Can Structural MRIs be used to Detect Amyloid Positivity using Deep Learning?

Abnormal β-amyloid (Aβ) accumulation in the brain is an early indicator of Alzheimer’s disease and is assessed using PET or CSF assays, both highly invasive procedures. As new anti-Alzheimer’s treatments have recently been developed to target amyloid pathology, there is great interest in understanding how well Aβ positivity (Aβ+) can be predicted from Tl-weighted brain MRI (T1w), as a less invasive alternative. For this Aβ+ prediction task, we compared (1) classical machine learning algorithms (including logistic regression, XGBoost and shallow artificial neural networks) versus (2) deep learning models, based on 2D and 3D convolutional neural networks. To combine the strengths of both shallow and deep neural networks, we also tested a hybrid ANN-CNN for the same amyloid prediction task. All models were trained on paired MRI/PET data collected from 762 elderly subjects (mean age: 75.1 yrs. ±7.6SD;394F/368M; 459 healthy controls, 67 with MCI and 236 with dementia) scanned as part of the Alzheimer’s Disease Neuroimaging Initiative. Results were compared based on each model’s balanced accuracy and F1 scores. Though the classical machine learning algorithms gave a balanced accuracy of 0.771, the hybrid network and the 3D CNN performed best- the CNN alone yielded a balanced accuracy of 0.815 with MRI-derived CSF maps as input. Although tests on more diverse data are warranted, deep learned models from standard MRI show promise for Aβ+ estimation, before considering more invasive procedures.