Which cortical regions affect the prediction of cerebrospinal fluid amyloid beta from PET images by novel residual neural network (ArcheD)

Abstract

AbstractBackgroundAmyloid‐ß (Aß) aggregation, as a biological hallmark of Alzheimer’s disease (AD), is a key component of biological classification of AD in the AT(N) framework. This research aimed to develop a new deep learning model for predicting Aß cerebrospinal fluid (CSF) measures from amyloid PET scans and explore which cortical regions are most important contributors in the model’s prediction.MethodWe developed a Residual Network model (ArcheD) to predict Aß CSF from amyloid PET images, and trained it on 1,870 PET scans from the Alzheimer’s Disease Neuroimaging Initiative. We registered PET scans to MNI152 template and used the Neuromorphometrics atlas to derive cortical gray matter (GM) regions. We extracted brain regions’ contributions to model decision‐making by guided backpropagation, and visualized these contributions as relevance heatmaps. We used bootstrapping, Welch’s t‐test and FDR correction in testing significance of brain regions’ relevance to predictions in the whole cohort, and separately in clinical diagnostic classes of AD, mild cognitive impairment (MCI), and cognitively normal (CN) cases.ResultArcheD showed high accuracy in predicting Aß CSF (r = 0.81; p<0.00001). According to the relevance analysis for all samples (n = 1,390) temporal (mean 0.106; 95%CI [0.105, 0.108]) and parietal (0.105; 95%CI [0.104, 0.107]) lobes were the most important GM regions for the model decision‐making. Occipital (0.098; 95%CI [0.098, 0.100]) and frontal lobes (0.091; 95%CI [0.089, 0.92]) showed noticeably fewer relevance values. Based on between‐classes comparison, these regions contribute significantly more to CSF prediction in AD and MCI individuals compared to CN (p<0.05). Analysis of 49 smaller regions showed an almost 7‐fold difference between the most and least contributing regions.ConclusionWe developed a deep residual neural network model that predicted Aß CSF from amyloid PET scans with high accuracy, and found that temporal and parietal lobes have the highest influence on Aß CSF predictions. In the next step, we will validate the model in independent data and aim to develop the model further for clinical use.