Quantifying and leveraging predictive uncertainty for medical image assessment.

Area of Science:

• Medical Imaging Analysis
• Machine Learning in Healthcare
• Radiology

Background:

• Medical image interpretation faces challenges like artifacts, occlusions, and limited contrast.
• High inter-rater variability exists in chest radiography due to inconclusive data and subjective disease definitions.
• Current machine learning models often provide overconfident predictions with poor generalization on unseen medical data.

Purpose of the Study:

• To develop a system that learns explicit uncertainty measures alongside probabilistic predictions for medical image classification.
• To address the inherent ambiguity in medical images from various radiologic exams.
• To improve the reliability and generalization of machine learning models in medical image analysis.

Main Methods:

• Proposed a system that estimates both classification probability and an explicit uncertainty measure.
• Utilized uncertainty-driven bootstrapping to filter training data.
• Conducted experiments on diverse medical imaging datasets including chest radiographs, 2D ultrasound, and MRI.

Main Results:

• Sample rejection based on predicted uncertainty significantly improved ROC-AUC by 8% (to 0.91) for chest radiograph abnormality classification with <25% rejection rate.
• Uncertainty-driven bootstrapping increased model robustness and accuracy.
• A multi-reader study confirmed that predictive uncertainty correlates with reader errors.

Conclusions:

• Explicitly modeling uncertainty is crucial for handling ambiguity in medical image interpretation.
• The proposed system enhances diagnostic performance and robustness in machine learning-based medical image analysis.
• Predictive uncertainty can serve as an indicator of potential human errors in image interpretation.