Longitudinal structural MRI-based deep learning and radiomics features for predicting Alzheimer's disease progression

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Gerontology

Background:

• Alzheimer's disease (AD) is the leading cause of dementia.
• Early diagnosis of mild cognitive impairment (MCI) is vital for managing progression and treatment.
• Structural MRI and deep learning (DL) show promise for detecting neurodegeneration.

Purpose of the Study:

• To evaluate the efficacy of DL models in predicting MCI to AD conversion using longitudinal MRI data.
• To identify key MRI-based biomarkers for early AD detection.

Main Methods:

• Utilized T1-weighted MRI scans from 228 MCI participants in the ADNI database.
• Employed a 3D Residual Network (ResNet3D) for single-timepoint analysis and a Long Short-Term Memory (LSTM) model with attention for longitudinal analysis.
• Extracted radiomics features from gray matter segmentation.

Main Results:

• A single-timepoint DL model achieved a c-index of ~0.70.
• Longitudinal analysis significantly improved prediction accuracy (c-index: 0.80-0.90) with high AUC (>0.85) for 2-5 year prediction windows.
• Gray matter surface-to-volume ratio and elongation were identified as key predictive radiomics features.

Conclusions:

• Advanced DL architectures combined with structural MRI are valuable for predicting MCI to AD conversion.
• This approach facilitates early risk stratification and personalized interventions for individuals at high risk of AD progression.
• Future research requires larger, diverse cohorts to validate findings and explore further enhancements.