Uncertainty-Disturbance and Symmetry-Guided Unsupervised Domain Adaptation for Cross-Center Dementia Classification in Brain Imaging

Abstract

This paper addresses the domain shift problem commonly encountered in cross-center and cross-device deployment for binary dementia imaging classification, and proposes an unsupervised domain adaptation framework that jointly improves target-domain generalization and prediction reliability through uncertainty disturbance and symmetry-guided learning. Built on adversarial domain alignment, the proposed method introduces uncertainty estimation to characterize the prediction risk of unlabeled target-domain samples and reduce the misleading influence of high-uncertainty samples during training, thereby improving the stability of cross-domain optimization. In addition, the framework leverages the anatomically consistent left–right hemispheric symmetry by decomposing image representations into symmetric and asymmetric components with corresponding constraints and adaptive fusion, encouraging the model to focus on morphological cues related to neurodegenerative changes rather than domain-specific imaging discrepancies. Bidirectional transfer experiments on ADNI and OASIS demonstrate the effectiveness of the proposed method. In the ADNI $\rightarrow $ OASIS setting, the proposed method achieves an Accuracy of 84.75%, Precision of 84.21%, F1-Score of 84.31%, and AUC of 85.01%. In the OASIS $\rightarrow $ ADNI setting, it obtains an Accuracy of 72.18%, Precision of 71.68%, F1-Score of 71.78%, and AUC of 72.48%. Moreover, t-SNE, Grad-CAM, and uncertainty visualizations show clearer intra-class aggregation, more consistent attention regions, and more reasonable uncertainty-related risk distributions. Ablation studies and hyperparameter sensitivity analyses further validate the complementary contributions and robustness of the uncertainty mechanism and symmetry-guided constraint, providing a transferable technical pathway for cross-center dementia-assisted classification.