Longitudinal Structural MRI-Based Deep Learning And Radiomics Features For Predicting Alzheimer\\\\\\\'s Disease Progression

Alzheimer\\\'s disease (AD), the leading cause of dementia worldwide, affects more than 55 million individuals and gen-erates annual healthcare costs exceeding two trillion USD [14]. A substantial proportion (30–40% per year) of pa-tients with mild cognitive impairment (MCI) progress to AD [2], making early and accurate prognostication essential for timely intervention, trial enrichment, and resource allocation. This paper presents a comprehensive review of a re-cent longitudinal MRI-based study by Aghajanian et al. [1], which integrates three-dimensional (3D) convolutional neural networks (CNNs), time-aware long short-term memory (T-LSTM) networks with attention mechanisms, and radiomics features to predict MCI-to-AD conversion using structural MRI. The cohort comprises 228 ADNI MCI participants with at least three T1-weighted MRI scans over an 18-month window (684 scans in total) [1]. A 3D Res-Net-18 backbone [9] extracts volumetric features, fed into a T-LSTM incorporating inter-scan intervals and attention mechanisms [10]. The best longitudinal model achieves a concordance index (c-index) of 0.90, with time-specific AUCs of 0.96, 0.94, and 0.89 for 2-, 3-, and 5-year conversion prediction, respectively, and an approximate 11-fold hazard ratio between high- and low-risk groups [1]. This review analyzes the methodology, highlights its strengths and weaknesses, and discusses key implications for clinical translation.