Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

NimbleReg: A light-weight deep-learning framework for diffeomorphic image registration

Antoine Legouhy, Ross Callaghan, Nolah Mazet, Vivien Julienne, Hojjat Azadbakht, Hui Zhang

TL;DR

NimbleReg addresses the computational burden of deep-learning-based non-linear image registration by representing anatomy as boundary surfaces of multiple regions and learning region-specific velocity fields with a shared PointNet backbone. It fuses these region-wise velocities into a global, diffeomorphic transformation using a log-Euclidean stationary velocity field framework via kernel convolution and Lie-group integration, ensuring diffeomorphism and invertibility. The approach is trained in a self-supervised manner with a Chamfer-based fit term and a simplex-aware velocity regularization, and evaluated on brain MRI where it matches state-of-the-art image-based methods while dramatically reducing memory usage. The method enables fast, cross-modality, segmentation-driven registration and offers a practical, scalable alternative to dense grid-based DL registration, with potential extensions to adaptive kernel bandwidth and higher-order integration.

Abstract

This paper presents NimbleReg, a light-weight deep-learning (DL) framework for diffeomorphic image registration leveraging surface representation of multiple segmented anatomical regions. Deep learning has revolutionized image registration but most methods typically rely on cumbersome gridded representations, leading to hardware-intensive models. Reliable fine-grained segmentations, that are now accessible at low cost, are often used to guide the alignment. Light-weight methods representing segmentations in terms of boundary surfaces have been proposed, but they lack mechanism to support the fusion of multiple regional mappings into an overall diffeomorphic transformation. Building on these advances, we propose a DL registration method capable of aligning surfaces from multiple segmented regions to generate an overall diffeomorphic transformation for the whole ambient space. The proposed model is light-weight thanks to a PointNet backbone. Diffeomoprhic properties are guaranteed by taking advantage of the stationary velocity field parametrization of diffeomorphisms. We demonstrate that this approach achieves alignment comparable to state-of-the-art DL-based registration techniques that consume images.

NimbleReg: A light-weight deep-learning framework for diffeomorphic image registration

TL;DR

NimbleReg addresses the computational burden of deep-learning-based non-linear image registration by representing anatomy as boundary surfaces of multiple regions and learning region-specific velocity fields with a shared PointNet backbone. It fuses these region-wise velocities into a global, diffeomorphic transformation using a log-Euclidean stationary velocity field framework via kernel convolution and Lie-group integration, ensuring diffeomorphism and invertibility. The approach is trained in a self-supervised manner with a Chamfer-based fit term and a simplex-aware velocity regularization, and evaluated on brain MRI where it matches state-of-the-art image-based methods while dramatically reducing memory usage. The method enables fast, cross-modality, segmentation-driven registration and offers a practical, scalable alternative to dense grid-based DL registration, with potential extensions to adaptive kernel bandwidth and higher-order integration.

Abstract

This paper presents NimbleReg, a light-weight deep-learning (DL) framework for diffeomorphic image registration leveraging surface representation of multiple segmented anatomical regions. Deep learning has revolutionized image registration but most methods typically rely on cumbersome gridded representations, leading to hardware-intensive models. Reliable fine-grained segmentations, that are now accessible at low cost, are often used to guide the alignment. Light-weight methods representing segmentations in terms of boundary surfaces have been proposed, but they lack mechanism to support the fusion of multiple regional mappings into an overall diffeomorphic transformation. Building on these advances, we propose a DL registration method capable of aligning surfaces from multiple segmented regions to generate an overall diffeomorphic transformation for the whole ambient space. The proposed model is light-weight thanks to a PointNet backbone. Diffeomoprhic properties are guaranteed by taking advantage of the stationary velocity field parametrization of diffeomorphisms. We demonstrate that this approach achieves alignment comparable to state-of-the-art DL-based registration techniques that consume images.

Paper Structure

This paper contains 14 sections, 1 equation, 3 figures.

Table of Contents

  1. Introduction
  2. Methods
  3. Training:
  4. Inference:
  5. Preprocessing
  6. Velocities estimation
  7. Integration into a diffeomorphic transformation
  8. Loss functions
  9. Experiments and results
  10. Evaluation strategies
  11. Data
  12. Implementation details
  13. Results
  14. Discussion and conclusion

Figures (3)

  • Figure 1: a) Synthetic segmentation and associated extracted surfaces. b) Velocities estimated by $f_\theta$ for each region and the associated SVF; c) Deformed points and grid after integration. d) Deformed points and grid without integration.
  • Figure 2: Diagrams for the proposed registration model at training (bottom-left) and at inference (right), and a zoom in on $f_\theta$'s architecture (top-left).
  • Figure 3: Quality of alignment metrics between moved and reference. Left: segmentation overlap (higher the better). Right: surface distance (lower the better). Lines are colored according to which datasets the images are from.