Abstract
Self-supervised learning enables the creation of algorithms that outperform supervised pre-training methods in numerous computer vision tasks. This paper provides a comprehensive overview of self-supervised learning applications across various X-ray modalities, including conventional X-ray, computed tomography, mammography, and dental X-ray. Apart from the application of self-supervised learning in the interpretation phase of X-ray images, the paper also emphasizes the critical role of self-supervised learning integration in the preprocessing and archiving phase. Furthermore, the paper explores the application of self-supervised learning in multi-modal scenarios, which represents a key future direction in developing machine learning-based applications across the field of medicine. Lastly, the paper addresses the main challenges associated with the development of self-supervised learning applications tailored for X-ray modalities. The findings from the reviewed literature strongly suggest that the self-supervised learning approach has the potential to be a “game-changer”, enabling the elimination of the current situation where many machine learning-based systems are developed but few are deployed in daily clinical practice.
1. Introduction
Radiology has been established for over 125 years and has significantly reduced mortality rates from various diseases such as pneumonia, cancer, coronary heart disease, and nonfatal myocardial infarction (Howell 2011; The National Lung Screening Trial Research Team 2011; Wake et al. 2011; Crummy et al. 2018; Imai et al. 2018). Currently, radiologists are faced with the challenge of interpreting a vast amount of imaging data. The immense task of interpreting medical images leads to radiologists experiencing fatigue and burnout, consequently raising the probability of medical errors (Bercovich and Javitt 2018). To address these challenges, there is an urge to develop algorithms. Implementing automated medical image analysis offers several advantages, including enhanced sensitivity for subtle findings, prioritization of time-sensitive cases, automation of routine tasks, and alleviating the scarcity of radiologists in remote areas and developing countries (Çallı et al. 2021).
The deployment of machine learning (ML) in radiology holds promise for improving many aspects of the radiology workflow (Pierre et al. 2023), potentially resulting in increased diagnostic accuracy and efficiency, optimized treatment plans, enhanced quality of care, and a potential reduction in healthcare-related expenditures. The currently developed deep learning-based systems for automated medical image interpretation have reached levels comparable to practicing radiologists’ performance in some tasks (Rajpurkar et al. 2017). Studies have also shown that using ML can assist physicians in identifying abnormalities in medical images more effectively (Leibig et al. 2022). Despite all these breakthroughs, several challenging aspects still need to be addressed. Deep learning relies on having access to a substantial amount of annotated data (Zhou et al. 2021). This issue is particularly problematic in medical applications because medical imaging datasets are significantly smaller (hundreds/thousands) compared to natural domain datasets (millions of samples). The annotation of medical imaging data adds another challenge as it requires a substantial amount of time, effort, and a team of experts (Mckinney et al. 2020). To tackle the issue of data scarcity, the transfer learning technique is a promising solution. This method involves pretraining systems on a large natural domain dataset and then employing them in the medical image domain. However, the domain gap between the two datasets is substantial because medical images have entirely different characteristics from natural images, such as low image quality, latent feature distribution, lower resolution, 3D form, and similar image content across images (Zhou et al. 2019; Zhuang et al. 2019; Li et al. 2021a). Data imbalance is another issue for deep learning algorithms. This imbalance is particularly noticeable in the context of rare diseases, where most of the dataset comprises “normal” data, while instances of the disease are statistically scarce. Additionally, the ML systems frequently struggle with the designated tasks when data are collected from different acquisition centers and devices, introducing variation within the medical imaging domain, such as signal-to-noise ratio. All these practical challenges critically stagnate the robustness and generalizability of transfer learning.
