جمان الجاسم

Brain image analysis plays a crucial role in the early diagnosis, treatment, an‎d monitoring of mental disorders. However, domain shift remains a major challenge in developing robust deep learning models for brain imaging. This issue arises due to variations in imaging protocols, equipment, an‎d environmental conditions, which can lead to performance degradation on unseen scans. Existing deep learning models heavily rely on large amounts of labeled data, which are difficult to obtain in the medical field due to high annotation costs an‎d privacy concerns. To address these challenges, we propose a novel Unsupervised Two-level Adapting Model (UTLAM) for Alzheimer’s disease classification. At the first level, UTLAM introduces an extended Mean Inter an‎d Intra Class Discrepancy (MIID) metric to statistically align both inter-class an‎d inter-domain discrepancies, thereby facilitating the pseudo-labeling of unlabeled target samples. At the second level, features from labeled source data an‎d pseudo-labeled target data are used within an adversarial learning framework to confuse the extractor, promoting the generation of domain-invariant features an‎d improving model generalization. Without requiring labeled target data, UTLAM offers a practical solution for Alzheimer’s disease classification by eliminating the need for expensive an‎d time-consuming annotations. We eva‎luate the model on stan‎dard Alzheimer’s disease benchmarks, including the ADNI an‎d AIBL databases. Experimental results demonstrate superior transfer learning performance compared to existing baselines in classifying cognitively normal (CN), Alzheimer’s disease (AD), an‎d mild cognitive impairment (MCI) subjects. Specifically, UTLAM achieves accuracies of (92.02%, 77.72%, an‎d 83.04%), (92.60%, 71.45%, an‎d 62.50%), an‎d (93.22%, 84.80%, an‎d 72.19%) for ADNI-1 → AIBL, ADNI-1 → ADNI-2, an‎d AIBL → ADNI-3 transfer learning tasks on CN vs. AD, MCI vs. AD, an‎d CN vs. MCI classifications, respectively.

: بيماري آلزايمر , يادگيري عميق , ؛ تطبيق دامنه؛ , تراز ويژگي؛ , يادگيري خصمانه

: Alzheimer’s disease , Deep learning , Domain adaptation , Feature alignment , Adversarial learning

Juman Aljasim

Rahman Farnoosh