Entropy‐guided contrastive learning for semi‐supervised medical image segmentation

Abstract

Abstract Accurately segmenting medical images is a critical step in clinical diagnosis and developing patient‐specific treatment plans. While supervised learning algorithms have achieved excellent performance in this area, they require a large amount of annotated data, which is often time‐consuming and difficult to obtain. As a result, semi‐supervised learning (SSL) has gained attention as it has the potential to alleviate this challenge by using not only limited labelled data but also a large amount of unlabelled data. A common approach in SSL is to filter out high‐entropy features and use the low‐entropy part to compute unsupervised loss. However, it is believed that the high‐entropy part is also beneficial for model training, and discarding it can lead to information loss. To address this issue, a simple yet efficient contrastive learning approach is proposed in this work for semi‐supervised medical image segmentation, called Entropy‐Guided Contrastive Learning Segmentation Network (EGCL‐Net). The proposed method separates the low‐entropy and high‐entropy features via the average of predictions, using contrastive loss to pull the intra‐class entropy representation distance close and push the inter‐class entropy representation distance away. Extensive experiments on the automated cardiac diagnosis challenge dataset, COVID‐19, and BraTS2019 datasets showed that: (1) EGCL‐Net can significantly improve performance by utilizing high‐entropy representation, and (2) the authors’ EGCL‐Net outperforms recent state‐of‐the‐art semi‐supervised methods in both qualitative and quantitative evaluations.