Spectral-conditioned hypernetworks for meta-learning neuroimaging normalization with concept-guided transparency
TL;DR
Coupling adaptive normalization with few-shot meta-learning and anatomical concept guidance provides a practical pathway toward robust and transparent AI-assisted neuroimaging diagnosis, and suggests that coupling adaptive normalization with few-shot meta-learning and anatomical concept guidance provides a practical pathway toward robust and transparent AI-assisted neuroimaging diagnosis.
Abstract
Multi-site neuroimaging studies are essential for developing clinically deployable artificial intelligence (AI) systems, yet deep learning models remain highly sensitive to scanner-induced domain shift. Variations in scanner manufacturer, field strength, acquisition protocol, reconstruction pipeline, and site-specific preprocessing can alter image intensity, texture, and spatial-frequency characteristics, thus degrading model performance when deployed on unseen scanners. This paper presents a spectral-conditioned hypernetwork framework for meta-learning neuroimaging normalization with concept-guided transparency. The proposed framework integrates a spectral-conditioned hypernetwork into the batch normalization layers of a three-dimensional convolutional encoder. For each input magnetic resonance imaging volume, a compact spectral embedding is extracted from the three-dimensional Fourier magnitude spectrum and used to dynamically generate the channel-wise normalization scale and shift parameters. This enables input-adaptive normalization without requiring explicit site labels or scanner metadata. To improve generalization to unseen imaging centers, this study formulates a bi-level meta-learning protocol in which small site-specific support sets adapt only the hypernetwork parameters, while the outer loop jointly optimizes the encoder and hypernetwork across multiple source sites. To enhance clinical interpretability, the proposed framework introduces a concept-guided regularization term that aligns gradient-weighted class activation mapping attribution maps with predefined anatomical regions of interest, including the hippocampus, ventricles, temporal pole, and entorhinal cortex. The framework is evaluated on a multi-site Alzheimer’s disease classification benchmark using leave-one-site-out validation. The proposed method improves diagnostic performance, reduces cross-site variability in the area under the receiver operating characteristic curve, and produces stronger alignment of attribution with clinically meaningful anatomical concepts compared to statistical harmonization, adversarial domain adaptation, meta-learning domain generalization, site-conditioned normalization, and concept-based interpretability baselines. These findings suggest that coupling adaptive normalization with few-shot meta-learning and anatomical concept guidance provides a practical pathway toward robust and transparent AI-assisted neuroimaging diagnosis.