@Inbook{Verenich2021,
author="Verenich, Edward
and Murshed, M. G. Sarwar
and Khan, Nazar
and Velasquez, Alvaro
and Hussain, Faraz",
editor="Sayed-Mouchaweh, Moamar",
title="Mitigating the Class Overlap Problem in Discriminative Localization: COVID-19 and Pneumonia Case Study",
bookTitle="Explainable AI Within the Digital Transformation and Cyber Physical Systems: XAI Methods and Applications",
year="2021",
publisher="Springer International Publishing",
address="Cham",
pages="125--151",
abstract="Early detection and diagnosis of the SARS-CoV-2 virus which causes COVID-19 remains a difficult and time-sensitive process. Though progress has been made in the development and administration of tests, the time required to obtain a confident response can be prohibitive in areas of dense outbreaks. For these reasons, there has been great interest in adopting machine learning solutions to detect COVID-19 in prospective patients and aid clinicians in its localization. However, the adoption of deep learning approaches for these tasks is faced with multiple key challenges. In this work, we present a novel discriminative localization approach based on the computation of Class Activation Maps (CAMs) that addresses the problem of class overlap and imbalance within training data. In particular, we train a Convolutional Neural Network (CNN) on viral pneumonia and COVID-19 labels and compute the difference in scaled activations of the features in the last convolutional layer of the CNN for both classes. In doing so, we exploit the feature similarity between the two diseases for training purposes, thereby mitigating issues arising due to the scarcity of COVID-19 data. In models with overlapping classes, this can yield a neuron co-adaptation problem wherein the model is less confident in the outputs of similar classes, resulting in large CAMs with decreased localization certainty. Increasing the final mapping resolution of the last convolutional layer can reduce the overall area of such CAMs, but does not address class overlap, because CAMs computed for such classes exhibit significant spatial overlap, making it difficult to distinguish which spatial regions contribute the most to a specific target class. Our approach mitigates this by computing directed differences in scaled activations of pneumonia and COVID-19 CAMs, amplifying these differences, and projecting the results on to the original image in order to aid clinicians in the early detection of COVID-19. Specifically, our method clearly delineates spatial regions of an image that the classifier model considers as the most relevant for a given classification, even when standard CAMs for multiple classes within the model exhibit significant spatial overlap due to similarity in feature activations.",
isbn="978-3-030-76409-8",
doi="10.1007/978-3-030-76409-8_7",
url="https://doi.org/10.1007/978-3-030-76409-8_7"
}