Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Regression-based Pelvic Pose Initialization for Fast and Robust 2D/3D Pelvis Registration

Yehyun Suh, J. Ryan Martin, Daniel Moyer

TL;DR

This work addresses the initialization sensitivity of optimization-based 2D/3D pelvic registration by introducing a regression-based initializer that provides a coarse but close-enough pose $\theta_0$ to start the optimization. The method integrates with differentiable DRR-based registration and evaluates three pose estimators, showing consistent reductions in iteration counts and improvements in rotational and translational RMSE/MAE across varying pose ranges. Key contributions include the design of three initializer variants using a ResNet-18 backbone trained with simulated poses and demonstrated robustness to pose variation, as well as a thorough experimental comparison. The approach has practical implications for faster, more reliable clinical pelvic registration, though limitations such as depth ambiguity, small dataset, and manual preprocessing point to avenues for future validation and automation.

Abstract

This paper presents an approach for improving 2D/3D pelvis registration in optimization-based pose estimators using a learned initialization function. Current methods often fail to converge to the optimal solution when initialized naively. We find that even a coarse initializer greatly improves pose estimator accuracy, and improves overall computational efficiency. This approach proves to be effective also in challenging cases under more extreme pose variation. Experimental validation demonstrates that our method consistently achieves robust and accurate registration, enhancing the reliability of 2D/3D registration for clinical applications.

Regression-based Pelvic Pose Initialization for Fast and Robust 2D/3D Pelvis Registration

TL;DR

This work addresses the initialization sensitivity of optimization-based 2D/3D pelvic registration by introducing a regression-based initializer that provides a coarse but close-enough pose to start the optimization. The method integrates with differentiable DRR-based registration and evaluates three pose estimators, showing consistent reductions in iteration counts and improvements in rotational and translational RMSE/MAE across varying pose ranges. Key contributions include the design of three initializer variants using a ResNet-18 backbone trained with simulated poses and demonstrated robustness to pose variation, as well as a thorough experimental comparison. The approach has practical implications for faster, more reliable clinical pelvic registration, though limitations such as depth ambiguity, small dataset, and manual preprocessing point to avenues for future validation and automation.

Abstract

This paper presents an approach for improving 2D/3D pelvis registration in optimization-based pose estimators using a learned initialization function. Current methods often fail to converge to the optimal solution when initialized naively. We find that even a coarse initializer greatly improves pose estimator accuracy, and improves overall computational efficiency. This approach proves to be effective also in challenging cases under more extreme pose variation. Experimental validation demonstrates that our method consistently achieves robust and accurate registration, enhancing the reliability of 2D/3D registration for clinical applications.

Paper Structure

This paper contains 11 sections, 4 equations, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Preliminaries and Optimization-based Pose Estimation
  5. Regression and Registration
  6. Experiments
  7. Data and Experimental Setting
  8. Models, Training, and Evaluation
  9. Results
  10. Discussion
  11. Conclusion

Figures (3)

  • Figure 1: This figure diagrams the iterative registration framework; our initializer is at left, agnostic to the optimization framework at center and right.
  • Figure 2: Box plots of the difference in RMSE between the (Original) and (Proposed,1) methods for both rotation and translation under different experimental conditions. Statistical significance is marked with based on single tail t-tests, * denotes $p < 0.0001$.
  • Figure 3: Comparison of initialization and registration results using ProST on 2 sample cases using the original (naive) initialization and the proposed initialization. The third column shows difference maps (moving minus target, red:positive, blue:negative).