Evaluation of effectiveness of pre-training method in chest X-ray imaging using vision transformer

Abstract

ABSTRACT The limited availability of medical images is a major limitation when using deep learning, which requires large amounts of data to improve performance. To address this problem, transfer learning has become the de facto standard, using convolutional neural networks (CNNs) previously trained on natural images, and fine-tuned on medical images. Recently, vision transformers (ViT), which require large annotated medical images, have been studied from various perspectives. In this study, we investigated an effective pre-training method for binary classification of COVID-19 using chest radiography (CXR) images. Our results showed that pre-training on natural images outperformed CXR images using the fine-tuning method. Pre-training on natural images was also able to capture more global features. Furthermore, this trend increased as the number of pre-trained on natural images increased. In summary, these results suggest that the fine-tuning method with a large number of natural images as pre-training had the best discrimination performance.