Deep self-cleansing for medical image segmentation with noisy labels

Area of Science:

• Medical imaging and computer-assisted diagnosis.
• Deep learning applications in healthcare.
• Image processing and analysis.

Background:

• Medical image segmentation is vital for diagnosis and surgical planning.
• Supervised deep learning methods are common but sensitive to noisy labels.
• Inaccurate labels, like imprecise boundaries, hinder model performance.

Purpose of the Study:

• To develop a robust segmentation framework mitigating noisy label impact.
• To enhance segmentation accuracy by preserving clean and cleansing noisy labels.
• To improve the reliability of deep learning models in medical imaging.

Main Methods:

• Introduced a deep self-cleansing segmentation framework.
• Utilized a Gaussian Mixture Model (GMM)-based label filtering module (LFM) to identify noisy labels.
• Employed a label cleansing module (LCM) to generate pseudo low-noise labels for training.

Main Results:

• Achieved significant segmentation performance improvements on clinical liver tumor and public cardiac datasets.
• Demonstrated a +7.31% boost in B-model and +12.36% improvement in L-model segmentation accuracy.
• Effectively suppressed noisy label interference, enhancing target structure modeling and segmentation robustness.

Conclusions:

• The proposed framework offers a solution for noisy labels in medical image segmentation.
• Integration of LFM and LCM effectively preserves clean labels and generates pseudo low-noise labels.
• Validated approach shows potential for improving disease diagnosis and surgical planning.