Semi-supervised liver segmentation based on local regions self-supervision

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Semi-supervised learning reduces annotation needs in medical image segmentation.
• Contrastive learning aids in utilizing unreliable predictions but can neglect anatomical structures.
• Integrating anatomical context is crucial for optimizing medical image segmentation.

Purpose of the Study:

• To propose a novel semi-supervised approach for liver segmentation using contrastive learning.
• To enhance the effectiveness of contrastive learning for medical image segmentation tasks.
• To leverage unlabeled data for improved liver segmentation accuracy.

Main Methods:

• Proposed a semi-supervised contrastive learning method with local regions self-supervision (LRS²).
• Utilized Shannon entropy to differentiate reliable and unreliable predictions, reducing representational dissimilarity within regional units.
• Introduced a dynamic reliability threshold and applied morphological operations (erosion, dilation) for refined sample selection.

Main Results:

• The LRS² method demonstrated satisfactory performance by effectively exploiting unreliable predictions.
• Achieved significant improvements in Dice coefficients compared to supervised VNet, with gains up to +6.11%.
• Reported Dice coefficients ranging from 93.31% to 95.12% across different labeled data proportions.

Conclusions:

• The proposed method successfully selects positive and negative samples from reliable regions, assigning anchor pixels in unreliable regions correctly.
• Incorporating anatomical structure through regional partitioning enhances the precision of sample information capture.
• Extensive experiments validate the effectiveness of the LRS² method for semi-supervised liver segmentation.