Figure 4
From: Multimodal fusion models for pulmonary embolism mortality prediction
Interpretation of the model predictions for the image and EHR modalities: (a) EHR selected feature set from the multimodality classifier. The features presented are the average over the test set. (b) Grad-Cam visualization results of correctly classified examples: pairs of 2D slices taken from 3D CTPA scans and their corresponding heatmap results. High activations (red, yellow and green) indicate the focus areas that the model selected for classification. Note the high activations in anatomically relevant regions such as the pulmonary trunk, RV and LV chamber, aorta, IVC and heart area.
