An Improvement in Medical Imaging via Self-Supervised Representation Learning

Abstract

This paper proposes Self-Supervised COVIDNet (SSL-COVIDNet), which is a deep neural network designed for the Coronavirus disease (COVID-19) diagnosis from chest CT images based on self-supervised learning. The main ingredient of our model is Momentum Contrast (MoCo), which is a selfsupervised learning algorithm with a contrastive loss. Unlike the traditional approach to this task, SSL-COVIDNet is designed to pretrain more general image representations from unlabeled images. As a result, the pre-trained SSLCOVIDNet model can be fine-tuned for downstream tasks such as COVID-19 diagnosis with some linear layers without significant task-specific architecture modifications. On the COVIDx CT-2 dataset, which is a diverse dataset for COVID-19 diagnosis, our model achieves an accuracy of approximately 100% and 98% for training and evaluation during the pretraining phase. The results suggest the feasibility to use self-supervised learning as an effective technique to leverage existing unlabeled medical images, pretrain generalized models for medical imaging, and then fine-tune it for the task of desire.