Detecting amyloid-beta positivity using regions of interest from structural magnetic resonance imaging

Background and purpose: Alzheimer disease (AD) is the most common type of dementia. Amyloid-beta (A beta) positivity is the main diagnostic marker for AD. A beta positron emission tomography and cerebrospinal fluid are widely used in the clinical diagnosis of AD. However, these methods only assess the concentrations of A beta, and the accessibility of these methods is thus relatively limited compared with structural magnetic resonance imaging (sMRI). Methods: We investigated whether regions of interest (ROIs) in sMRIs can be used to predict A beta positivity for samples with normal cognition (NC), mild cognitive impairment (MCI), and dementia. We obtained 846 A beta negative (A beta-) and 865 A beta positive (A beta+) samples from the Alzheimer's Disease Neuroimaging Initiative database. To predict which samples are A beta+, we built five machine learning models using ROIs and apolipoprotein E (APOE) genotypes as features. To test the performance of the machine learning models, we constructed a new cohort containing 97 A beta- and 81 A beta+ samples. Results: The best performing machine learning model combining ROIs and APOE had an accuracy of 0.798, indicating that it can help predict A beta+. Furthermore, we searched ROIs that could aid our prediction and discovered that an average left entorhinal cortical region (L-ERC) thickness is an important feature. We also noted significant differences in L-ERC thickness between the A beta- and A beta+ samples even in the same diagnosis of NC, MCI, and dementia. Conclusions: Our findings indicate that ROIs from sMRIs along with APOE can be used as an initial screening tool in the early diagnosis of AD.