Cognitive Explainers of Graph Neural Networks Based on Medical Concepts

Although deep neural networks (DNN) have achieved state-of-the-art performance in various fields, some unexpected errors are often found in the neural network, which is very dangerous for some tasks requiring high reliability and high security. The non-transparency and unexplainably of Convolutional Neural Networks (CNN) still limit its application in many fields, such as medical care and finance. Despite current studies that have been committed to visualizing the decision process of DNN, most of these methods focus on the low level and do not take into account the prior knowledge of medicine. In this work, we propose an interpretable framework based on key medical concepts, enabling CNN to explain from the perspective of doctors' cognition. We propose an interpretable automatic recognition framework for the ultrasonic standard plane, which uses a concept-based graph convolutional neural network to construct the relationships between key medical concepts, to obtain an interpretation consistent with a doctor's cognition. Extensive experiments have empirically shown that our model can meaningfully explain the decision of the classifier and provide quantitative support.