Coarse-Refined Consistency Learning Using Pixel-Level Features for Semi-Supervised Medical Image Segmentation.

Area of Science:

• Medical image analysis
• Computer vision
• Machine learning

Background:

• Pixel-level annotations for medical image segmentation are costly and time-consuming.
• Semi-supervised learning (SSL) reduces annotation burden by using unlabeled data.
• Existing SSL methods often underutilize labeled data by not leveraging pixel-level information.

Purpose of the Study:

• To propose an innovative Coarse-Refined Network with pixel-wise Intra-patch ranked loss and patch-wise Inter-patch ranked loss (CRII-Net).
• To improve medical image segmentation accuracy, particularly with scarce labeled data.
• To enhance segmentation of challenging regions like blurred boundaries and low-contrast lesions.

Main Methods:

• Developed CRII-Net, incorporating a coarse-refined consistency constraint for stable targets.
• Extracted pixel-level and patch-level features to maximize labeled data utilization.
• Introduced Intra-Patch Ranked Loss (Intra-PRL) for object boundaries and Inter-Patch Ranked Loss (Inter-PRL) for low-contrast lesions.

Main Results:

• CRII-Net demonstrated superiority on two medical image segmentation tasks.
• Achieved at least a 7.49% improvement in Dice Similarity Coefficient (DSC) with only 4% labeled data compared to SOTA methods.
• Outperformed other methods in segmenting hard samples/regions, validated by quantitative and visualization results.

Conclusions:

• CRII-Net effectively addresses the challenge of limited labeled data in medical image segmentation.
• The proposed ranked loss functions significantly improve segmentation of complex and low-contrast regions.
• CRII-Net offers a promising solution for efficient and accurate medical image segmentation.