Abstract / Summary
Deep learning has become a promising strategy for improving disease features in medical imaging, but its application in neurodegenerative diseases is still limited. Subjective cognitive decline (SCD) is a key preclinical stage of Alzheimer’s disease (AD), but its recognition is challenging due to subtle and heterogeneous changes in the brain. A total of 234 normal controls and 229 individuals with SCD were included from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database. Preprocess structured T1 weighted MRI data using standardized procedures. We extracted MRI radiomics features and vision transformer-based deep learning features. To integrate complementary information, we propose a Hybrid Vision-Radiomics Aggregation (HVRA) framework based on Minimum Redundancy Maximum Relevance (mRMR) derived feature selection and weighted feature concatenation. Consensus features were identified through the feature selection procedure, and model performance was evaluated using a repeated cross-validation strategy within the ADNI cohort to assess internal validation performance. The HVRA model achieved the best performance with an accuracy of 88.98% ± 3.42% and an AUC of 0.941, outperforming all unimodal methods. Compared with conventional fusion strategies, including direct concatenation and learnable fusion approaches, HVRA demonstrated superior classification performance. Four consensus features were consistently selected and demonstrated significant correlations with cognitive performance. Furthermore, these features showed associations with cognitive performance, suggesting their potential relevance for characterizing SCD-related brain alterations. Integrating deep learning and radiomics features provides a more comprehensive characterization of brain alterations associated with SCD, supporting its potential for improving SCD classification.