Fig. 3: Grad-CAM* resultant saliency maps for five representative COVID-19 patients from testing set.

All images are of correctly predicted positive by 3D model. Within the heatmap, areas of red indicate activation of the algorithm related with COVID-19 prediction. a, b Images and (f, g) associated maps from Hubei, China cohort. c Image and (h) associated map from Tokyo, Japan cohort. d Image and (i) associated map from an advanced case in Milan, Italy Center #1. Note activation in non-consolidating areas for prediction of COVID-19, indicating specific features independent of pneumonia-related consolidation are learned. e Image and (j) associated map of an advanced case in Milan, Italy Center #2. Note: case (e) represents an unseen testing center from training/validation centers. *footnote: Grad-CAM images are produced from preprocessed input data, including cropping to lung region and resizing to fixed dimension, which may result in visible changes to anatomic aspect ratio.