Organizations: Faculty of Engineering and Natural Sciences (VPA Lab), Sabanci University, Istanbul, Türkiye. · Department of Computer Science, Boston University, Boston, MA, USA. · Department of Medicine, Boston University Chobanian & Avedisian School of Medicine; Department of Computer Science and Faculty of Computing and Data Sciences, Boston University, Boston, MA, USA.
Despite progress in deep learning for Alzheimer's disease (AD) diagnostics, models trained on structural magnetic resonance imaging (sMRI) often do not perform well when applied to new cohorts due to domain shifts from varying scanners, protocols and patient demographics. AD, the primary driver of dementia, manifests through progressive cognitive and neuroanatomical changes like atrophy and ventricular expansion, making robust, generalizable classification essential for real-world use. While convolutional neural networks and transformers have advanced feature extraction via attention and fusion techniques, single-domain generalization (SDG) remains underexplored yet critical, given the fragmented nature of AD datasets. To bridge this gap, we propose Extended MixStyle (EM), a higher-order extension of MixStyle that incorporates skewness and kurtosis into feature-statistic perturbation to mimic diverse distributional variations. Trained on sMRI data from the National Alzheimer's Coordinating Center (NACC; n=4,647) to differentiate participants with normal cognition (NC) from those with mild cognitive impairment (MCI) or AD and tested on three unseen cohorts (total n=3,126), EM yields enhanced cross-domain performance, with an average numerical gain of 2.4 percentage points in macro-F1 over the 3D U-Net baseline, underscoring its potential for improving cross-cohort generalization in sMRI-based AD classification. The source code is available at https://github.com/zobia111/Extended-Mixstyle.
Figures & tables
Figure 1: Higher-order feature statistics across cohorts. Bar plots of skewness and kurtosis computed from intermediate 3D U-Net feature maps across four sMRI cohorts. For each dataset, channel-wise feature distributions were aggregated over test samples and their higher-order moments were summarized. Error bars denote the 95% confidence intervals (CIs). Clear variability in skewness and kurtosis was observed across cohorts, indicating that feature distributions differ beyond mean and standard deviation.
Figure 2: EM integration into a 3D U-Net encoder and its internal operation. EM receives a batch of feature maps as input, and produces a batch of mixed feature maps as output. In detail, the batch (x1,x2,…,xB) are obtained from Layer 2, and are used to compute per-channel feature statistics m . For each sample b , the mean ( μb ), standard deviation ( σb ), skewness ( γb ), and kurtosis ( κb ) are computed (depending on the specific variant, EM1 or EM2 ). These statistics are randomly paired (e.g. between xb and x~b ), and mixed through a random sample-specific mixing coefficient λb producing mixed statistics mbmix . Each feature map is then normalized using its original statistics and renormalized using its corresponding mixed statistics, thus generating mixed feature maps that are forwarded to the next encoder layer.
Dataset
Group (Participants)
Age, years
Gender
mean ± std
male count
NACC [ 19 ]
NC (n=2524)
69.8 ± 9.9
871 (34.5%)
MCI (n=1175)
74.0 ± 8.7
555 (47.2%)
AD (n=948)
75.0 ± 9.1
431 (45.5%)
ADNI [ 20 ]
NC (n=481)
74.3 ± 6.0
235 (48.9%)
MCI (n=971)
72.8 ± 7.7
572 (58.9%)
Table 1: Participant demographics across sMRI cohorts. 3D MRI and demographic data were obtained from NACC, ADNI, AIBL, and OASIS-3. Participants were grouped as NC, MCI, or AD, with mean age and number of male participants reported where available.
Methods
ADNI
AIBL
OASIS-3
ACC
SEN
SPE
F1
ACC
SEN
SPE
F1
ACC
SEN
SPE
F1
Baseline ∗ [ 46 ]
49.47
0.563
0.744
0.508
70.80
0.574
0.805
0.575
66.45
0.578
0.844
0.534
MixUp [ 27 ]
46.34
0.548
0.732
0.476
75.03
0.593
0.821
0.595
65.21
0.562
0.838
0.523
RSC [ 33 ]
47.17
0.546
0.732
0.483
68.22
0.583
0.817
0.561
63.04
0.549
0.830
0.514
CCSDG [ 31 ]
47.39
0.570
0.740
0.488
73.22
0.578
0.820
0.573
67.54
0.587
0.840
0.539
MixStyle [ 18 ]
46.62
0.558
0.737
0.476
66.26
0.575
0.811
0.538
64.44
0.565
0.838
0.518
Table 2: Cross-dataset generalization performance on external cohorts. Models were trained on NACC and evaluated on ADNI, AIBL, and OASIS-3. All values in this table are means across three random seeds (1, 2, and 42); ∗ marks methods for which corresponding 10-seed mean results are reported in Table 3 . Best results are in bold.
Dataset
Method
1
2
3
4
5
6
7
8
9
42
μ±σ
ADNI
Baseline
0.523
0.516
0.477
0.444
0.476
0.491
0.485
0.480
0.490
0.485
0.487 ± 0.02
EM 1
0.508
0.528
0.482
0.516
0.510
0.520
0.516
0.513
0.520
0.521
0.513 ± 0.01
EM 2
0.509
0.517
0.501
0.485
0.497
0.504
0.501
0.499
0.503
0.497
0.501 ± 0.01
AIBL
Baseline
0.558
0.597
0.525
0.527
0.549
0.564
0.558
0.554
0.563
0.571
0.557 ± 0.02
EM 1
0.638
0.643
0.587
0.590
0.597
0.612
0.606
0.602
0.611
0.606
0.609 ± 0.02
EM 2
0.560
0.540
0.500
0.538
0.531
0.546
0.540
0.536
0.545
0.580
0.542 ± 0.02
Table 3: Macro-F1 scores across 10 repeated runs. Individual seed results and mean ± standard deviation are reported for the 3D U-Net baseline, EM 1 , and EM 2 across external cohorts.
Dataset
Comparison
F1 1
F1 2
Δ F1
95% CI
Raw p
Holm p
ADNI
EM 1 vs EM 2
0.5134
0.5012
0.0122
[-0.0061, 0.0304]
0.1952
0.2020
EM 1 vs Baseline
0.5134
0.4865
0.0269
[0.0051, 0.0492]
0.0152
0.0456
EM 2 vs Baseline
0.5012
0.4865
0.0147
[-0.0027, 0.0322]
0.1010
0.2020
AIBL
EM 1 vs EM 2
0.6090
0.5417
0.0673
[0.0309, 0.1039]
0.0002
0.0006
EM 1 vs Baseline
0.6090
0.5566
0.0525
[0.0119, 0.0927]
0.0126
0.0252
EM 2 vs Baseline
0.5417
0.5566
-0.0149
[-0.0454, 0.0161]
0.3450
0.3450
Table 4: Statistical significance analysis of F1 across 10 random seeds. Paired stratified bootstrap testing with 10,000 resamples was used. Pairwise differences are reported with 95% confidence intervals, raw p -values, and Holm-adjusted p -values. Bold values indicate comparisons that remain statistically significant after Holm correction ( p<0.05 ).
Layers
EM
ADNI
AIBL
OASIS-3
ACC
SEN
SPE
F1
ACC
SEN
SPE
F1
ACC
SEN
SPE
F1
Layer 1
EM 1
50.90
0.562
0.745
0.523
70.49
0.595
0.822
0.586
62.42
0.599
0.829
0.512
EM 2
50.79
0.552
0.740
0.519
72.31
0.613
0.832
0.612
60.24
0.550
0.823
0.490
Layer 2
EM 1
50.30
0.575
0.748
0.519
76.39
0.614
0.836
0.629
68.32
0.588
0.844
0.540
EM 2
49.42
0.568
0.742
0.508
66.71
0.613
0.822
0.560
64.75
0.601
0.840
0.538
Layer 3
EM 1
52.16
0.542
0.738
0.528
60.51
0.551
0.796
0.537
53.88
0.558
0.813
0.476
Table 5: Effect of EM placement within the encoder on model generalization performance. The proposed modules EM 1 and EM 2 were inserted after different encoder blocks of the 3D U-Net. Each configuration corresponds to EM modules applied after individual layers or combinations of layers within the encoder. Best results are in bold.
Figure 3: Hyperparameter sensitivity analysis of Extended MixStyle. Top row: macro-F1 sensitivity to the higher-order weighting parameters βskew and βkurt across ADNI, AIBL, and OASIS-3. Bottom row: macro-F1 sensitivity of EM 1 to the MixStyle interpolation parameter α and mixing probability p . Low-to-moderate higher-order weighting generally provides more stable performance, while several neighboring α - p configurations achieve comparable results across cohorts, indicating that the selected settings are not isolated hyperparameter choices and remain robust to moderate parameter variation.
Pooling
ADNI
AIBL
OASIS-3
ACC
SEN
SPE
F1
ACC
SEN
SPE
F1
ACC
SEN
SPE
F1
EM 1
GAP
50.30
0.575
0.748
0.519
76.39
0.614
0.836
0.629
68.32
0.588
0.844
0.540
Max
49.97
0.591
0.756
0.512
72.91
0.575
0.813
0.573
67.33
0.567
0.842
0.533
Attention
43.05
0.553
0.731
0.434
70.95
0.580
0.814
0.554
61.58
0.532
0.838
0.530
EM 2
Table 6: Effect of pooling strategy on cross-dataset generalization performance. GAP, max pooling, and attention-based pooling are compared for EM 1 and EM 2 . Results show that GAP provides the most stable overall performance across external cohorts. Best results within each EM variant are in bold.
Variant
ADNI
AIBL
OASIS-3
ACC
SEN
SPE
F1
ACC
SEN
SPE
F1
ACC
SEN
SPE
F1
Fixed EM 1
50.30
0.575
0.748
0.519
76.39
0.614
0.836
0.629
68.32
0.588
0.844
0.540
Learnable EM 1
49.64
0.578
0.748
0.510
76.09
0.571
0.832
0.585
71.27
0.581
0.840
0.533
Fixed EM 2
49.42
0.568
0.742
0.508
66.71
0.613
0.822
0.560
64.75
0.601
0.840
0.538
Learnable EM 2
49.25
0.558
0.741
0.506
71.25
0.573
0.824
0.581
61.49
0.625
0.836
0.519
Table 7: Fixed vs. learnable higher-order weighting parameters. NACC-selected fixed coefficients provide more consistent overall performance, whereas learnable coefficients improve only selected cohort-specific cases. Best results within each pair are in bold.
Figure 4: Grad-CAM visualizations on 3D sMRI samples across cohorts. The figure presents NC, MCI and AD subjects from ADNI (top row), AIBL (middle row), and OASIS-3 (bottom row). For each cohort, columns show: original sMRI scans, MixStyle baseline, EM 1 based on mean, standard deviation, and skewness, and EM 2 extending EM 1 with kurtosis.
Figure 5: t-SNE visualizations of sMRI embeddings under different training settings. Data were drawn from three cohorts: ADNI, AIBL and OASIS-3. The vanilla 3D U-Net (a) shows more cohort-specific separation, with AIBL forming relatively compact clusters. EFDM (b) increases inter-cohort overlap. EM 1 (c) shows the strongest intermixing among the external cohorts, while EM 2 (d) also improves overlap relative to the baseline, though with somewhat less uniform mixing.
Methods
ADNI
AIBL
OASIS-3
ACC
SEN
SPE
F1
ACC
SEN
SPE
F1
ACC
SEN
SPE
F1
Baseline [ 46 ]
81.16
0.626
0.859
0.572
91.53
0.620
0.955
0.636
84.16
0.570
0.958
0.683
EFDM [ 34 ]
78.29
0.558
0.840
0.510
90.92
0.468
0.969
0.552
82.29
0.518
0.953
0.637
EM 1
83.56
0.645
0.884
0.614
91.51
0.481
0.974
0.575
81.37
0.451
0.969
0.592
EM 2
81.43
0.615
0.865
0.573
92.32
0.608
0.966
0.654
84.01
0.554
0.962
0.675
Table 8: One-to-all (AD vs. all) analysis to evaluate AD classification performance across cohorts. Performance is compared among the 3D U-Net baseline, EFDM (best performing competing method), and the proposed EM 1 and EM 2 variants. Best results are in bold.
Task
Method
ADNI
AIBL
OASIS-3
ACC
SEN
SPE
F1
ACC
SEN
SPE
F1
ACC
SEN
SPE
F1
MCI vs NC
Baseline
47.90
0.374
0.691
0.490
71.30
0.481
0.762
0.369
63.09
0.629
0.631
0.170
EM 1
56.81
0.565
0.573
0.636
66.76
0.647
0.672
0.406
61.41
0.666
0.611
0.172
EM 2
56.40
0.518
0.656
0.613
69.24
0.529
0.727
0.376
67.18
0.740
0.667
0.212
MCI vs AD
Baseline
67.38
0.739
0.649
0.556
70.65
0.556
0.823
0.624
63.68
0.626
0.714
0.752
EM 1
68.95
0.796
0.649
0.586
70.71
0.683
0.725
0.671
65.84
0.647
0.740
0.769
Table 9: Binary classification performance for clinically relevant cross-dataset diagnostic tasks. Classification results are reported for MCI vs. NC and MCI vs. AD on the external ADNI, AIBL, and OASIS-3 cohorts. The baseline 3D U-Net is compared with EM 1 and EM 2 . Best results within each binary task and cohort are in bold.
Method
ADNI
AIBL
OASIS-3
NACC
ACC
SEN
SPE
F1
ACC
SEN
SPE
F1
ACC
SEN
SPE
F1
ACC
SEN
SPE
F1
Source Dataset: ADNI
Baseline
-
-
-
-
48.26
0.497
0.764
0.453
49.22
0.494
0.782
0.425
54.36
0.538
0.774
0.535
EM 1
-
-
-
-
58.69
0.529
0.792
0.509
55.43
0.561
0.814
0.476
56.69
0.547
0.782
0.548
EM 2
-
-
-
-
57.33
0.562
0.788
0.534
50.77
0.526
0.793
0.442
55.70
0.545
0.778
0.548
Source Dataset: AIBL
Table 10: Cross-source domain generalization experiments. Models are trained on one source cohort and evaluated on the remaining target cohorts. For each source setting, the baseline 3D U-Net is compared with EM 1 and EM 2 . Best results are in bold.
Method
Training Time
Overhead
Peak GPU
/Epoch (s)
(%)
Memory (GB)
Baseline [ 46 ]
659.62±2.82
0.00
8.69
MixUp [ 27 ]
759.62±2.82
15.16
8.76
MixStyle [ 18 ]
805.19±14.42
22.07
8.88
EM 1
856.96±4.30
29.92
9.10
EM 2
875.73±8.40
32.76
9.58
Table 11: Computational efficiency comparison of the baseline and augmentation-based methods. Training time is reported as mean ± standard deviation over three measured epochs.
Cross-domain generalization studies how machine-learning models retain reliable performance when data, sensors, populations, or tasks differ from training.
Although Alzheimer's disease detection via MRIs has advanced significantly thanks to contemporary deep learning models, challenges such as class imbalance, protocol variations, and limited dataset diversity often hinder their generalization capacity. To address this issue, this article focuses on the single domain generalization setting, where given the data of one domain, a model is designed and developed with maximal performance w.r.t. an unseen domain of distinct distribution. Since brain morphology is known to play a crucial role in Alzheimer's diagnosis, we propose the use of learnable pseudo-morphological modules aimed at producing shape-aware, anatomically meaningful class-specific augmentations in combination with a supervised contrastive learning module to extract robust class-specific representations. Experiments conducted across three datasets show improved performance and generalization capacity, especially under class imbalance and imaging protocol variations. The source code will be made available upon acceptance at https://github.com/zobia111/SDG-Alzheimer.
Zobia Batool, Diala Lteif, Vijaya B. Kolachalama +2
VPALab, Faculty of Engineering and Natural Sciences, Sabanci University, Istanbul, Türkiye
Alzheimer's Disease (AD) is a progressive neurodegenerative disorder that affects millions of older adults, with prevalence expected to rise significantly in the coming years. Early diagnosis, particularly during the mild cognitive impairment (MCI) stage, is critical for timely intervention. Structural Magnetic Resonance Imaging (sMRI) has emerged as a key modality for detecting AD-related brain changes, but traditional graph-based approaches often struggle with modality and inter-site heterogeneity, limiting diagnostic performance. In this paper, we propose Graph Matching Network for Alzheimer's Disease Diagnosis (GMN4AD), designed to model interactions between heterogeneous brain graphs derived from neuroimaging data. Unlike conventional methods that treat each brain graph independently, GMN4AD leverages graph matching to capture cross-graph relationships, enhancing diagnostic precision. Furthermore, we introduce a test-time domain adaptation strategy that combines contrastive learning to mitigate domain shifts during inference. Extensive experiments on three public AD datasets demonstrate that GMN4AD achieves superior performance compared to state-of-the-art methods, offering a robust and generalizable solution for AD diagnosis.
Chen Zhao, Huan Huang, Yixin Xie +2
Department of Computer Science, Kennesaw State University, Marietta, GA, 30060 · Department of Information Technology, Kennesaw State University, Marietta, GA, 30060 · School of Data Science and Analytics, Kennesaw State University, Marietta, GA, 30060 +1
Accurate assessment of Alzheimer's disease stages from MRI is important for understanding disease progression and supporting the development of computer-aided analysis tools. Deep learning models have shown promising results for Alzheimer's disease classification from MRI. However, when evaluated on single datasets, it is difficult to determine whether the observed performance reflects consistent architectural behavior or dataset-specific effects. In this work, we present a multi-dataset benchmarking framework with standardized preprocessing and evaluation. We compare five architectures: ResNet-50, ReXNet-150, Swin Transformer, ViT-Base, and ViT-Large. The models are evaluated across three MRI datasets representing severe class imbalance, train-test distribution mismatch, and large-scale augmentation. The results show broadly consistent performance patterns across datasets, with ViT-Large achieving the best performance on all three datasets and reaching up to 99.84% accuracy and 99.86% macro F1. Performance differences between architectures are most pronounced under distribution mismatch and considerably smaller in the augmented-data setting. Across datasets, lower class-level performance is observed primarily among adjacent early-stage categories, particularly Very Mild and Mild impairment. These findings highlight the importance of evaluating deep learning architectures across multiple data conditions rather than relying on single-dataset performance.
Hafija Aktar, Molly Croes, Frank S. Menniti +1
Department of Computer Science, University of Rhode Island, Kingston, RI 02881, USA · Department of Biomedical and Pharmaceutical Sciences, George & Anne Ryan Institute for Neuroscience, University of Rhode Island, Kingston, RI 02881, USA