Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Data-Driven Tissue- and Subject-Specific Elastic Regularization for Medical Image Registration

Anna Reithmeir, Lina Felsner, Rickmer Braren, Julia A. Schnabel, Veronika A. Zimmer

TL;DR

This work tackles intra-patient medical image registration by replacing global physics-based regularization with data-driven, tissue- and subject-specific elasticity parameters. It introduces two hypernetworks: a global model that learns tissue-aware Lamé parameters and a spatially adaptive model that applies per-voxel elasticity maps derived from segmentation, enabling registration with regionally appropriate biomechanical constraints. Elasticity parameters are optimized at test time via grid search guided by class-wise Dice scores, obviating the need to retrain the networks for new subjects. Across NLST, ACDC, and L2R-Lung datasets, subject-specific spatially adaptive regularization yields higher registration quality in most tissues and metrics, though some 2D folding increases and parameter interpretability remains a limitation.

Abstract

Physics-inspired regularization is desired for intra-patient image registration since it can effectively capture the biomechanical characteristics of anatomical structures. However, a major challenge lies in the reliance on physical parameters: Parameter estimations vary widely across the literature, and the physical properties themselves are inherently subject-specific. In this work, we introduce a novel data-driven method that leverages hypernetworks to learn the tissue-dependent elasticity parameters of an elastic regularizer. Notably, our approach facilitates the estimation of patient-specific parameters without the need to retrain the network. We evaluate our method on three publicly available 2D and 3D lung CT and cardiac MR datasets. We find that with our proposed subject-specific tissue-dependent regularization, a higher registration quality is achieved across all datasets compared to using a global regularizer. The code is available at https://github.com/compai-lab/2024-miccai-reithmeir.

Data-Driven Tissue- and Subject-Specific Elastic Regularization for Medical Image Registration

TL;DR

This work tackles intra-patient medical image registration by replacing global physics-based regularization with data-driven, tissue- and subject-specific elasticity parameters. It introduces two hypernetworks: a global model that learns tissue-aware Lamé parameters and a spatially adaptive model that applies per-voxel elasticity maps derived from segmentation, enabling registration with regionally appropriate biomechanical constraints. Elasticity parameters are optimized at test time via grid search guided by class-wise Dice scores, obviating the need to retrain the networks for new subjects. Across NLST, ACDC, and L2R-Lung datasets, subject-specific spatially adaptive regularization yields higher registration quality in most tissues and metrics, though some 2D folding increases and parameter interpretability remains a limitation.

Abstract

Physics-inspired regularization is desired for intra-patient image registration since it can effectively capture the biomechanical characteristics of anatomical structures. However, a major challenge lies in the reliance on physical parameters: Parameter estimations vary widely across the literature, and the physical properties themselves are inherently subject-specific. In this work, we introduce a novel data-driven method that leverages hypernetworks to learn the tissue-dependent elasticity parameters of an elastic regularizer. Notably, our approach facilitates the estimation of patient-specific parameters without the need to retrain the network. We evaluate our method on three publicly available 2D and 3D lung CT and cardiac MR datasets. We find that with our proposed subject-specific tissue-dependent regularization, a higher registration quality is achieved across all datasets compared to using a global regularizer. The code is available at https://github.com/compai-lab/2024-miccai-reithmeir.
Paper Structure (12 sections, 1 equation, 5 figures, 1 table)

This paper contains 12 sections, 1 equation, 5 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methods
  3. Training of Global and Spatially Adaptive Networks
  4. Global Model:
  5. Spatially adaptive Model:
  6. Elasticity Parameter Identification
  7. Registration with Spatially Adaptive Elastic Regularization
  8. Experiments and Results
  9. Implementation Details and Datasets
  10. Analysis of Subject-Specific Elasticity Parameters
  11. Comparison of Regularization Methods
  12. Discussion and Conclusion

Figures (5)

  • Figure 1: Overview of our method. We train a globally and spatially adaptive network. After training, the optimal tissue-specific elasticity parameters are estimated with the global network. The spatially adaptive network then predicts the deformation field for the registration of a new image pair that follows the physical properties specified by the parameter values.
  • Figure 2: The subject-specific tissue-wise regularization parameters $\lambda,\mu$ that have been estimated with the global network for the NLST dataset. Each blue dot corresponds to one test sample. The results show that the optimal parameters vary between tissue types and subjects.
  • Figure 3: Qualitative results for the NLST (top) and ACDC (bottom) datasets. The left column shows the moving (M) and fixed (F) images. From left to right: Warped moving image, predicted deformation, Jacobian determinant (blue: positive values, red: negative values), difference image between fixed and warped image, segmentation map boundaries. The results show that the subject-specific spatially adaptive regularization (elas-sas) leads to more plausible deformation than the global regularization (elas-g).
  • Figure 4: The subject-specific tissue-wise regularization parameters $\lambda,\mu$ for the ACDC and L2R-Lung datasets that are estimated with the trained global model. In Fig. \ref{['fig:first']}, the classes correspond to the right ventricle (RV), myocardium (LV-Myo), and left ventricle blood pool (LV-BP).
  • Figure 5: Further visual registration examples of the NLST (top) and ACDC (bottom) test dataset. The left column shows the moving (M) and fixed (F) images. From left to right: Warped moving image, predicted deformation, Jacobian determinant (blue: positive values, red: negative values), difference image between fixed and warped image, segmentation map boundaries.