Region Uncertainty Estimation for Medical Image Segmentation With Noisy Labels.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning for 3D medical image segmentation requires extensive annotated data, which is costly and time-consuming to obtain.
• Foundation models like SAM offer sparse annotation capabilities but struggle with organs exhibiting blurred boundaries.
• Existing methods face challenges in robustly handling noisy labels in medical image segmentation tasks.

Purpose of the Study:

• To develop a region uncertainty estimation framework for Computed Tomography (CT) image segmentation using noisy labels.
• To mitigate the impact of imperfect annotations and reduce the cost of medical image annotation.
• To improve the performance of segmentation models in low-resource and remote scenarios.

Main Methods:

• A sample-stratified training strategy is proposed, prioritizing high-quality information at each stage.
• A boundary-guided regional uncertainty estimation module is designed to assess sample confidence.
• Sample-to-voxel level processing is employed to propagate reliable supervision information to noisy data.

Main Results:

• The proposed method demonstrates superior performance compared to existing approaches across multiple CT datasets under various noise conditions.
• The framework effectively mitigates the impact of noisy annotations on segmentation accuracy.
• Significant improvements in segmentation performance are observed in scenarios with imperfect annotations.

Conclusions:

• The developed framework offers a reliable label propagation strategy for medical image segmentation with noisy labels.
• This approach reduces annotation costs and enables robust model training, enhancing segmentation performance.
• The study paves the way for applying medical segmentation foundation models in resource-limited settings.