Semi-URF: Progressive Uncertainty-Aware Region Filtering and Fusion for Semi-Supervised Medical Image Segmentation.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• Semi-supervised learning for medical image segmentation often discards uncertain unlabeled data, limiting learning from complex regions crucial for clinical accuracy.
• This conservative approach hinders the development of robust AI models by avoiding challenging areas.

Purpose of the Study:

• To introduce Semi-URF, an Uncertainty-Aware Region Filtering and Fusion framework, designed to exploit uncertainty in unlabeled data as a supervisory signal.
• To improve the learning from challenging regions in medical image segmentation, thereby reducing the need for extensive expert annotations.

Main Methods:

• Uncertainty Distribution Adaptive Thresholding (UDAT) adaptively filters reliable and unreliable regions from unlabeled data.
• Bidirectional Uncertainty-Consistent Exchange (BUCE) enables learning from unreliable regions by exchanging data patches between labeled and unlabeled sets.
• Frequency-Enhanced Feature Fusion (FEFF) module utilizes wavelet transforms and cross-attention to enhance boundary detail perception.

Main Results:

• Semi-URF demonstrates superior performance compared to state-of-the-art methods, particularly when annotated data is scarce.
• The framework effectively utilizes informative pixels from unlabeled data by adaptively managing uncertainty.
• Significant reduction in reliance on costly expert annotations for medical image segmentation tasks.

Conclusions:

• Semi-URF transforms data uncertainty from a limitation into a valuable learning resource for semi-supervised medical image segmentation.
• The proposed framework offers a more efficient and cost-effective approach to developing AI models for clinical applications.
• This method holds promise for advancing AI in healthcare by improving segmentation accuracy with reduced annotation burden.