ST-LP: self-training and label propagation for semi-supervised classification

Due to the particularly high costs of manual data labeling, especially in the field of medical imaging, and the necessity for specialized knowledge, there is a growing interest in semi-supervised methods. In this paper, a novel framework of Self-Training with Label Propagation (ST-LP) is proposed for semi-supervised classification. It integrates self-training and label propagation to address the challenge of limited labeled data in classification tasks, a concern exacerbated by the especially expensive nature of data labeling in the medical domain. Our method involves leveraging two soft pseudo-labels generated from a pre-training fine-tuned model and label propagation scheme as inputs for a pseudo-label prediction module. Subsequently, confident predictions from this model are selected as pseudo-labeled data. The effectiveness of our approach is demonstrated through experiments conducted on diverse datasets, including the MNIST dataset and two medical classification datasets: ISIC2018 and MURA. Experimental results demonstrate that our method consistently achieved comparable or outstanding results when dealing with large amounts of unlabeled data.