Semi-supervised image segmentation using a residual-driven mean teacher and an exponential Dice loss.

Area of Science:

• Computer Vision
• Medical Image Analysis
• Machine Learning

Background:

• Semi-supervised segmentation is crucial for reducing the need for extensive expert annotations in computer vision and medical imaging.
• Existing methods like the Mean Teacher (MT) framework provide a foundation but can be further optimized for performance.
• Accurate segmentation of structures like the Left Atrium (LA) and optic cup (OC) is vital for clinical diagnosis and research.

Purpose of the Study:

• To develop a novel residual-driven semi-supervised segmentation method (RDMT) to enhance the efficiency and accuracy of image segmentation.
• To improve the performance of the Mean Teacher (MT) framework through model-level residual perturbation and an exponential Dice (eDice) loss.
• To validate the RDMT method on challenging medical imaging datasets for Left Atrium and optic cup segmentation.

Main Methods:

• Introduced a novel model-level residual perturbation integrated into the exponential moving average (EMA) scheme of the Mean Teacher (MT) framework.
• Employed an exponential Dice (eDice) loss function to improve the network's sensitivity to object boundaries.
• Applied the RDMT method to segment 3D Left Atrium (LA) using V-Net and 2D optic cup (OC) using U-Net on public LASC and REFUGE datasets.

Main Results:

• The RDMT method achieved an average Dice score of 0.8776 for LA segmentation with 10% labeled data.
• The RDMT method achieved an average Dice score of 0.7751 for OC segmentation with 20% labeled data.
• RDMT significantly outperformed the baseline Mean Teacher (MT) and demonstrated competitive performance against other state-of-the-art semi-supervised methods.

Conclusions:

• The developed residual-driven semi-supervised segmentation method (RDMT) effectively enhances segmentation accuracy while minimizing the requirement for labeled data.
• The integration of residual perturbation and eDice loss offers a significant improvement over the standard Mean Teacher framework.
• RDMT shows strong potential for practical applications in medical image analysis, particularly for complex structures like the Left Atrium and optic cup.