Label Noise in Pathological Segmentation Is Overlooked, Leading to Potential Overestimation of Artificial...

Area of Science:

• Digital pathology
• Medical imaging
• Artificial intelligence

Background:

• Artificial intelligence (AI) significantly impacts medical imaging, with semantic segmentation vital for digital pathology.
• Pathology segmentation AI models rely on pathologist annotations, which can contain underexplored label noise.
• The effects of various label noise types on AI model training in pathology are not well understood.

Purpose of the Study:

• To evaluate the impact of artificial label noise on pathology segmentation AI models.
• To investigate the susceptibility of AI models to different types and intensities of label noise.
• To identify discrepancies between apparent and true performance scores due to label noise.

Main Methods:

• Combined a survey of public datasets with the synthesis of artificial label noise.
• Developed modules to simulate four types of artificial label noise at varying intensities.
• Trained deep learning models on datasets with simulated noise and evaluated their performance using clean test data.

Main Results:

• AI models demonstrated high susceptibility to overfitting label noise, particularly boundary-dependent noise (dilation, shrinkage).
• Significant discrepancies were observed between apparent and true performance scores, with boundary-altering noise causing the most pronounced overestimation.
• Random noise combinations further degraded model generalization capabilities.

Conclusions:

• Label noise critically impacts the reliability and generalizability of AI in digital pathology.
• Standardized methods for quantifying and mitigating label noise are needed.
• Developing robust benchmarks with noise-inclusive datasets and enhancing annotation quality are essential for clinical AI adoption.