PROJECT TITLE :

Mix-and-Interpolate A Training Strategy to Deal With Source-Biased Medical Data

ABSTRACT:

According to reports, the coronavirus disease 2019 (COVID-19) had infected more than 127 million people and been responsible for over 2.5 million deaths across the globe as of the 31st of March, 2021. It is essential to make a prompt diagnosis of COVID-19, both for the treatment of individual patients and for the containment of this extremely contagious disease. As a result of the realization of the clinical value of non-contrast chest computed tomography (CT) for the diagnosis of COVID-19, deep learning (DL) based automated methods have been proposed in order to assist radiologists in reading the massive quantities of CT exams that have been performed as a direct result of the pandemic. The problem of data source bias is one that is often overlooked when training deep convolutional neural networks for COVID-19 classification using real-world multi-source data. In this work, we address this problem and attempt to find a solution. The problem of data source bias refers to the circumstance in which certain sources of data only include a single class of data. Training with such source-biased data may cause DL models to learn to distinguish data sources rather than COVID-19. We propose the MIx-and-Interpolate (MINI) training strategy as a solution to this issue because it is conceptually straightforward, straightforward to implement, efficient, and effective. The proposed MINI approach generates volumes of the absent class by combining the samples collected from various hospitals, which enlarges the sample space of the original source-biased dataset. This results in the MINI approach generating volumes of the absent class. Experimental results on a large collection of real patient data (1,221 COVID-19 and 1,520 negative CT images, and the latter consisting of 786 community acquired pneumonia and 734 non-pneumonia) from eight hospitals and health institutions show that: 1) MINI can improve COVID-19 classification performance upon the baseline (which does not deal with the source bias), and 2) MINI is superior to competing methods in terms of the extent to which it improves classification performance.