COVID-19 diagnosis using model agnostic meta-learning on limited chest X-ray images

ABSTRACT

In the past year, detection of Coronavirus infection has demonstrated itself to be a challenging task. The gold standard for detection, real-time reverse transcription polymerase chain reaction (RT-PCR) testing, has a few shortcomings, including high false negative rates, long turn-around times, and limited availability. Applying machine learning for automatic analysis on chest X-rays can overcome these issues, but the limited amount of data with which to train inhibits development of robust deep neural networks. In this paper, we demonstrate the feasibility of performing few-shot learning to classify COVID-19 chest X-rays by utilizing a Model-Agnostic Meta-Learning (MAML) algorithm. We compare the improved variant of MAML, named MAML++, to other state-of-the-art machine learning strategies and demonstrate the robust and superior performance in classification accuracy. In addition, we explore the effect of the number of images made available to the sub-learners used for training MAML++ and show that increasing the number of images leads to diminishing returns in performance. Lastly, we compare MAML++ to the original MAML algorithm and discuss the shortcomings of MAML-based algorithms in classification problems. © 2021 ACM.