ArcheD, a residual neural network for prediction of cerebrospinal fluid amyloid-beta from amyloid PET images

Abstract

Detection and measurement of amyloid-beta (Abeta) aggregation in the brain is a key factor for early identification, diagnosis and progression of Alzheimer's disease (AD). We aimed to develop a novel deep learning model that aims to predict Abeta cerebrospinal fluid (CSF) concentration directly from amyloid PET images, independent of the tracer, brain reference region or preselected regions of interest. We used 1870 Abeta PET images and CSF measurements from the Alzheimer's Disease Neuroimaging Initiative to train and validate a convolutional neural network ('ArcheD') with residual connections. We evaluated the performance of ArcheD in relation to the standardized uptake value ratio (SUVR) of cortical Abeta with cerebellum as a reference region and measures of episodic memory. To interpret the trained neural network model, we identified the brain regions which the model considered as most informative for CSF prediction and compared the importance of these regions in clinical-based (cognitively normals, subjective memory complaint, mild cognitive impairment & AD) and biological-based (Abeta-positive versus Abeta-negative) classifications. ArcheD-predicted Abeta CSF values had strong correlations with measured Abeta CSF values (r=0.81; p<0.001). ArcheD-based Abeta CSF was correlated with SUVR (r<-0.53, q<0.01) and episodic memory measures (0.34 < r < 0.46; q<1x10-10) in all participants except in those with AD. We investigated the importance of brain areas to the ArcheD decision-making process and found that for both clinical and biological classifications, cerebral white matter regions significantly (q<0.01) contributed to CSF prediction, specifically in non-symptomatic and early stages of AD. However, brain stem, subcortical areas, cortical lobes, limbic lobe, and basal forebrain contributed significantly more to late disease stages (q<0.01). Considering cortical gray matter separately, the parietal lobe was the strongest predictor of CSF amyloid levels in those with prodromal or early AD. In those patients with AD, the temporal lobe played a more crucial role in the prediction of Abeta CSF from PET images. In summary, we developed a novel neural network 'ArcheD' that reliably predicted Abeta CSF concentration from Abeta PET scans. ArcheD may contribute to clinical practice in determining Abeta CSF levels and improving AD early detection. Further studies are needed to validate and tune the model for clinical use.