Integrating standard and native spaces for radiomics and brain network analysis in Alzheimer's disease diagnosis and prognosis.

INTRODUCTION

Structural MRI analysis for Alzheimer's disease (AD) is limited by balancing group-level comparability in standard space with anatomical fidelity in native space. We therefore propose a multi-space, hybrid-feature framework, integrating radiomics and network metrics from both spaces to classify AD and predict mild cognitive impairment (MCI) progression.

METHODS

An integrated dual-space analytical framework was applied to T1-weighted MRI data. Models were developed on 1,477 participants from Alzheimer's Disease Neuroimaging Initiative (ADNI) and externally tested on an independent cohort of 1,349 participants from National Alzheimer's Coordinating Center (NACC). The framework extracts parallel radiomic and graph-based network features from both Montreal Neurological Institute (MNI) standard space and native space. These features were used to build machine learning models for three-class diagnosis (NC vs. MCI vs. AD) and 6-year prognostic prediction of MCI-to-AD conversion. For each task, the models using standard-space, native-space, and combined-space features were systematically compared. Model interpretation was performed using Shapley Additive Explanations (SHAP), and the features were validated against established AD biomarkers.

RESULTS

The combined-space model demonstrated superior performance in both diagnostic classification (Macro-Averaged AUC: 0.96 in ADNI cohort, 0.94 in NACC cohort) and prognostic prediction of MCI-to-AD conversion (C-index: 0.83; HRs: 7.60, 95%CIs: 4.57-12.64). The extracted features in the ADNI cohort demonstrated significant correlations with APOE ε4 genotype, cognitive scores, and CSF biomarkers.

CONCLUSION

Integrating multi-scale features from both standard and native spaces enhances AD diagnosis and prognosis accuracy more effectively than conventional single-space analysis.