Robust T-Loss for medical image segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Medical image segmentation is crucial for diagnosis and treatment planning.
• Noisy labels in medical datasets are a common challenge, impacting model performance.
• Existing loss functions struggle with significant label noise.

Purpose of the Study:

• Introduce T-Loss, a novel loss function for robust medical image segmentation.
• Address the challenge of noisy masks in medical imaging datasets.
• Improve segmentation accuracy and resilience to annotation errors.

Main Methods:

• Derived T-Loss from the negative log-likelihood of the Student-t distribution.
• Utilized a single, dynamically optimized parameter to control sensitivity to noisy labels.
• Conducted extensive experiments on public skin lesion and lung segmentation datasets.
• Simulated various types of label noise to test robustness.

Main Results:

• T-Loss significantly outperformed traditional loss functions in Dice scores on skin lesion and lung segmentation tasks.
• Demonstrated remarkable resilience to simulated label noise, mimicking human annotation errors.
• Showcased the adaptive nature of the T-Loss parameter in preventing noisy memorization during training.

Conclusions:

• T-Loss offers a robust and effective solution for medical image segmentation with noisy labels.
• Its adaptive parameter control makes it a promising alternative for real-world medical imaging applications.
• The proposed method enhances segmentation accuracy and reliability in the presence of data imperfections.