A Deep Learning Model to Predict a Diagnosis of Alzheimer Disease by Using ¹⁸F-FDG PET of the Brain

Abstract

To develop and validate a deep learning algorithm that predicts the final diagnosis of Alzheimer disease (AD), mild cognitive impairment, or neither at fluorine 18 (18 F) fluorodeoxyglucose (FDG) PET of the brain and compare its performance to that of radiologic readers. Materials and Methods: Prospective 18 F-FDG PET brain images from the Alzheimer's Disease Neuroimaging Initiative (ADNI) (2109 imaging studies from 2005 to 2017, 1002 patients) and retrospective independent test set (40 imaging studies from 2006 to 2016, 40 patients) were collected. Final clinical diagnosis at follow-up was recorded. Convolutional neural network of InceptionV3 architecture was trained on 90% of ADNI data set and tested on the remaining 10%, as well as the independent test set, with performance compared to radiologic readers. Model was analyzed with sensitivity, specificity, receiver operating characteristic (ROC), saliency map, and t-distributed stochastic neighbor embedding. Results: The algorithm achieved area under the ROC curve of 0.98 (95% confidence interval: 0.94, 1.00) when evaluated on predicting the final clinical diagnosis of AD in the independent test set (82% specificity at 100% sensitivity), an average of 75.8 months prior to the final diagnosis, which in ROC space outperformed reader performance (57% [four of seven] sensitivity, 91% [30 of 33] specificity; P , .05). Saliency map demonstrated attention to known areas of interest but with focus on the entire brain. Conclusion: By using fluorine 18 fluorodeoxyglucose PET of the brain, a deep learning algorithm developed for early prediction of Alzheimer disease achieved 82% specificity at 100% sensitivity, an average of 75.8 months prior to the final diagnosis.

Predicting Alzheimer’s disease progression using deep recurrent neural networks

“…In this study, we proposed a machine learning algorithm to predict multimodal AD markers (e.g., ventricular volume, cognitive scores, etc) and clinical diagnosis of individual participants for every month up to six years into the future. Most previous work has focused on a "static" variant of the problem, where the goal is to predict a single timepoint (Duchesne et al, 2009;Stonnington et al, 2010;Zhang and Shen, 2012;Moradi et al, 2015;Albert et al, 2018;Ding et al, 2018) or a set of pre-specified timepoints in the future (regularized regression; (Wang et al, 2012;Johnson et al, 2012;McArdle et al, 2016;Wang et al, 2016)). By contrast, our goal is the longitudinal prediction of clinical diagnosis and multimodal AD markers at a potentially unlimited number of timepoints into the future 1 , as defined by The Alzheimer's Disease Prediction Of Longitudinal Evolution (TADPOLE) challenge , which arguably a more relevant and complete goal for tasks, such as prognosis and cohort selection.…”

Quantifying brain metabolism from FDG‐PET images into a probability of Alzheimer's dementia score

“…Most relevant to our work, 3D CNNs have shown success in the classification of DAT using MRI (Hosseini‐Asl, Gimel'farb, & El‐Baz, ; Payan & Montana, ). For FDG‐PET, however, existing deep learning studies have employed 2D CNNs which do not take full advantage of the spatial topographic patterns inherent in FDG‐PET images (Ding et al, ; Liu, Cheng, & Yan, ). Neural networks are often described as black boxes.…”

Predicting Alzheimer’s Disease by Hierarchical Graph Convolution from Positron Emission Tomography Imaging

“…The intrinsic nature of connectivity disorder of AD further motivates studies that can identify the optimized representation for brain images [13], which can be generalized as the task of learning low-dimensional manifold embedded in a highdimensional space [14]. While deep learning methods such as Convolutional Neural Network (CNN) can effectively learn the lower-to-higher representation of images and be used for AD classification [15], [16], it is important to incorporate irregular structures (comparing with regular image grid) of graph into the deep learning scheme [17]. By defining PET data extracted from Regions of Interest (ROIs) as signals on the nodes of a graph, we can perform signal filtering and representation learning on graph, similar to the signal filtering and feature extraction (e.g.…”