Semi-supervised CT image segmentation via contrastive learning based on entropy constraints.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning methods for computed tomography (CT) segmentation require extensive labeled data, which is costly and time-consuming to acquire.
• Existing semi-supervised methods struggle with unreliable pseudo-labels, potentially degrading model performance.

Purpose of the Study:

• To develop an effective semi-supervised learning method for CT image segmentation that leverages unlabeled data.
• To improve the accuracy and efficiency of medical image segmentation by addressing the limitations of labeled data scarcity.

Main Methods:

• A novel semi-supervised network model integrating Convolutional Neural Networks (CNNs) and Transformers for capturing local and global image features.
• Implementation of an entropy-constrained contrastive learning loss to effectively utilize unlabeled data while discarding unreliable samples.
• Inclusion of a residual squeeze and excitation module in the student network to enhance feature representation.

Main Results:

• The proposed method demonstrated superior performance on the COVID-19 CT dataset compared to state-of-the-art semi-supervised techniques.
• Achieved a 2.3% improvement in Dice Similarity Coefficient (DSC) and a 2.5% improvement in Jaccard Index (JC).
• Reduced the Hausdorff Distance 95th percentile (HD95) by 1.9 mm, indicating improved boundary accuracy.

Conclusions:

• The developed semi-supervised approach effectively enhances CT image segmentation by fusing CNN and Transformer architectures with entropy-constrained contrastive learning.
• This method significantly improves the utilization of unlabeled medical images, offering a more efficient alternative to fully supervised approaches.