Fig. 7

Deep learning with convolutional neural networks accurately predicts CTRL vs. AD astrocytes and microglia. a Architecture of the convolutional neural networks (CNNs) used for deep learning of image features from astrocyte and microglial profiles (see details in Methods section). Receiver operating characteristic (ROC) curves demonstrate the performance of the CNN model to predict the diagnosis of CTRL vs. AD based on the features of (b) astrocytic or (d) microglial high-plex images. Histograms show the within-group proportions of (c) astrocytes or (e) microglia as a function of the AD classification probability stratified by their true label (i.e., CTRL or AD). Note that both CTRL astrocytes and CTRL microglia (blue bars) tend to have a probability of AD diagnosis closer to zero, whereas AD glia (red bars) tend to be correctly classified as AD