Abstract
This paper introduces a novel framework for "slice-to-volume" registration that simultaneously identifies the correct plane within a 3D volume and accounts for non-rigid, elastic deformations. By formulating the problem as a discrete Markov Random Field (MRF) with a 5-dimensional label space, the authors achieve sub-millimeter accuracy in monomodal scenarios and robust performance in challenging multimodal (Ultrasound to MRI) tasks.
Introduction
In image-guided surgery, clinicians frequently need to map real-time 2D "slices" (e.g., Ultrasound) onto high-resolution 3D preoperative scans (e.g., CT or MRI). This 2D-3D registration presents a dual challenge: finding the precise rigid alignment (the "Where") and accounting for non-rigid elastic deformations caused by physiological processes like breathing or surgical intervention (the "Squish"). Most existing methods address these sequentially or assume a rough initial alignment. This paper proposes a method to tackle both simultaneously, which is critical for accuracy in dynamic surgical environments.
Methods
The proposed framework utilizes a locally over-parametrized graphical model based on Markov Random Fields (MRFs). Instead of a global search for plane parameters, the problem is decomposed by overlaying a grid of nodes on the 2D image. Each node is assigned a label from a 5-dimensional space consisting of 3D displacement ($d_x, d_y, d_z$) for local elastic deformation and 2D orientation ($\phi, \theta$) to define the local plane orientation.
To ensure physical plausibility, the model employs pairwise potentials:
- Distance Preservation ($F_1$): Regularizes neighboring nodes to maintain relative distances, preventing unrealistic image tearing.
- Plane Preservation ($F_2$): Penalizes nodes that deviate from their neighbors' planes, encouraging a quasi-planar grid while allowing for local flexibility.
Optimization is performed using the FastPD algorithm, which efficiently finds near-optimal solutions in complex, non-submodular energy landscapes.
Experimental Evaluation
The method was evaluated on two distinct datasets:
- Monomodal Heart MRI: Temporal 3D MRI volumes where 2D slices were extracted and subjected to synthetic noise and transformations.
- Multimodal Brain MRI/US: Utilizing the MNI BITE database, this task involved registering intraoperative Ultrasound to preoperative MRI. This required Mutual Information as a similarity metric due to the disparate intensity patterns of the two modalities.
Key Results
The framework demonstrated high accuracy and robustness across both test cases:
- Monomodal Accuracy: Average errors were less than 1mm for translation and less than 0.013 radians (~0.74°) for rotation, with DICE coefficients reaching 0.93.
- Multimodal Robustness: In brain registration, the method improved DICE scores by 0.05 and decreased Contour Mean Distance (CMD) by 0.4mm, maintaining DICE scores above 0.7 even with low-resolution Ultrasound data.
- Efficiency: Registration times ranged from 4 to 10 minutes, which, while not yet real-time, represents a significant improvement over many non-rigid registration frameworks.
Discussion and Conclusion
The primary innovation of this work is the localization of global parameters into a labeling problem solvable via standard MRF toolkits. The framework is metric-independent and avoids error propagation by solving for plane location and deformation in a single step.
Current limitations include the computational weight of the 5D label space and the approximation of coplanarity via pairwise terms. Future work will explore third-order potentials to more accurately enforce rigid plane structures and further optimize discrete sampling to achieve real-time performance for surgical applications.
Summary updated