Learning robust medical image segmentation from multi-source annotations.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Multi-source annotations are common in medical image segmentation to reduce noise and bias.
• Learning from varied annotations presents challenges due to inherent uncertainties and variances.
• Existing methods struggle to effectively leverage multi-source data without compromising accuracy.

Purpose of the Study:

• To develop a robust method for training segmentation networks with multi-source annotations.
• To address the challenge of annotation uncertainty in medical image segmentation.
• To improve the generalization and reliability of segmentation models.

Main Methods:

• Proposed Uncertainty-guided Multi-source Annotation Network (UMA-Net) incorporating pixel and image-level uncertainty estimation.
• Developed an Annotation Uncertainty Estimation Module (AUEM) for pixel-wise uncertainty and weighted segmentation loss.
• Introduced a Quality Assessment Module (QAM) for image-level quality assessment and an auxiliary predictor for low-quality samples.

Main Results:

• UMA-Net demonstrated effectiveness and feasibility across multiple datasets (2D X-ray, fundus, 3D MRI, QUBIQ).
• The method successfully guided training using uncertainty estimation, improving segmentation performance.
• Preservation of knowledge from low-quality samples via an auxiliary predictor without impacting the primary predictor.

Conclusions:

• UMA-Net offers a robust solution for learning from multi-source annotations in medical image segmentation.
• The uncertainty-guided approach effectively mitigates challenges posed by annotation variance.
• The proposed method shows significant potential for enhancing medical image segmentation tasks.