Core Concepts
Self-supervised learning, particularly contrastive learning, can effectively learn useful feature representations from medical images without relying on scarce annotated data.
Abstract
The review investigates several state-of-the-art predictive and contrastive self-supervised learning (SSL) algorithms originally developed for natural images, as well as their adaptations and optimizations for medical image analysis tasks.
Key highlights:
Supervised deep learning on manually annotated data has seen significant progress in computer vision, but its application in medical image analysis is limited by the scarcity of high-quality annotated data.
Self-supervised learning (SSL) is an emerging solution to address this challenge, with contrastive SSL being the most successful approach.
Predictive learning tasks, such as relative position prediction, solving jigsaw puzzles, and rotation prediction, can learn structural and contextual semantics from medical images.
Contrastive SSL methods like context-instance contrast, instance-instance contrast, and temporal contrast can effectively learn useful feature representations from medical images without relying on labeled data.
The review discusses the methodologies of these predictive and contrastive SSL approaches, their adaptations for medical image analysis, and the current limitations and future directions in this field.
Stats
"Medical image datasets often have fewer image samples despite large variability in the image visual attributes between them, e.g., the number of images in the medical image datasets varying from one thousand to one hundred thousand."
"Natural image datasets often have over 1 million images (e.g., ImageNet)."
Quotes
"SSL, as its name implies, creates supervisory information that is derived from the data itself."
"Contrastive learning encourages learning feature representation with inter-class separability and intra-class compactness, which can assist in classifier learning."