Structural tensor and frequency guided semi-supervised segmentation for medical images.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Semi-supervised semantic segmentation reduces reliance on pixel-level annotations by using limited labeled and abundant unlabeled data.
• Existing methods often overlook object structural information and frequency-domain properties, focusing primarily on spatial augmentations.

Purpose of the Study:

• To investigate the utility of structural and frequency information for semi-supervised medical image segmentation.
• To enhance segmentation performance by integrating these novel data perspectives.

Main Methods:

• Introduced a novel structural tensor loss (STL) for spatial domain feature learning, enforcing object consistency.
• Proposed a frequency-domain alignment loss (FAL) to capture and align features across augmented samples in the frequency domain.

Main Results:

• The proposed method, integrating STL and FAL, demonstrated superior performance over state-of-the-art semi-supervised approaches.
• Experiments were conducted on diverse medical imaging datasets including MRI, CT, and ultrasound, showing significant improvements in Dice similarity coefficient.

Conclusions:

• The novel approach effectively improves semi-supervised medical image segmentation performance.
• This method has the potential to significantly decrease the demand for manual medical image labeling.