Fig. 4: Explainability methods applied to CathAI algorithms.
From: CathAI: fully automated coronary angiography interpretation and stenosis estimation
a GradCAM applied to CathAI classification of primary anatomic structure. Two original angiogram images are shown (left), alongside corresponding images highlighted by GradCAM (right) showing areas of greater importance for algorithm decisions. GradCAM-highlighted areas focused around the left coronary artery within the angiogram image, with blue color indicating lowest importance, yellow color indicating medium importance and red color indicating highest importance to CathAI Algorithm 2’s prediction. b LOVI Saliency Maps of CathAI prediction of coronary stenosis severity. Original angiogram images (top) and corresponding images with LOVI saliency maps (bottom). White pixels represent greater contribution to CathAI’s (Algorithm 4) prediction, showing that Algorithm 4 focused on pixels near the region of coronary artery stenosis in most cases. LOVI Layer Ordered Visualization of Information.
