Fig. 4: Grad-CAM and guided grad-CAM AI explainability applied to DNNs for three echo tasks.
From: Multiview deep learning improves detection of major cardiac conditions from echocardiography
Grad-CAM and guided grad-CAM heat maps showing the class-weighted activations of the final convolutional layer in our single-view DNNs for (top to bottom) LV/RV abnormalities (VD), diastolic dysfunction (DD), and valvular regurgitation (Valve) using A2c, A4c, and PLAX views for VD and DD, and using A4c, A5c, and PLAX views for valve. For each panel, the left image is the original echo frame, the middle image is the grad-CAM, and the right image is the guided grad-CAM. Brighter red (grad-CAM) or pink (guided grad-CAM) areas indicate areas of greater importance for that DNN’s prediction from that frame.
