Typicality- and instance-dependent label noise-combating: a novel framework for simulating and combating real-world...

Area of Science:

• Machine Learning
• Computer Vision
• Medical Image Analysis

Background:

• Neural networks struggle with real-world noisy labels, particularly instance-dependent label noise (IDN) common in medical imaging.
• Existing methods for instance-dependent label noise (IDN) do not account for sample typicality, limiting their effectiveness.
• Atypical medical samples often receive incorrect labels, necessitating noise-handling methods that consider sample characteristics.

Purpose of the Study:

• To introduce a more realistic noise simulation model called typicality- and instance-dependent label noise (TIDN).
• To develop a novel framework designed to combat TIDN in neural network training.
• To improve the robustness and accuracy of models trained on noisy datasets, especially in medical image classification.

Main Methods:

• Developed a typicality- and instance-dependent label noise (TIDN) simulation by measuring sample distance to decision boundaries.
• Introduced a TIDN-attention module to learn the label transition matrix from latent ground truth to observed noisy labels.
• Proposed a recursive algorithm for correcting predictions using the learned transition matrix.

Main Results:

• Demonstrated that TIDN more accurately reflects real-world label noise compared to instance-independent label noise (IIN) and instance-dependent label noise (IDN).
• The proposed TIDN-combating framework achieved superior classification performance on datasets with simulated TIDN.
• The framework also showed improved performance on actual real-world noisy medical image data.

Conclusions:

• Typicality- and instance-dependent label noise (TIDN) provides a more realistic benchmark for evaluating label noise robustness.
• The developed TIDN-combating framework effectively addresses complex real-world label noise scenarios in medical image classification.
• This work advances the field of learning with noisy labels by proposing a more accurate noise model and a robust training strategy.