Review of disease identification methods based on computed tomography imagery

Abstract

Methods and approaches to computational diagnosis of various pulmonary diseases via automated analysis of chest images performed with computed tomography were reviewed. Google Scholar database was searched with several queries focused on deep learning and machine learning chest computed tomography imagery analysis studies published during or after 2017. A collection of 39 papers was collected after screening the search results. The collection was split by publication date into two separate sets based on the date being prior to or after the start of the COVID-19 pandemic. Information about the size of the dataset used in the study, classification categories present in it, primary classification target, employed approaches and architectures, metrics used to judge the performance, and the values of those metrics were collected for each paper in the set of discovered studies. Full collected data, including the citation, on every paper was provided in two tables respective to their publication date being prior or after COVID-19. Popular methodologies with the best metrics were identified, outlined, and described. The selected methodologies were compared by their accuracies in various papers found during this study. The comparison table of the found accuracies was provided. A best-performing approach was selected based on the found accuracies. As of this review, ResNet, its variations, and the architectures built upon it have the most promising results, with VGG and Xception being close contenders. The complications with reviewing existing studies in the field are outlined, the most important of them being the diversity in the way that dataset size is described, as well as diversity in the metrics employed, making a comparison between many individual papers impossible or at least lowering the quality of such a comparison. Metrics commonly used to measure the performance of machine learning approaches used in the found studies are outlined and described. Further research direction is proposed, with an emphasis on multi-class classification, modularity, and disease progress prediction. This proposition is guided by finding that most of the studies found focus on single class classification. Additionally, almost none of the studies discuss disease progression, and almost all of the studies discuss rigid solutions which are hardly extendable for future diseases and other classification methods.