Comparative evaluation of MRI radiomic models for Alzheimer’s disease prognosis

Abstract

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder in which early detection is critical for targeted intervention. This thesis evaluates the prognostic value of structural magnetic resonance imaging (sMRI) volumetry, high-dimensional radiomics, and cognitive assessments, independently and in combination, for predicting the progression from mild cognitive impairment (MCI) to AD approximately two years after baseline. Participants were classifed as stable MCI (sMCI) and progressive MCI (pMCI) based on a two-year diagnostic follow-up. T1-weighted MRI scans and cognitive assessments from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) were analyzed. Automated segmentation and radiomic extraction pipelines were developed to extract volumetric and radiomic features. Explainable machine learning models based on logistic regression with elastic net regularization were trained using nested cross-validation and evaluated on an independent hold-out test set. The results demonstrated that both sMRI-derived biomarkers and cognitive assessments contain valuable prognostic information. Evaluated on the independent test set, the region of interest (ROI) radiomic model achieved the highest predictive performance among the sMRI models (AUC=0.79). However, radiomic models did not demonstrate statistically signifcant improvements over the conventional volumetric model (AUC=0.73, DeLong’s test p=0.087). While the best performance was achieved by a multimodal model combining ROI radiomic features and cognitive assessments (AUC=0.82), the model combining standard volumetric features with cognitive assessments (AUC=0.80) was identifed as the optimal, most clinically practical solution. The fndings suggest that volumetric features and cognitive assessments remain practical and efective biomarkers for AD prognosis. Furthermore, the developed computational pipeline demonstrates the potential for integrating explainable prognostic models into future clinical decision support systems for risk stratifcation of MCI patients.