COVID-CBR: A Deep Learning Architecture Featuring Case-Based Reasoning for Classification of COVID-19 from Chest X-Ray Images

Background and Objectives: This study aims to assist rapid accurate diagnosis of COVID-19 based on chest x-ray (CXR) images to provide supplementary information, leading to screening program for early detection of COVID-19 based on CXR images by developing an interpretable, robust and performant AI system. Methods: A case-based reasoning approach built upon autoencoder deep learning architecture is applied to classify COVID-19 from other non-COVID-19 as well as normal subjects from chest x-ray images. The system integrates the interpretation and decision-making together by producing a set of profiles that in appearance resemble the training samples and hence explain the outcome of classifications. Three classes are studied, which are COVID-19 (n=250), other non-COVID-19 diseases (NCD) (n=384), including TB and ARDS, and normal (n=327). Results: This COVID-CBR system sustains the average sensitivity and specificity of 93.1±3.58% and 96.1±4.10% respectively for classification of these three classes. In comparison with the current state of the art, including COVID-Net, VGG-16 and other explainable AI systems, the developed COVID-CBR system appears to perform similar or better when classifying multi-class categories. Conclusion: This paper presents a case-based reasoning deep learning system for detection of COVID19 from chest x-ray images. Comparison with several state of the art systems is conducted. Although the improvement tends to be marginal, especially for VGG-16, the novelty of this work manifests its interpretable feature building upon case-based reasoning, leading to revealing this viral insight and hence ascertaining more effective treatment and drugs while maintaining being transparent. Furthermore, different from several other current explainable networks that highlight key regions or the points of an input that activate the network, i.e. heat maps, this work is constructed upon whole training images, i.e. case-based, whereby each training image belongs to one of the case clusters.