Fig. 5: Saliency maps.
From: Deep learning predicts all-cause mortality from longitudinal total-body DXA imaging
a Integrated gradients pixel-level saliency maps. Low mortality (top row) and high mortality (bottom row) sample scans from test set. Participants are 73 year old White males with similar body measurements. Rows from left to right depict: high-energy scan, low-energy scan, single-record combined-modality saliency maps overlayed on scan. b Grad-CAM higher-level convolutional feature maps for the same participants depicted in a. Feature maps depicted are all from convolutional layers within the dense block of the combined-modality single-record model before the last pooling layer.
