An Effective Semi-Supervised Approach for Liver CT Image Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning models for medical image segmentation require large annotated datasets, posing a significant bottleneck.
• Existing semi-supervised methods often struggle with generating high-quality pseudo-labels for effective data expansion.

Purpose of the Study:

• To develop a deep semi-supervised approach for liver CT image segmentation under a very low annotated-data paradigm.
• To enhance the quality of pseudo-labels for expanding training datasets in medical image segmentation.

Main Methods:

• A novel pseudo-labeling algorithm is proposed, utilizing mean operation on features from labeled data to generate class representations.
• Pseudo-labels for unlabeled data are created by measuring distances between feature vectors and class representations.
• Pseudo-label optimization techniques and a random patch strategy for unlabeled data are incorporated to improve segmentation accuracy.

Main Results:

• The proposed method significantly improves liver CT image segmentation accuracy in a low-data regime.
• Achieved more competitive results compared to existing semi-supervised methods on the LiTS dataset with limited labeled slices.
• Demonstrated the effectiveness of enhanced pseudo-labeling and data augmentation strategies.

Conclusions:

• The developed deep semi-supervised method effectively addresses the challenge of limited annotated data in medical image segmentation.
• The approach offers a promising solution for improving liver CT segmentation by generating higher-quality pseudo-labels.
• This work contributes to advancing semi-supervised learning techniques in the field of medical image analysis.