SSL-CL-RUS - A Semi-Supervised Framework for Renal Ultrasound Segmentation in CKD: Combining Pseudo-Label Guided...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Nephrology

Background:

• Chronic kidney disease (CKD) poses a global health burden, requiring accurate diagnosis.
• Renal ultrasound is vital for CKD assessment, but image quality degradation in advanced stages hinders kidney structure segmentation.
• Current segmentation methods struggle with the reduced image quality in later CKD stages.

Purpose of the Study:

• To develop a novel semi-supervised segmentation framework for renal ultrasound images in CKD.
• To improve the accuracy of kidney structure segmentation, particularly in advanced CKD stages (G4 and G5).
• To leverage both labeled and unlabeled data for enhanced segmentation performance.

Main Methods:

• Proposed a semi-supervised segmentation framework integrating Pseudo-Label Guided Contrastive Loss (PLGCL) with the ST++ learning strategy.
• Utilized a combination of labeled and unlabeled ultrasound data for model training.
• Conducted extensive experimental evaluations against state-of-the-art techniques.

Main Results:

• The proposed framework significantly outperformed existing methods in kidney segmentation accuracy across all CKD stages.
• Achieved substantially higher Dice coefficients, with notable improvements in late-stage CKD (G4 and G5).
• Demonstrated the effectiveness of combining contrastive learning and pseudo-labeling for challenging ultrasound segmentation tasks.

Conclusions:

• The novel framework offers enhanced delineation of renal structures in ultrasound, crucial for advanced CKD.
• Improved segmentation fidelity supports more accurate disease progression assessments and clinical decision-making.
• This approach has the potential to improve patient outcomes through reliable kidney morphological evaluations.