FSE-CNN: Fusion of Training-Trajectory Neural Evidence and Spatial-Scale Morphology for Cross-Cohort Brain MRI Classification

Structural magnetic resonance imaging supports the characterization of Alzheimer’s disease (AD)-related cognitive impairment. But existing cross-cohort AD classification methods remain limited in addressing acquisition and population shifts when source data cannot be revisited and target-side adaptation is excluded. In particular, current cross-cohort AD classification may underutilize diverse disease-related information learned from source data, while explicit characterization of regional, scale-dependent gray-matter morphology remains limited. To address these gaps, we propose FSE-CNN for source-free cross-cohort classification of AD-related cognitive status. First, the neural pathway derives decision evidence from model states retained during source training, with the inference configuration determined exclusively from source data. Second, the spatial-scale geometric evidence encoding module characterizes gray-matter morphology through subject-relative spatial organization and finite physical scales, providing complementary geometric evidence for cross-cohort classification. Finally, the resulting neural and geometric evidence are combined through protocol-specific fusion. We evaluated FSE-CNN on two cross-cohort protocols. In OASIS-2 → OASIS-1, FSE-CNN achieved 83.48% accuracy, 85.14% balanced accuracy, and 88.60% AUC, numerically exceeding Harmonizing Flows by 1.70, 1.26, and 0.39 percentage points, respectively. In ADNI → OASIS-2, it achieved 85.79% accuracy, 85.55% balanced accuracy, and 85.03% AUC, with corresponding gains of 0.71, 0.67, and 0.21 percentage points. Overall, these results demonstrate the effectiveness of FSE-CNN for cross-cohort AD-related cognitive-status classification.