Year: 2024
Type: conference-paper
Abstract: Medical image segmentation is crucial for medical image analysis, disease diagnosis, and treatment planning. Traditional methods often struggle due to the time and effort required for data labelling, making it costly and labour-intensive. To address this, self-supervised learning presents a promising alternative by leveraging unlabeled data for training. In this study, we explore the application of the Barlow Twins model, a self-supervised learning method, in medical image segmentation. We employ two pathology image datasets, consisting of both labelled and unlabeled data, and pre-train the model using image augmentations such as scaling and noise. The pre-trained parameters are then used to train a UNet model for segmentation tasks. Our findings demonstrate that pre-training with a combination of labelled and unlabeled data significantly improves the model's generalisation and feature expressiveness. Larger datasets used during pre-training further enhance segmentation accuracy, as shown by improved Dice coefficients. The results confirm that self-supervised learning, particularly with substantial amounts of unlabeled data, can lead to better segmentation performance. In conclusion, this study underscores the potential of self-supervised learning in medical image segmentation, especially when labelled data is scarce. Future research could focus on using larger unlabeled datasets and optimising model architecture to further improve segmentation outcomes.
Authors: Zhuofan Meng, Junbo Huo, Wenzhe Huang, Nikesh Bajaj, Pengfei Fan
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4409829919