-timepoint weighting: a voxel-wise relational weight matrix is computed within each time point to preserve individual variability in local brain structures; (2) Inter-timepoint weighting: during temporal fusion, channel-wise features are first adaptively aggregated within each branch, and then inter-branch weights are learned to balance and exchange information across different time points.

To further strengthen the discriminative power of spatial features, the 3D coordinate attention module decomposes attention weights along the axial, sagittal, and coronal planes. Through this joint spatial modeling strategy, the module adaptively amplifies signal responses in lesion-associated regions while suppressing interference from irrelevant brain areas.

Finally, the classification module maps the high-level feature representations to clinical diagnostic labels via fully connected layers. Experimental results demonstrate that the proposed architecture—through its combination of multi-scale feature extraction, dynamic cross-modal fusion, and anatomy-driven attention enhancement—significantly improves the model’s ability to identify early-stage pathological features of AD.

Given that the original input data consists of three-dimensional medical images, a Down_sample module was designed and implemented to reduce feature dimensionality and accelerate model training. As illustrated in Fig. 5b, this module first applies both max pooling and average pooling to the input data, enabling the capture of prominent local features and global contextual information, respectively. The pooled feature maps are then passed through a convolutional layer with a 5×5 kernel, which performs spatial dimension compression. This step further reduces feature dimensionality while enriching channel-wise information, thereby preserving more semantic content.

Fig. 5. Main module design.

a Fusion_group module. b Down_sample module. c Fusion_channel module. d Coord-3D module.

In the feature fusion stage, the Fusion_group module is introduced to ensure comprehensive feature extraction from each input branch prior to fusion. This module employs convolutional kernels of varying sizes (e.g., 1×1, 3×3, and 5×5) to extract multi-scale features with diverse receptive fields. This design allows the model to capture both fine-grained details and broader contextual dependencies. Finally, a 3 × 3 convolution is applied to fuse the extracted multi-scale features, generating a more representative and discriminative feature output.

The Fusion_channel module is responsible for integrating feature information from multiple branches. In this study, two types of inputs are considered: (1) data from the same subject at different time points, and (2) segmented brain tissue data—CSF, GM, and WM—from the same subject at a single time point. Taking the second case as an example, each input branch undergoes a 1 × 1 convolution for initial channel-level feature fusion, resulting in a feature map of shape (B, 1, D, H, W). A softmax function is then used to compute a voxel-wise relational weight matrix, and the feature map is reshaped to (B, D×H×W, 1). The weight matrix is multiplied with the original feature map, applying voxel-wise relational weights across all channels within each branch. This operation yields a set of relative channel weights through weighted summation of all voxels in each channel. The channel weight information from multiple branches is concatenated and passed through a sigmoid function to generate inter-branch channel weights. This process ensures that, while retaining intra-branch channel characteristics, the model also learns the relative importance of different branches. Finally, during multi-branch fusion, the learned weights are applied to the original data, producing the final weighted feature map.

The Coord-3D module is a coordinate-aware attention mechanism that enhances the model’s feature extraction capability by incorporating both spatial and channel information. As shown in Fig. 5c, the module begins by decomposing the input feature map along the x, y, and z spatial axes. Taking the x-axis as an example, global average pooling is performed along this axis to extract two-dimensional spatial encoding information. This encoding captures both aggregated voxel information along the axis and relative spatial positioning. The pooled features are then reshaped into a tensor of shape (B, C, H×W). A similar process is applied to the y and z axes. The resulting feature vectors are concatenated along the channel dimension, forming a unified representation of shape (B, C, H×W + D×W + D×H). A 1×1 convolution is applied to compute channel-wise attention weights, followed by activation using a sigmoid function to generate the attention map. Finally, this attention map is reshaped and multiplied with the original input feature map, yielding a weighted feature representation that integrates both spatial positional and channel information.

Explainability method

In numerous research fields, understanding the reasons behind a model’s predictions is as crucial as the accuracy of those predictions. In this study, we opted to employ SHapley Additive exPlanations (SHAP)⁴⁷ for the interpretability analysis of our model.

SHAP constitutes a unified framework for interpreting machine learning models. It evaluates the contribution of each feature by calculating the marginal contribution of each feature across all possible prediction tasks and averaging these values. This process enables the assignment of a value to each feature, indicating its importance in a specific prediction. We can treat each voxel in an MRI as a unique feature and construct high-resolution heatmaps for the input MRI by assigning SHAP values to individual voxels on a voxel-by-voxel basis. For brain parcellation and visualization, we followed the approach of Qiu et al.⁴⁸, using the FNIRT command to perform nonlinear registration with the Hammersmith Brain Atlas. Upon manual review, the registration for all ADNI samples was satisfactory.

For the method of selecting highly correlated brain regions, the contribution of each of the 95 brain regions to the model’s prediction was quantified using SHAP values. To identify the most influential regions for AD classification, we implemented the following procedure at each time point (m0, m12, m24): First, all regions were ranked in descending order based on their mean absolute SHAP value. Second, we calculated the cumulative contribution of the regions from the top of the ranking. The key regions were defined as the smallest set of regions required for the cumulative contribution to exceed 60% of the total. Applying this threshold identified 15 key regions at each time point, which collectively accounted for 60–69% of the model’s decision, while the remaining 80 regions played a comparatively minor role. These regions were thus considered the primary factors associated with AD classification.

VBM analysis

VBM was employed to perform a two-sample t-test on whole-brain GM volume between the AD group and the CN control group. Data preprocessing was conducted using the SPM12 and CAT12 toolboxes. The specific procedures are as follows:

1. The original DICOM-format data were converted to NIfTI format and imported into the SPM software. All 3D T1-weighted images were spatially normalized to the standard space of the Montreal Neurological Institute (MNI).

2. The normalized images were segmented using the CAT12 toolbox to obtain GM, WM, and CSF maps, along with regional volume data for each tissue type.

3. The GM images were smoothed using an 8 mm full-width at half-maximum (FWHM) Gaussian kernel to enhance the signal-to-noise ratio.

4. Statistical modeling was then applied to the smoothed GM images. A two-sample t-test was conducted to compare the AD and CN groups, with total intracranial volume and age included as covariates. Multiple comparisons were corrected using the Family-Wise Error (FWE) correction method. Clusters with p < 0.05 and a minimum cluster size greater than 100 voxels were considered statistically significant.

Statistical Analysis

To validate the visualization results of the model and to demonstrate that the identified brain regions are closely associated with the progression of (AD, a systematic statistical analysis was conducted based on GM volume data obtained through VBM.

Specifically, to assess the importance of these regions in the longitudinal progression of AD, the Mann–Whitney U test was employed to evaluate differences in GM volume between the AD and CN groups at different time points. In addition, Pearson correlation coefficients were calculated to examine the relationships between GM volume in the identified regions and cognitive performance as measured by Mini-Mental State Examination (MMSE) and Clinical Dementia Rating (CDR) scores.

Furthermore, to explore the relationship between AD progression and aging in terms of temporal changes in GM volume, correlations between GM volume increments and time intervals were also computed.

Experimental Settings

All experiments in this study were conducted using the PyTorch framework on an NVIDIA GeForce RTX 3080Ti GPU. The model was trained using the Adam optimizer with an initial learning rate of 0.001 and a batch size of 4. The maximum number of training epochs was set to 300, with an early stopping strategy to prevent overfitting. Specifically, training was terminated if the validation accuracy did not improve for 30 consecutive epochs. To avoid premature stopping when the model was still learning, the training AUC was monitored: if the validation accuracy remained unchanged but the training AUC continued to increase, the early stopping counter was reset, allowing further training to preserve the best model.

Author contributions

W.C. and J.D.J.H. conceived the project. JS is responsible for data collection and processing. J.S., W.C., and J.D.J.H. conducted experimental design and data analysis. All authors participated in writing or revising the manuscript.

Data availability

The data in this study is from the ADNI database (https://adni.loni.usc.edu) and the AIBL database (https://aibl.csiro.au). Both datasets are publicly available. The relevant code is available from the authors upon reasonable request.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41540-026-00666-7.