An Explainable Deep Learning Method for Microwave Head Stroke Localization

Abstract

In this article, an explainable deep learning scheme is proposed to tackle microwave imaging for the task of multiple object localisation. Deep learning has been involved in solving microwave imaging tasks due to its strong pattern recognition capabilities. However, the lack of explainability of the model's predictions makes it infeasible to deploy deep learning models in practical applications such as stroke detection and localisation as the model is a black box, the confidence of the output is unknown as they cannot be verified. This article aims to alleviate this concern by applying the gradient-weighted class activation map (Grad-CAM), an explainable artificial intelligence technique, together with the Delay-Multiply-And-Sum (DMAS) algorithm to spatially explain the deep learning model. The Grad-CAM method highlights the important parts of the input signal for decision making and the important parts are mapped to the image domain to provide a more intuitive understanding of the model. This article concludes that the deep learning model learns from reliable information and provides outputs which have a physical basis.