Semi-supervised medical image segmentation based on multi-stage iterative training and high-confidence...

Area of Science:

• Computer Vision
• Machine Learning
• Medical Image Analysis

Background:

• Semi-supervised learning is crucial for image segmentation due to high annotation costs.
• Existing methods often struggle with low-confidence pseudo-labels from perturbed inputs.
• Multi-branch co-training structures can lead to gradient interference and performance degradation.

Purpose of the Study:

• To develop a novel semi-supervised segmentation framework to enhance pseudo-label reliability.
• To address the challenges of low-confidence pseudo-labels and gradient interference in existing methods.
• To improve data utilization efficiency and model stability in medical image segmentation.

Main Methods:

• A multi-stage training strategy distinguishing labeled and unlabeled data training.
• A Balanced Uncertainty Adjustment Module (BUAM) for improved pseudo-label generation.
• Minimizing multi-branch gradient interference and adverse effects of input perturbations.

Main Results:

• The proposed framework significantly enhances pseudo-label reliability.
• The multi-stage approach reduces the negative impact of input perturbations and gradient interference.
• The framework demonstrates superior performance on ISIC and Cardiac MRI datasets, outperforming state-of-the-art methods.

Conclusions:

• The novel multi-stage semi-supervised segmentation framework offers a clear advantage over existing methods.
• The BUAM module effectively improves pseudo-label generation and data utilization.
• The approach enhances model stability and achieves state-of-the-art results in medical image segmentation.