A Simple Self-labeling Method for Semi-supervised Medical Image Segmentation

Ye Zhu

Keywords: Semi-supervise learning Pseudo labels

Abstract:

Leveraging a few labeled images and a large number of unlabeled images is crucial for medical image segmentation since labeling the medical data can be very expensive and time-consumed. Therefore we introduce a naïve but simple method to utilize the massive unlabeled
medical images for better training. We first use all the labeled data to train a basic model, then use this pre-trained model to infer the unlabeled images to get pseudo-labels, and finally use all the obtained pseudo-labels and the original labels as the ground truth of all images, and retrain the model from scratch to acquire the final model. We believe this is a simple but effective way to utilize the massive number of unlabeled images and experiments were performed to evaluate such method.