UA-UNet: Uncertainty Aware Pseudo-Label Generation in Residual U-Net for Medical Image Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Supervised learning for medical image segmentation requires extensive annotated data, which is costly and time-consuming to acquire.
• Semi-supervised learning (SSL) utilizes both labeled and unlabeled data but often suffers from unreliable predictions due to ignored model uncertainty and noise.
• Existing SSL methods struggle with complex medical images and out-of-domain generalization.

Purpose of the Study:

• To introduce UA-UNet, an uncertainty-guided teacher-student model for robust multi-class medical image segmentation.
• To address limitations of current SSL methods, including disregarded model uncertainty, noise susceptibility, and poor adaptability to complex or out-of-domain data.

Main Methods:

• Developed UA-UNet, an uncertainty-guided teacher-student architecture based on Residual U-Net.
• Incorporated uncertainty estimation to guide high-quality pseudo-label generation from an ensemble of teacher models.
• Combined consistency regularization and pseudo-labeling to mitigate high-uncertainty regions and improve segmentation accuracy.

Main Results:

• UA-UNet achieved superior performance on the KiTS23 dataset, outperforming 10 other methods.
• The model obtained a Dice score of 0.901 and an IoU of 0.891.
• Generated uncertainty maps that enhance the interpretability of segmentation results.

Conclusions:

• UA-UNet offers a robust solution for semi-supervised medical image segmentation, particularly when labeled data is scarce.
• The uncertainty-guided approach improves segmentation accuracy and provides valuable interpretability through uncertainty maps.
• The model demonstrates strong performance and adaptability for complex medical imaging tasks.