Heteroscedastic Uncertainty Estimation Framework for Unsupervised Registration

Area of Science:

• Medical image analysis
• Deep learning
• Computer vision

Background:

• Unsupervised medical image registration commonly uses loss functions assuming uniform noise.
• Real-world medical images exhibit heteroscedastic noise, degrading registration accuracy.
• Existing methods are susceptible to noise-induced outliers, limiting performance.

Purpose of the Study:

• To develop a framework for heteroscedastic image uncertainty estimation in unsupervised registration.
• To adaptively reduce the influence of high-uncertainty regions during registration.
• To improve the accuracy and robustness of medical image registration.

Main Methods:

• Proposed a framework with collaborative training for displacement and variance estimators.
• Introduced a novel image fidelity weighting scheme using signal-to-noise ratios.
• Tested the framework on two registration architectures across three medical image datasets.

Main Results:

• The proposed method consistently outperformed baseline approaches.
• The framework produced sensible uncertainty estimates.
• Demonstrated consistent improvements in displacement estimation accuracy.

Conclusions:

• The developed framework effectively handles heteroscedastic noise in medical image registration.
• Adaptive uncertainty estimation improves registration accuracy by mitigating outlier influence.
• The approach offers a robust solution for unsupervised medical image registration challenges.