Self-Supervised Pretraining From Handcrafted Features for chest X-ray classification

Abstract

Modern convolutional neural networks require a large amount of human labeled data during training process. Prior work demonstrates that this problem can be addressed using self-supervised learning. This paper presents a novel self-supervised pretraining approach, which has been shown to be beneficial for the quality and stability of training process in case of domain-specific datasets with a small amount of labeled data