Confidence Calibration of a Medical Imaging Classification System That is Robust to Label Noise

Area of Science:

• Artificial Intelligence
• Medical Imaging Analysis
• Machine Learning

Background:

• Neural network calibration is vital for medical applications, ensuring predicted probabilities align with actual outcomes.
• Existing calibration methods often require holdout data, which is challenging to obtain with reliable labels in medical imaging.
• Label noise in medical datasets can significantly impact model performance and clinical decision-making.

Purpose of the Study:

• To develop a robust network calibration procedure for neural networks that is resilient to label noise.
• To enable accurate probability predictions in medical analysis despite data imperfections.
• To improve the reliability of clinical decisions based on AI models.

Main Methods:

• A novel calibration method is proposed that directly addresses label noise during the training process.
• The method leverages the relationship between noisy and clean label confusion matrices.
• It involves estimating the noise level and using it to derive network accuracy for calibration.

Main Results:

• The proposed method achieves calibration performance comparable to methods using reliable labels, despite the presence of label noise.
• Demonstrated robustness in achieving accurate probability estimates even with imperfectly labeled medical image data.
• Successfully integrated noise estimation into a noise-robust training framework.

Conclusions:

• The developed calibration technique offers a practical solution for noisy medical image datasets.
• It enhances the trustworthiness of neural network predictions in critical medical applications.
• This approach paves the way for more reliable AI-driven diagnostics and treatment planning.