Hierarchical uncertainty estimation for learning-based registration in neuroimaging.

Area of Science:

• Medical Imaging
• Neuroimaging
• Artificial Intelligence

Background:

• Deep learning-based image registration shows high accuracy in medical imaging, especially for human neuroimaging using MRI.
• Current uncertainty estimation methods, like Monte Carlo dropout, are generic and do not leverage domain-specific spatial modeling.

Purpose of the Study:

• To propose a principled framework for propagating uncertainties from spatial locations to global transformation models in image registration.
• To improve the accuracy of brain MRI scan registration by incorporating uncertainty-aware fitting.

Main Methods:

• Developed a framework to propagate epistemic or aleatoric uncertainties from local spatial estimates to global transformation models.
• Utilized a Gaussian distribution for local uncertainty modeling and explored hierarchical uncertainty propagation.
• Compared the proposed uncertainty estimates with Monte Carlo dropout against reference registration error.

Main Results:

• The proposed uncertainty estimates showed a much better correlation with registration error than Monte Carlo dropout.
• Uncertainty-aware fitting of transformations significantly enhanced the registration accuracy of brain MRI scans.
• Demonstrated the utility of sampling from the posterior distribution of transformations for downstream neuroimaging tasks.

Conclusions:

• The proposed method offers a principled and effective way to estimate and propagate uncertainties in deep learning-based image registration.
• This approach leads to improved registration accuracy and provides a foundation for uncertainty propagation in downstream neuroimaging applications.