Slice-to-volume deformable registration: efficient one-shot consensus between plane selection and in-plane deformation

Enzo Ferrante, Vivien Fecamp, Nikos Paragios

Introduction

In the context of image-guided surgery and biopsies, real-time 2D imaging (such as ultrasound) must be accurately registered to high-resolution pre-operative 3D volumes (such as CT or MRI). This process, known as slice-to-volume registration, is critical for providing clinicians with spatial context. However, it presents a significant computational challenge because it requires solving two problems simultaneously: identifying the correct 2D plane within the 3D volume (plane selection) and accounting for non-rigid tissue movement (deformable alignment).

Traditional approaches often suffer from "combinatorial explosions" when attempting to search high-dimensional spaces for both position and deformation, making them unsuitable for real-time clinical applications. This paper presents a novel decoupling strategy that achieves state-of-the-art accuracy with a 3.5x speedup over previous methods.

The Decomposed Graphical Model

The core innovation of this work is the use of a decomposed graphical model based on Markov Random Fields (MRFs). Rather than searching a massive 5D space, the problem is split into two lower-dimensional label spaces:

  1. The Plane Graph: This graph handles rigid parameters, including rotation and translation, to locate the optimal slice within the 3D volume.
  2. The Deformation Graph: This graph manages the non-rigid, in-plane "warping" required to align the 2D image with the selected 3D slice.
Schematic of the decomposed graphical model illustrating the mapping between the 3D search space and the 2D deformation plane.

These two graphs are interconnected; each control point on the registration grid is represented by nodes in both graphs that communicate to reach a consensus. This allows for a "one-shot" optimization process using Loopy Belief Propagation (LBP). By decoupling the search spaces, the computational complexity is reduced from the square of the product of the labels to the square of the largest label space, effectively reducing the search space size by approximately 10x.

Experimental Validation

The proposed framework was validated using two distinct datasets to test its versatility:

  • Cardiac MRI (Monomodal): 200 cases derived from 3D MRI sequences of a beating heart. Artificial noise and deformations were introduced to create a ground truth for accuracy testing.
  • Neuro-Ultrasound (Multimodal): Real-world data from the MNI BITE database, involving the fusion of pre-operative MRI volumes with intra-operative ultrasound images of brain tumors.

The framework is "metric-agnostic," allowing the use of Sum of Absolute Differences (SAD) for monomodal data and Mutual Information (MI) for multimodal data without altering the underlying algorithm.

Results and Performance

The decoupling strategy yielded significant improvements in both speed and efficiency:

  • Computational Speed: The method achieved a 3.5x speedup compared to previous MRF-based approaches. Processing times for the heart dataset dropped from 220 seconds to 60 seconds, while brain data registration was reduced from 10 minutes to 3 minutes.
  • Accuracy: High accuracy was maintained despite the increased speed. In the cardiac dataset, the mean rotation error was recorded at 0.0036 radians. In the brain dataset, the DICE coefficient—a metric for spatial overlap—showed significant improvement, indicating successful capture of complex tissue deformations.
  • Robustness: The model demonstrated resilience to image anisotropy, a common issue in medical imaging where resolution varies significantly across different axes.

Conclusion and Future Work

This paper demonstrates that decoupling plane selection from in-plane deformation significantly enhances the efficiency of slice-to-volume registration without sacrificing accuracy. While the current model is highly effective, it is primarily focused on in-plane deformations and remains dependent on a reasonable initial "guess" for the slice position.

Future research directions include incorporating biophysical modeling to better simulate organ-specific movements and utilizing machine learning to automatically optimize energy function parameters. These advancements could further improve the realism and robustness of registration in complex clinical environments.

Summary updated

Abstract

Purpose This paper introduces a novel decomposed graphical model to deal with slice-to-volume registration in the context of medical images and image-guided surgeries. Methods We present a new non-rigid slice-to-volume registration method whose main contribution is the ability to decouple the plane selection and the in-plane deformation parts of the transformation—through two distinct graphs—toward reducing the complexity of the model while being able to obtain simultaneously the solution for both of them. To this end, the plane selection process is expressed as a local graph-labeling problem endowed with planarity satisfaction constraints, which is then directly linked with the deformable part through the data registration likelihoods. The resulting model is modular with respect to the image metric, can cope with arbitrary in-plane regularization terms and inherits excellent properties in terms of computational efficiency. Results The proof of concept for the proposed formulation is done using cardiac MR sequences of a beating heart (an artificially generated 2D temporal sequence is extracted using real data with known ground truth) as well as multimodal brain images involving ultrasound and computed tomography images. We achieve state-of-the-art results while decreasing the computational time when we compare with another method based on similar techniques. Conclusions We confirm that graphical models and discrete optimization techniques are suitable to solve non-rigid slice-to-volume registration problems. Moreover, we show that decoupling the graphical model and labeling it using two lower-dimensional label spaces, we can achieve state-of-the-art results while substantially reducing the complexity of our method and moving the approach close to real clinical applications once considered in the context of modern parallel architectures.
BibTeX
@article{ferrante2015slice,
  author = {Enzo Ferrante and Vivien Fecamp and Nikos Paragios},
  title = {{Slice-to-volume deformable registration: efficient one-shot consensus between plane selection and in-plane deformation}},
  journal = {Int. J. Comput. Assist. Radiol. Surg},
  year = {2015},
  doi = {10.1007/s11548-015-1205-2},
}
Figure