The detection of Alzheimer's Disease (AD) using structural Magnetic Resonance Imaging (MRI) and Machine Learning (ML) often focuses on late-stage atrophy patterns. End-to-end deep learning models address this by considering MRI signal intensities. However, their explainability components typically focus on attention regions, neglecting underlying patterns. This work overcomes both problems by training and explaining time-to-event models utilizing Radiomics features. SHapley Additive exPlanations (SHAP) and high-level explanations were combined to interpret the effects of MRI texture, shape, and volumes, as well as neuro-psychological and cognitive tests, and socio-demographic features on the AD risk score. All models were trained and internally validated on the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset. External validation was performed on the Australian Imaging Biomarkers and Lifestyle flagship study of Ageing (AIBL) and the Open Access Series of Imaging Studies version 3 (OASIS-3). The results demonstrate that Radiomics features add value to models trained on cognitive tests, socio-demographics, genetics, and MRI volumes, particularly for long-term AD predictions. On average, the Radiomics-based models slightly outperformed the comparison models by between 0.11% points and 3.02% points in terms of Brier scores for the eight-year prediction. Despite varying data distributions during external validation, the models demonstrate moderate to high reproducibility. The analysis of Radiomics features uncovered complex associations with AD, including tissue with complex texture in the left entorhinal cortex, an irregular shape of the right amygdala, and a fine-granular texture of the left middle temporal gyrus. All models showed reasonable concordance with the Voxel-Based Morphometry (VBM).
Bloch et al. (Sat,) studied this question.