Dual structure-aware image filterings for semi-supervised medical image segmentation.

Area of Science:

• Medical image analysis
• Computer vision
• Machine learning

Background:

• Semi-supervised image segmentation is crucial for leveraging unlabeled data.
• Existing methods often use image or model-level variations, but overlook medical image structure.
• Prior structure information in medical images remains underexplored in segmentation.

Purpose of the Study:

• To propose novel dual structure-aware image filterings (DSAIF) for image-level variations in semi-supervised medical image segmentation.
• To effectively utilize the inherent structure of medical images for improved segmentation.
• To alleviate confirmation bias and overfitting to noisy pseudo-labels.

Main Methods:

• Developed dual structure-aware image filterings (DSAIF) based on connected filtering and dual contrast invariant Max-tree and Min-tree representations.
• Proposed a novel connected filtering to remove topologically equivalent nodes without siblings in Max/Min-trees, preserving critical structures.
• Applied DSAIF to mutually supervised networks to decrease consensus on erroneous predictions for unlabeled images.

Main Results:

• The proposed DSAIF method effectively reduces consensus errors on unlabeled data for mutually supervised networks.
• This approach alleviates confirmation bias and overfitting to noisy pseudo-labels.
• Extensive experiments on three benchmark datasets show significant and consistent performance improvements over state-of-the-art methods.

Conclusions:

• Dual structure-aware image filterings (DSAIF) offer a novel and effective image-level variation strategy for semi-supervised medical image segmentation.
• The method successfully leverages prior structural information in medical images, outperforming existing approaches.
• The proposed technique enhances segmentation performance by mitigating common issues like confirmation bias.