CLEF: Contrastive Learning of Equivariant Features in CT Images

Ilya Kuleshov, Vera Soboleva, Mikhail Goncharov

Keywords: Self-supervised learning, Biomedical image segmentation, Contrastive learning

Abstract: This work focuses on developing a self-supervised method of pretraining on biomedical images. The pretrained models are then fine-tuned on a small labelled dataset. We show, that using contrastive learning along with an equivariance loss and a loss, designed by us to maximise the features' information, we manage to improve quality in comparison to a fully-supervised baseline. Our method of pretraining achieves an average dice score of 0.86, reducing the baseline error by 20\%.