Deep learning detection of informative features in [18F] flortaucipir PET for Alzheimer’s disease classification

Abstract

AbstractBackgroundDeep learning has been applied to disease classification using neuroimaging data. To date, most classification of Alzheimer’s disease (AD) or mild cognitive impairment (MCI) by deep learning has focused on MRI, FDG PET, and amyloid PET. Here we developed a deep learning classification model using tau PET measured with [18F]flortaucipir to classify AD or MCI against cognitively normal older adults (CN). We also visualized the process of the deep learning classification to examine whether it could detect informative features or patterns that would potentially improve diagnostic classification.MethodWe downloaded 458 tau PET images (196 CN, 196 MCI, and 66 AD) from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) and included only one scan per individual. SPM12 was used to process the tau PET data using standard techniques. We used a 3D convolution neural network (CNN) method for the classification, and applied a layer‐wise relevance propagation (LRP) algorithm to identify informative features and to visualize the classification results (Fig. 1). Five‐fold cross validation was applied, where 70% of the entire data set was used for model training, 20% for model testing, and 10% for independent validation.ResultOur deep learning‐based classification model distinguished AD from CN with an average accuracy of 96.2% (sensitivity: 95.4%; specificity: 96.9%) in an independent validation set, whereas MCI vs. CN showed 64.2% accuracy (sensitivity: 48.6%; specificity: 82.4%) (Table1). Visualization of deep learning decisions using the LRP algorithm identified patterns similar to the regions showing significant diagnostic group differences by conventional SPM analysis. Fig. 2 compared SPM contrasts to relevance heatmaps of 3D‐CNN classification.ConclusionOur results showed that tau PET‐based classification yielded higher accuracies than other imaging modalities and the LRP approach detected informative features in tau PET for AD classification. MCI classification was much less accurate reflecting the known heterogeneity within this group. The LRP‐based deep learning strategy can help visualize patterns of tau deposition in AD, MCI, or CN to assist with identification of features and patterns. Future directions include multi‐modality analysis and prognostic prediction.