Explaining Deep Learning Models for COVID-19 Detection with Grad-CAM and Novel Use of PCA

Abstract

Machine learning and more specifically deep learning has achieved remarkable results in a range of computer vision tasks such as classification. Despite this, their black-box nature means researchers are largely unable to explain and interpret the decisions these systems make. Researchers use various techniques to explain deep learning classification models, e.g. Class Activation Maps (CAM) and Gradient Weighted Class Activation Maps (Grad-CAM) which produce heat maps of the input image highlighting the regions that contribute most to the model's decision. In this paper we present a novel technique based on Principal Component Analysis (PCA) to explain deep learning model decisions at a higher level, with results similar to those produced by Grad-CAM. This technique is applied directly to our dataset of COVID-19 blood test images, and we compare the PCA results with Grad-CAM using the convolutional neural network model we developed using the same dataset. As the PCA is applied to the dataset directly, no deep learning model needs to be trained allowing for faster and simpler computation than techniques such as Grad-CAM while producing similar explanation results. The results indicated that the reconstructed PCA map using the first two principal components and Grad-CAM have a similarity score of 85.7% and 71.4% respectively for COVID-19 positive and negative images, with an average similarity score of 78.6%.