Nested Continual Learning with Elastic Weight Consolidation for Progressive Alzheimer’s Disease Classification from MRI
- 1Department of Electrical and Electronics Engineering, Istanbul University-Cerrahpaşa, İstanbul, Türkiye
- 2Biomedical Device Technology, Namık Kemal University, Tekirdağ, Türkiye
Abstract
Introduction/ Objective: MRI-based Alzheimer’s staging is often modeled as separate tasks, although clinical assessment progresses from screening to more refined staging. This study developed a continual-learning framework for retaining earlier diagnostic knowledge.
Methods: A nested multi-head framework was implemented for sequential learning of related binary classification tasks from structural MRI slices. Because verified subject identifiers were unavailable in the Mendeley dataset, the primary experiments were kept at the level of slice-based methodological analysis. Subject-level external validation was additionally carried out on OASIS-VBM with subject-wise three-fold cross-validation.
Results: In the Mendeley slice-level evaluation, the proposed EWC + replay model retained high post-sequence discrimination, with AUC values of 0.976, 0.985, and 1.000 for T1, T2, and T3. The ECE values were 0.034, 0.057, and 0.187. When the model was tested at the subject level on OASIS, AUC values decreased to 0.753, 0.748, and 0.749. This difference shows that the slice-level results should be interpreted cautiously.
Discussion: The findings support the use of the proposed framework for studying sequential retention and forgetting in hierarchical AD staging. However, Mendeley results are methodological findings at the slice level and should not be interpreted as patient-level clinical performance.
Conclusion: This study presents a nested continual-learning framework for hierarchical Alzheimer’s disease staging. The framework was tested with ablation, calibration, baseline, and subject-level validation analyses. The results show that it can be used to study retention across related AD staging tasks. Future studies using subject-wise longitudinal cohorts, 3D models, and multimodal data may further improve its clinical relevance.
Related
- Multimodal patch-based deep learning for early Alzheimer’s disease classification using MRI and PET
- An uncertainty-aware multimodal deep learning framework for Alzheimer’s disease classification using MRI and electronic health records
- Prospective validation of an uncertainty-aware machine learning model for secondary progressive multiple sclerosis classification
- Multi-class Alzheimer’s disease classification using VGG19 transfer learning with hybrid Harris Hawks Optimization and random forest
- Hybrid deep learning and meta learning for Alzheimer disease classification via multimodal radiomic and visual features
BibTeX
@article{akgundogdu2026nested,
title = {{Nested Continual Learning with Elastic Weight Consolidation for Progressive Alzheimer’s Disease Classification from MRI}},
author = {Abdurrahim Akgündoğdu and Şerife Çeli̇kbaş},
journal = {Current Medical Imaging},
year = {2026},
month = aug,
doi = {10.2174/0115734056512153260820074645},
url = {https://doi.org/10.2174/0115734056512153260820074645},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Nested Continual Learning with Elastic Weight Consolidation for Progressive Alzheimer’s Disease Classification from MRI
AU - Abdurrahim Akgündoğdu
AU - Şerife Çeli̇kbaş
PY - 2026
DA - 2026/08/31
JO - Current Medical Imaging
DO - 10.2174/0115734056512153260820074645
UR - https://doi.org/10.2174/0115734056512153260820074645
AB - Introduction/ Objective: MRI-based Alzheimer’s staging is often modeled as separate tasks, although clinical assessment progresses from screening to more refined staging. This study developed a continual-learning framework for retaining earlier diagnostic knowledge. Methods: A nested multi-head framework was implemented for sequential learning of related binary classification tasks from structural MRI slices. Because verified subject identifiers were unavailable in the Mendeley dataset, the primary experiments were kept at the level of slice-based methodological analysis. Subject-level external validation was additionally carried out on OASIS-VBM with subject-wise three-fold cross-validation. Results: In the Mendeley slice-level evaluation, the proposed EWC + replay model retained high post-sequence discrimination, with AUC values of 0.976, 0.985, and 1.000 for T1, T2, and T3. The ECE values were 0.034, 0.057, and 0.187. When the model was tested at the subject level on OASIS, AUC values decreased to 0.753, 0.748, and 0.749. This difference shows that the slice-level results should be interpreted cautiously. Discussion: The findings support the use of the proposed framework for studying sequential retention and forgetting in hierarchical AD staging. However, Mendeley results are methodological findings at the slice level and should not be interpreted as patient-level clinical performance. Conclusion: This study presents a nested continual-learning framework for hierarchical Alzheimer’s disease staging. The framework was tested with ablation, calibration, baseline, and subject-level validation analyses. The results show that it can be used to study retention across related AD staging tasks. Future studies using subject-wise longitudinal cohorts, 3D models, and multimodal data may further improve its clinical relevance.
SN - 1573-4056
ER -