Abstract
In a world withstanding the waves of a raging pandemic, respiratory disease detection from chest radiological images using machine-learning approaches has never been more important for a widely accessible and prompt initial diagnosis. A standard machine-learning disease detection workflow that takes an image as input and provides a diagnosis in return usually consists of four key components, namely input preprocessor, data irregularities (like class imbalance, missing and absent features, etc.) handler, classifier, and a decision explainer for better clarity. In this study, we investigate the impact of the three primary components of the disease-detection workflow leaving only the deep image classifier. We specifically aim to validate if the deep classifiers may significantly benefit from additional preprocessing and efficient handling of data irregularities in a disease-diagnosis workflow. To elaborate, we explore the applicability of seven traditional and deep preprocessing techniques along with four class imbalance handling approaches for a deep classifier, such as ResNet-50, in the task of respiratory disease detection from chest radiological images. While deep classifiers are more capable than their traditional counterparts, explaining their decision process is a significant challenge. Therefore, we also employ three gradient visualization algorithms to explain the decision of a deep classifier to understand how well each of them can highlight the key visual features of the different respiratory diseases.