Learning a Single Network for Robust Medical Image Segmentation With Noisy Labels.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Robust segmentation with noisy labels is crucial in medical imaging due to annotation challenges.
• Existing methods often require multiple networks and are inflexible, failing to address coarse boundary labels.

Purpose of the Study:

• To propose a novel framework, SEAMAL, for robust segmentation with noisy labels.
• To achieve single-network robustness for both 2D and 3D medical image segmentation.
• To address the coarse boundary label problem for improved segmentation precision.

Main Methods:

• Developed a Simultaneous Edge Alignment and Memory-Assisted Learning (SEAMAL) framework.
• Introduced a Memory-assisted Selection and Correction (MSC) module for pixel-wise label reliability assessment.
• Incorporated an Edge Detection Branch (EDB) with a thinning loss for precise boundary learning.

Main Results:

• SEAMAL demonstrates robust learning for segmentation with noisy labels.
• The framework achieves applicability in both Set-HQ-knowable and Set-HQ-agnostic scenarios.
• SEAMAL significantly outperforms previous methods in extensive experiments.

Conclusions:

• SEAMAL offers a flexible, single-network solution for noisy-label medical image segmentation.
• The proposed methods effectively handle noisy labels and coarse boundary issues.
• SEAMAL represents a significant advancement in robust medical image segmentation.