G-T correcting: an improved training of image segmentation under noisy labels

Area of Science:

• Medical image analysis
• Machine learning in healthcare
• Computer vision

Background:

• Deep learning for medical image segmentation relies on expert annotations.
• Obtaining expert annotations is challenging and costly.
• Non-expert annotations introduce label noise, degrading segmentation performance.

Purpose of the Study:

• To enhance the segmentation performance of neural networks trained with noisy labels.
• To develop a robust framework for identifying and correcting inaccurate annotations in medical imaging datasets.

Main Methods:

• A two-stage framework, "G-T correcting," was proposed.
• The "G" stage uses a positive feedback method and Gaussian mixture model to recognize noisy labels via loss histograms.
• The "T" stage employs confident correcting and early learning strategies for label correction.

Main Results:

• The "G-T correcting" framework achieved over 90% accuracy in recognizing noisy labels.
• The method improved the network's DICE coefficient to 91% on simulated and real-world noisy data.
• Demonstrated significant enhancement in segmentation performance despite label noise.

Conclusions:

• The proposed "G-T correcting" method effectively addresses the challenge of noisy labels in medical image segmentation.
• This approach offers a promising solution for improving the reliability and accuracy of AI models in clinical settings.
• The framework shows good clinical application prospects for data-driven medical image analysis.