A deep learning MRI approach outperforms other biomarkers of prodromal Alzheimer's disease.

BACKGROUND

The three core pathologies of Alzheimer's disease (AD) are amyloid pathology, tau pathology, and neurodegeneration. Biomarkers exist for each. Neurodegeneration is often detected by neuroimaging, and we hypothesized that a voxel-based deep learning approach using structural MRI might outperform other neuroimaging methods.

METHODS

First, we implement an MRI-based deep learning model, trained with a data augmentation strategy, which classifies Alzheimer's dementia and generates class activation maps. Next, we tested the model in prodromal AD and compared its performance to other biomarkers of amyloid pathology, tau pathology, and neuroimaging biomarkers of neurodegeneration.

RESULTS

The model distinguished between controls and AD with high accuracy (AUROC = 0.973) with class activation maps that localized to the hippocampal formation. As hypothesized, the model also outperformed other neuroimaging biomarkers of neurodegeneration in prodromal AD (AUROC = 0.788) but also outperformed biomarkers of amyloid (CSF Aβ = 0.702) or tau pathology (CSF tau = 0.682), and the findings are interpreted in the context of AD's known anatomical biology.

CONCLUSIONS

The advantages of using deep learning to extract biomarker information from conventional MRIs extend practically, potentially reducing patient burden, risk, and cost.