Reduction of Annotation Effort in Medical Image Analysis Based on Self-supervised Learning

Self-supervised learning is increasingly valued in medical imaging, with the hope of solving the problem of scarce high-quality labeled medical image datasets. This study proposes a self-supervised learning training framework for single-view and multi-view images, which combines multi-task and semi-supervised loss applied to medical image classification and detection problems. It can save more than 50% of labeled training data and significantly reduce the demand for medical images and labeled data. Meanwhile, the proposed method achieves high performance in the classifications.