An Investigation of Feature-based Nonrigid Image Registration using Gaussian Process

Siming Bayer, Ute Spiske, Jie Luo, Tobias Geimer, William M. Wells Iii, Martin Ostermeier, Rebecca Fahrig, Arya Nabavi, Christoph Bert, Ilker Eyüpoglo, Andreas K. Maier

Abstract

Feature-based deformable registration (FDR) is essential for clinical applications like intraoperative image updates and adaptive treatment planning. Traditional FDR algorithms interpolate sparse feature correspondences to estimate dense displacement fields but often lack reliability metrics. This paper proposes treating the deformation field as a Gaussian Process (GP), enabling the simultaneous estimation of dense displacement and a corresponding uncertainty map. Evaluation across synthetic, phantom, and clinical datasets demonstrates that GP-based interpolation performs on par with state-of-the-art B-spline methods while providing critical mathematical uncertainty estimates that can guide clinical decision-making.

Introduction

In clinical environments such as neurosurgery or radiation therapy, anatomical structures frequently undergo nonrigid deformations—for instance, "brain shift" during surgery. Feature-based Deformable Registration (FDR) is commonly used to compensate for these shifts by identifying sparse landmarks and interpolating the dense displacement field. While standard methods like B-splines or Thin-Plate Splines are computationally efficient, they operate as "black boxes," offering no indication of local registration reliability. This study investigates Gaussian Processes (GP) as a framework to provide both accurate registration and a quantifiable "confidence score" for clinicians.

Methodology

The proposed approach models the deformation field as a Gaussian Process.

  1. Feature Correspondence: Sparse point matches between images serve as prior information for the valid deformation.
  2. GP Interpolation: A Squared Exponential (SE) kernel is utilized to interpolate the movement of every voxel based on these sparse points.
  3. Hyperparameter Optimization: The GP's performance is highly sensitive to its length-scale ($l$) and standard deviation ($\sigma$). Three strategies were compared:
  4. MEAN: Initialization based on mean displacement statistics.
  5. NML: Minimizing the Negative Log Marginal Likelihood.
  6. DGS (Discrete Grid Search): Optimizing parameters via a grid search to minimize Root Mean Squared Error (RMSE).
  7. Uncertainty Mapping: Unlike standard methods, the GP framework naturally yields a covariance matrix; its diagonal provides a spatial uncertainty map, indicating where the model's predictions are mathematically least certain.

Quantitative and Qualitative Evaluation

The study utilized synthetic brain models, a physical head phantom, and clinical data from brain surgery and brachytherapy to evaluate the GP framework against traditional B-spline interpolation.

Performance Comparison

Quantitative results indicate that GP-based interpolation, particularly using the DGS strategy, achieves accuracy comparable to industry-standard B-splines. As shown in Figure 2, metrics such as Error per Voxel, Modified Hausdorff Distance (MHD), and intensity differences ($|\Delta HU|$) for DGS are competitive with, and in some cases slightly outperform, B-spline results.

Quantitative comparison of registration performance across MEAN, NML, DGS, and B-Spline methods using Error per Voxel, MHD, and intensity difference metrics.

The Value of Uncertainty

The primary advantage of the GP approach is the generation of uncertainty maps. Visual analysis (Figure 6) reveals that while MEAN and NML methods often appear "overconfident" (showing high certainty even in areas with sparse data), the DGS method provides a more realistic representation of mathematical uncertainty. This spatial feedback allows clinicians to identify regions where the registration might be less reliable, a feature absent in traditional B-spline methods.

Qualitative comparison of uncertainty maps for different GP hyperparameter settings (a-c) and the resulting registration (d).

Conclusion

Gaussian Process-based registration offers a robust alternative to traditional interpolation methods. By matching the accuracy of B-splines while providing an integrated uncertainty map, GPs offer significant clinical potential for improving the safety and reliability of image-guided interventions. Future work will focus on optimizing computational efficiency for real-time applications and conducting clinician-led studies to assess the practical impact of uncertainty visualization on surgical outcomes.

Summary updated

Abstract

For a wide range of clinical applications, such as adaptive treatment planning or intraoperative image update, feature-based deformable registration (FDR) approaches are widely employed because of their simplicity and low computational complexity. FDR algorithms estimate a dense displacement field by interpolating a sparse field, which is given by the established correspondence between selected features. In this paper, we consider the deformation field as a Gaussian Process (GP), whereas the selected features are regarded as prior information on the valid deformations. Using GP, we are able to estimate the both dense displacement field and a corresponding uncertainty map at once. Furthermore, we evaluated the performance of different hyperparameter settings for squared exponential kernels with synthetic, phantom and clinical data respectively. The quantitative comparison shows, GP-based interpolation has performance on par with state-of-the-art B-spline interpolation. The greatest clinical benefit of GP-based interpolation is that it gives a reliable estimate of the mathematical uncertainty of the calculated dense displacement map.
BibTeX
@inproceedings{bayer2020investigation,
  author = {Siming Bayer and Ute Spiske and Jie Luo and Tobias Geimer and William M. Wells III and Martin Ostermeier and Rebecca Fahrig and Arya Nabavi and Christoph Bert and Ilker Eyüpoglo and Andreas K. Maier},
  title = {{An Investigation of Feature-based Nonrigid Image Registration using Gaussian Process}},
  booktitle = {Bildverarbeitung für die Medizin},
  year = {2020},
  url = {https://openreview.net/forum?id=fmAkBPzYfG},
}
Figure