Novel residual neural network approach for analyzing amyloid PET scans: Associations with cerebrospinal fluid beta amyloid and episodic memory

Abstract

AbstractBackgroundDiagnosing Alzheimer’s disease (AD) based on the biological definition depends on accurately determining amyloid‐β (Aβ) status. In this research, we aimed to create a deep learning model for predicting Aβ cerebral spinal fluid (CSF) measures directly from amyloid PET scans, and compare the result with AD clinical hallmarks.MethodWe developed a novel residual neural network (ResNet) model to predict CSF Aβ from amyloid PET scans and trained the model using 1,870 scans from the Alzheimer’s Disease Neuroimaging Initiative (ADNI). Episodic memory (EM) was measured with immediate and delayed recall in the Rey Auditory Verbal Learning Test (AVLTim/AVLTdel) and in Logical Memory (LMim/LMdel) subtest from the Wechsler Memory Scale. We investigated associations of predicted CSF values with EM measures and compared our model against the traditional approach of determining cortical amyloid with a standardized uptake value ratio (SUVR) measure. We used Pearson correlations to investigate the associations of cortical amyloid with CSF and AD clinical markers. Further, we looked at these correlations separately in different clinically defined groups (cognitively normal, mild cognitive impairment (MCI), and AD).ResultResNet accurately predicted Aβ CSF measures (r = 0.93, p<1e‐10). SUVR significantly correlated with clinical CSF (r←0.65, p<1e‐10) and with ResNet model predicted CSF (r←0.67, p<1e‐10). The predicted CSF is also significantly associated with EM (LMim and LMdel, r = 0.46; AVLTim, r = 0.40; AVLTdel, r = 0.34; p<1e‐10). Correlations in those with AD were generally weaker (SUVR, r←0.37; EM, 0.02<r<0.29) than in those with MCI (SUVR, r←0.70; EM, 0.27<r<0.43) or in cognitively healthy (SUVR, r←0.56; EM, 0.16<r<0.28).ConclusionWe developed to our knowledge the first ResNet model, which predicts Aβ CSF measures with high accuracy. Predicted CSF values strongly correlate with clinical AD hallmarks. However, there are many questions that warrant further study, and we aim to investigate our model in relation to early prediction and diagnosis of AD, and in studying the discordance between biological and clinical AD classifications.