Single Domain Generalization for Alzheimer's Detection from 3D MRIs with Pseudo-Morphological Augmentations and Contrastive Learning
Organizations: VPALab, Faculty of Engineering and Natural Sciences, Sabanci University, Istanbul, Türkiye
Abstract
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.
Figures & tables
| Dataset | Group (Participants) | Age, years | Gender |
|---|---|---|---|
| mean std | male count | ||
| NACC [ NACC ] | Non-AD (n=3699) | 71.1 9.7 | 1426 (38.6%) |
| AD (n=948) | 75.0 9.1 | 431 (45.5%) | |
| ADNI [ ADNI ] | Non-AD (n=1452) | 73.3 7.3 | 807 (55.6%) |
| AD (n=369) | 74.9 7.8 | 203 (55.0%) | |
| AIBL [ AIBL ] | Non-AD (n=582) | 72.9 6.5 | 256 (44.0%) |
| Method | OASIS-3 | ADNI | AIBL | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| ACC | F1 | SEN | SPE | MCC | ACC | F1 | SEN | SPE | MCC | ACC | F1 | SEN | SPE | MCC | |
| Baseline [ Models_Genesis ] | 70.96 | 0.697 | 0.850 | 0.650 | 0.458 | 64.47 | 0.586 | 0.664 | 0.640 | 0.247 | 79.43 | 0.665 | 0.722 | 0.804 | 0.389 |
| CCSDG [ CCSDG ] | 72.20 | 0.700 | 0.756 | 0.707 | 0.429 | 66.34 | 0.588 | 0.583 | 0.684 | 0.222 | 72.31 | 0.599 | 0.696 | 0.727 | 0.293 |
| Mixup [ Mixup ] | 73.76 | 0.715 | 0.762 | 0.727 | 0.454 | 68.37 | 0.613 | 0.634 | 0.696 | 0.275 | 79.27 | 0.652 | 0.658 | 0.811 | 0.354 |
| RSC [ paper11 ] | 72.05 | 0.704 | 0.813 | 0.681 | 0.454 | 65.84 | 0.595 | 0.648 | 0.661 | 0.253 | 78.52 | 0.654 | 0.709 | 0.796 | 0.371 |
| EFDM [ zhang2022exact ] | 74.84 | 0.724 | 0.756 | 0.745 | 0.468 | 63.32 | 0.568 | 0.602 | 0.641 | 0.199 | 84.27 | 0.694 | 0.608 | 0.875 | 0.404 |
| Dataset | Comparison | Method 1 F1 | Method 2 F1 | F1 | 95% CI | Raw | Holm |
|---|---|---|---|---|---|---|---|
| (mean SD) | (mean SD) | ||||||
| ADNI | Ours vs Baseline | 0.088 | [0.070, 0.106] | 0.001 | 0.009 | ||
| Ours vs AAGN | 0.085 | [0.064, 0.104] | 0.001 | 0.009 | |||
| Ours vs FM-SDG | 0.017 | [-0.003, 0.038] | 0.106 | 0.515 | |||
| AIBL | Ours vs Baseline | 0.031 | [-0.005, 0.072] | 0.106 | 0.515 | ||
| Ours vs AAGN | 0.074 | [0.022, 0.125] | 0.009 | 0.054 |
| Configuration | OASIS-3 | ADNI | AIBL | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| ACC | F1 | SEN | SPE | MCC | ACC | F1 | SEN | SPE | MCC | ACC | F1 | SEN | SPE | MCC | |
| Baseline [ Models_Genesis ] | 70.96 | 0.697 | 0.850 | 0.650 | 0.458 | 64.47 | 0.586 | 0.664 | 0.640 | 0.247 | 79.43 | 0.665 | 0.722 | 0.804 | 0.389 |
| Baseline + SupCon | 65.37 | 0.634 | 0.699 | 0.634 | 0.306 | 62.60 | 0.569 | 0.645 | 0.621 | 0.216 | 80.94 | 0.679 | 0.722 | 0.821 | 0.410 |
| Baseline + Morphology | 66.15 | 0.641 | 0.705 | 0.643 | 0.320 | 62.22 | 0.566 | 0.650 | 0.615 | 0.215 | 80.33 | 0.671 | 0.709 | 0.816 | 0.395 |
| With Fixed Morphology | 68.01 | 0.660 | 0.731 | 0.659 | 0.358 | 63.76 | 0.581 | 0.664 | 0.631 | 0.239 | 78.97 | 0.662 | 0.734 | 0.797 | 0.390 |
| w/o Soft Morphology | 75.78 | 0.734 | 0.762 | 0.756 | 0.484 | 71.88 | 0.636 | 0.596 | 0.750 | 0.298 | 85.17 | 0.704 | 0.608 | 0.885 | 0.421 |
| SupCon Weight ( ) | |||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| OASIS-3 | ADNI | AIBL | |||||||||||||
| ACC | F1 | SEN | SPE | MCC | ACC | F1 | SEN | SPE | MCC | ACC | F1 | SEN | SPE | MCC | |
| 0.3 | 74.38 | 0.726 | 0.813 | 0.714 | 0.486 | 72.60 | 0.665 | 0.737 | 0.723 | 0.383 | 83.21 | 0.698 | 0.696 | 0.851 | 0.432 |
| 0.5 | 79.35 | 0.765 | 0.741 | 0.816 | 0.535 | 80.18 | 0.721 | 0.650 | 0.840 | 0.450 | 87.44 | 0.748 | 0.696 | 0.899 | 0.511 |
| 0.7 | 72.20 | 0.679 | 0.596 | 0.776 | 0.361 | 77.10 | 0.710 | 0.772 | 0.771 | 0.461 | 85.48 | 0.720 | 0.671 | 0.880 | 0.459 |
| 0.9 | 77.64 | 0.753 | 0.777 | 0.776 | 0.520 | 75.89 | 0.672 | 0.604 | 0.798 | 0.359 | 82.75 | 0.694 | 0.696 | 0.845 | 0.425 |
| Method | OASIS-3 | ADNI | AIBL | NACC | ||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| ACC | F1 | SEN | SPE | MCC | ACC | F1 | SEN | SPE | MCC | ACC | F1 | SEN | SPE | MCC | ACC | F1 | SEN | SPE | MCC | |
| Source Dataset: OASIS-3 | ||||||||||||||||||||
| Baseline [ Models_Genesis ] | - | - | - | - | - | 68.15 | 0.595 | 0.542 | 0.717 | 0.220 | 73.98 | 0.610 | 0.684 | 0.747 | 0.304 | 73.90 | 0.667 | 0.669 | 0.757 | 0.365 |
| Ours | - | - | - | - | - | 67.66 | 0.593 | 0.553 | 0.708 | 0.221 | 74.13 | 0.605 | 0.646 | 0.754 | 0.285 | 78.72 | 0.708 | 0.651 | 0.822 | 0.427 |
| Source Dataset: ADNI | ||||||||||||||||||||
| Baseline [ Models_Genesis ] | 61.80 | 0.608 | 0.762 | 0.557 | 0.292 | - | - | - | - | - | 56.58 | 0.492 | 0.772 | 0.538 | 0.201 | 63.46 | 0.582 | 0.688 | 0.621 | 0.251 |
| Method | Params (M) | GFLOPs | Train Time/Epoch (s) |
|---|---|---|---|
| Baseline [ Models_Genesis ] | 7.6 | 769.871 | 652.76 |
| AAGN [ jiang2024anatomy ] | 7.6 | 769.871 | 690.76 |
| Ours | 7.7 | 769.871 | 2012.14 |
Alzheimer's Disease Classification
Alzheimer’s disease classification uses clinical, behavioral, and brain-imaging data to distinguish disease stages and predict progression.
See all 99 papers