Explainable deep learning for myocardial perfusion imaging with a Grad‑CAM‑based loss function

As deep learning methods become more frequent in medical imaging, model explainability becomes essential in order to ensure trust. Gradient-weighted class activation mapping (Grad-CAM) is a widely used technique for visualizing which regions of an input contribute to a network’s decision. In this work, we introduce a novel Grad-CAM–based loss function that leverages prior knowledge of target regions to constrain and improve the relevance assigned by Grad-CAM. The approach is applied to myocardial perfusion imaging (MPI), a standard examination method for evaluating suspected coronary artery disease (CAD). When CAD is suspected, invasive coronary angiography is usually performed to assess the degree of stenosis using quantitative coronary angiography (QCA). A convolutional neural network (CNN) was trained to predict QCA values for the three main coronary arteries of the heart (LAD, RCA, LCx) from MPI. When training the CNN with the proposed loss function, area under the receiver operating characteristics curve (AUC) values of 79.2%, 93.1%, 86.2% were achieved for LAD, RCA, and LCx respectively. The proposed loss function remains competitive to standardized functions, while also producing heatmaps which less frequently assign relevance to incorrect regions.