Exploring Inherent Consistency for Semi-Supervised Anatomical Structure Segmentation in Medical Imaging.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Semi-supervised learning is crucial for medical image segmentation due to high costs of labeled data.
• Existing methods primarily use image-level spatial consistency for regularization.
• Inherent anatomical properties of structures are often overlooked in current techniques.

Purpose of the Study:

• To introduce and evaluate a novel semi-supervised learning method incorporating inherent consistency for anatomical structure segmentation.
• To leverage latent representations capturing inherent anatomical properties for improved segmentation.
• To enhance the anatomical plausibility of segmentation results.

Main Methods:

• Projecting predictions and ground-truth into an embedding space to derive latent representations.
• Designing two inherent consistency constraints to align these latent representations.
• Integrating the plug-and-play method with existing semi-supervised segmentation techniques.

Main Results:

• The proposed method demonstrates good generalizability across three public datasets (ACDC, LA, Pancreas).
• Experimental results show superior performance compared to several state-of-the-art methods.
• The approach effectively improves segmentation performance and anatomical plausibility.

Conclusions:

• Incorporating inherent consistency is a promising direction for semi-supervised medical image segmentation.
• The proposed method offers a flexible and effective way to enhance segmentation accuracy and anatomical realism.
• This work provides a valuable contribution to advancing automated medical image analysis.