URCA: Uncertainty-based region clipping algorithm for semi-supervised medical image segmentation.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Supervised learning for image segmentation requires extensive labeled data, which is costly and time-consuming in medical imaging.
• Semi-supervised methods show promise but struggle with low-confidence pseudo-labels and distribution bias between labeled and unlabeled data.

Purpose of the Study:

• To improve the accuracy of semi-supervised medical image segmentation models.
• To address the challenges of low-confidence pseudo-labels and distribution bias in semi-supervised learning.

Main Methods:

• Proposed an Uncertainty-based Region Clipping Algorithm for semi-supervised medical image segmentation.
• Utilized Monte Carlo Dropout for uncertainty estimation and employed diverse loss functions and Non-Maximum Suppression for model diversity.
• Introduced a novel module to generate new samples by masking low-confidence pixels, enhancing pseudo-label confidence and data distribution.

Main Results:

• Achieved superior performance over state-of-the-art methods on ACDC and BraTS2019 benchmarks.
• Attained an average Dice score of 87.86% and HD95 of 4.214 mm on the ACDC dataset.
• Reached an average Dice score of 84.79% for brain tumor segmentation with a HD score of 10.13 mm.

Conclusions:

• The proposed method significantly enhances the accuracy of semi-supervised medical image segmentation.
• Demonstrated the model's superiority on both 2D and 3D medical image datasets.
• Code is publicly available for reproducibility and further research.