Calibrating ensembles for scalable uncertainty quantification in deep learning-based medical image segmentation.

Area of Science:

• Medical Image Analysis
• Machine Learning
• Computer Vision

Background:

• Uncertainty quantification is crucial for reliable automated image analysis, especially with deep learning models.
• Current methods for uncertainty quantification in deep learning models are often computationally expensive and do not scale well.
• Existing approaches struggle to accurately approximate classification probabilities in complex, high-dimensional problems.

Purpose of the Study:

• To develop a scalable and intuitive framework for uncertainty quantification in medical image segmentation.
• To address the limitations of classical uncertainty quantification techniques in deep learning.
• To provide uncertainty measurements that approximate classification probabilities for improved model interpretability and utility.

Main Methods:

• Proposed a novel, scalable framework for uncertainty quantification in medical image segmentation.
• Demonstrated the failure of classical methods (e.g., dropout, ensembling) to approximate classification probabilities.
• Utilized k-fold cross-validation to eliminate the need for separate calibration datasets.

Main Results:

• The proposed framework yields uncertainty measurements that approximate classification probabilities.
• Classical uncertainty quantification approaches were shown to be inadequate for accurate probability approximation.
• K-fold cross-validation effectively replaced the need for held-out calibration data.

Conclusions:

• The developed framework offers a scalable and accurate solution for uncertainty quantification in medical image segmentation.
• The method supports enhanced active learning and human-machine collaboration by generating reliable pseudo-labels.
• This work advances the reliability and applicability of deep learning in medical imaging analysis.