Self-supervised representation learning for histopathological images using contrastive learning
Ekta Gupta, Varun Gupta · Computational Methods in Science and Technology · 2024
Artificial Intelligence (AI) is becoming popular in the field of histopathological imaging. There isn’t a lot of annotated medical image data, and it is time consuming task to get this data annotated. The self-supervised learning (SSL) is an upcoming approach used for the unlabelled data. This paper suggests an SSL-based approach to classify breast cancer histopathological images. Self-supervised contrastive learning is used to learn the representations from unlabelled datasets and then used the same representations for the classification tasks of histopathological images of the labelled datasets. Two publically aavailable breast cancer datasets are used in the proposed approach: BreakHis and BCHD. Our method obtained an accuracy of 95.96% for the BreakHis dataset and 91.4% for the BCHD dataset. These results show that self-supervised learning can help learn representations of images, even if there isn’t labelled data and also this approach can help to make AI models work better in medical imaging applications.