ESSL-Polyp: A Robust Framework of Ensemble Semi-supervised Learning in Polyp Segmentation

We propose a robust framework called ESSL-Polyp combining ensemble and semi-supervised learning to improve polyp segmentation accuracy. The intuition starts from our previous experiments with semi-supervised learning on polyp segmentation. Following that, the semi-supervised models usually generalize better than supervised models with the same amount of training data, especially in out-of-domain datasets. In this paper, instead of using all labeled data, we split it into k-fold sub-datasets with labeled and unlabeled parts to train corresponding semi-supervised models. The ensemble of semi-supervised models is utilized to generate final precise predictions. We achieve an average of 0.8557 Dice score on five popular benchmark datasets, including Kvarsir, CVC-ClinicDB, ETIS-LaribPolypDB, CVC-ColonDB, and CVC-300. Meanwhile, the supervised baseline using the same training dataset only has an average Dice score of 0.8264. Our method especially yields superior performance compared to the supervised approach in out-of-domain datasets such as ETIS-LaribPolypDB, CVC-ColonDB, and CVC-300. The source code and pre-trained models are available at https://sal.vn/essl-polyp .