Answers are grounded in this paper and cite their sources.
Machine learning study demonstrates improved Alzheimer disease classification in brain MRI scans, indicating the benefit of autoencoder pretraining over standard transfer-learning models.
Accurate and early detection of Alzheimer’s Disease (AD) remains a critical challenge due to the heterogeneous nature of structural and clinical disease manifestations. Existing unimodal deep learning methods often fail to capture this heterogeneity, while traditional multimodal fusion strategies rely on static feature concatenation that limits discriminative power. To address these challenges, this study proposes MHA-FusionNet, a multimodal hybrid attention network that integrates structural MRI slices with structured clinical variables for robust AD classification. The framework employs modality-specific encoders, embedding harmonisation layers, and a cross-modal attention mechanism to dynamically weight MRI and clinical contributions. Extensive experiments were conducted using CNN, VGG16, and ResNet50 baselines with and without PCA serving as comparative models against the MRI-based configuration of the proposed architecture. While PCA improved unimodal performance, all baseline architectures plateaued at 80–83% accuracy and AUC values between 0.86 and 0.91. The MRI-based classifier achieved a patient-level accuracy of 91% and an AUC of 0.92, with McNemar’s test confirming the statistical significance of this improvement over every baseline ( p < 0.001 in all cases, Bonferroni-corrected). Ablation analysis confirmed that autoencoder pretraining and data augmentation each contribute significantly to this performance. A parallel clinical-data classifier, evaluated separately since the two datasets could not be patient-paired, achieved 94.4% binary accuracy. These findings demonstrate that a rigorously validated, pretrained MRI classifier can substantially outperform standard transfer-learning baselines on this task; clinical deployment would require external validation on an independent cohort, prospective testing, and comparison against clinician-level diagnostic accuracy, none of which are performed in this study.
No takes yet. Share an insight, caveat, or question.
Karthiga et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: