Hard Sample Aware Noise Robust Learning for Histopathology Image Classification.

Area of Science:

• Medical Image Analysis
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Histopathology image classification using deep learning aids cancer diagnosis.
• Noisy labels in manual annotations can negatively impact model training and diagnostic accuracy.

Purpose of the Study:

• To develop a robust deep learning method for histopathology image classification that addresses the challenge of noisy labels.
• To improve the accuracy and reliability of automated cancer diagnosis in the presence of data imperfections.

Main Methods:

• Introduced a hard sample aware noise robust learning method.
• Developed an easy/hard/noisy (EHN) detection model based on sample training history.
• Integrated EHN into a self-training architecture for gradual label correction and proposed a noise suppressing and hard enhancing (NSHE) scheme.

Main Results:

• The proposed method effectively distinguishes informative hard samples from harmful noisy labels.
• Achieved superior performance compared to state-of-the-art methods on both synthetic and real-world noisy datasets.
• Demonstrated applicability to real-world noisy datasets without requiring a clean subset.

Conclusions:

• The novel noise robust learning method significantly enhances histopathology image classification accuracy.
• The approach offers a practical solution for real-world cancer diagnosis scenarios with inevitable label noise.
• The method preserves more clean samples and improves overall diagnostic reliability.