Reliability Metrics of Explainable CNN based on Wasserstein Distance for Cardiac Evaluation

Abstract In recent works, convolutional neural networks (CNN) have been used in the non-invasive examination of the cardiac region for estimating pulmonary artery wedge pressure (PAWP) from chest radiographs.Moreover, because CNNs are able to output activated regions, physicians can estimate PAWP along with reasons. However, when new patient radiograph data are fed into the CNN, there is a possibility that activated regions that contain areas other than the cardiac appear.In this case, although we expect a large estimation error, it is not well known.Therefore, we verify this hypothesis by distance theory and statistic approaches.In particular, we build the probability distributions for the cardiac region and the regression activation map (RAM) and measure the similarity between these distributions by Wasserstein distance (WSD).When the CNN estimates PAWP from areas other than the cardiac region, the WSD value is high.Therefore, WSD is a reliability metrics for explainable CNN.We created two groups, normal and anomaly classes, based on WSD values.Chest radiographs which had a high WSD were assigned to the anomaly class, and those with a low WSD were assigned to the normal class. By comparing the normal and anomaly classes based on the PAWP estimation error, we confirmed that the errors from the anomaly class were higher than those of the normal class.Therefore, physicians need to be aware that there might be large estimation errors when activated regions contain areas other than the cardiac.