UTLAM: Unsupervised Two-Level Adapting Model for Alzheimer’s Disease Classification

The design of an accurate cross-domain model for Alzheimer disease AD classification from MRI scans faces critical challenges, including domain shifts caused by acquisition protocol variations. To address this issue, we propose a novel unsupervised two-level adapting model for Alzheimer’s disease classification using 3D MRI scans. In the first level, we introduce an extended mean inter- and intra-class discrepancy metric, which statistically aligns both inter-class and inter-domain discrepancies, enabling pseudo-labeling of the unlabeled samples. The second level integrates labeled source and pseudo-labeled target features into an adversarial learning, encouraging the feature extractor to generate domain-invariant representations, thereby improving model generalizability. The proposed model uses standard Alzheimer’s disease benchmarks, including ADNI and AIBL databases. Experimental results demonstrate UTLAM’s superior transfer learning capability compared to the existing baselines in identifying cognitive normal CN, AD, and mild cognitive impairment in MCI subjects. Notably, UTLAM achieves classification accuracies of (92.02%, 77.72%, and 83.04%), (92.60%, 71.45%, and 62.50%), and (93.22%, 84.80%, and 72.19%) on (CN vs. AD, MCI vs. AD, and CN vs. MCI) classifications via ADNI-1 to AIBL, ADNI-1 to ADNI-2, and AIBL to ADNI-3 transfer learnings, respectively. Without relying on a labeled target, UTLAM offers a highly practical solution for Alzheimer’s disease classification.