Spatially decoupled reliable mutual learning for semi-supervised medical image segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Semi-Supervised Learning (SSL) in medical image segmentation faces challenges with spatial context overfitting and confirmation bias from noisy pseudo-labels.
• Existing SSL methods struggle to generalize effectively due to reliance on fixed background contexts and susceptibility to inaccurate pseudo-labels.

Purpose of the Study:

• To propose a robust framework, Spatially Decoupled Reliable Mutual Learning (SDRML), to address limitations in SSL for medical image segmentation.
• To enhance segmentation accuracy and robustness, particularly in scenarios with limited labeled data.

Main Methods:

• Introduced Spatially Decoupled Reliable Mutual Learning (SDRML) framework.
• Implemented a Spatial Decoupling strategy using translation consistency to focus on intrinsic anatomical features.
• Developed a Reliable Mutual Learning mechanism with a Confident Regional Cross-entropy loss to filter low-confidence pseudo-labels and mitigate confirmation bias.

Main Results:

• SDRML demonstrated significant performance improvements over state-of-the-art methods on ACDC (2D MRI), Left Atrium (3D MRI), and Pancreas-CT datasets.
• The framework showed superior robustness and segmentation accuracy in data-scarce conditions, performing well with as little as 10% labeled data.
• Experiments confirmed the effectiveness of spatial decoupling and reliable pseudo-label filtering.

Conclusions:

• SDRML effectively resolves spatial dependency and noise accumulation issues inherent in SSL for medical image segmentation.
• The proposed method offers a highly effective solution for medical image segmentation tasks requiring minimal annotations, leveraging spatial decoupling and noise filtering.