A case study: impact of target surface mesh size and mesh quality on volume-to-surface registration performance in hepatic soft tissue navigation

Georges Hattab, Carina Riediger, Jürgen Weitz, Stefanie Speidel

Abstract

In laparoscopic liver surgery, accurate registration between pre-operative 3D models and intra-operative surfaces is essential for effective Augmented Reality (AR) guidance. This study investigates the minimum data requirements and mesh quality standards necessary for robust non-rigid registration using Physics-Based Shape Matching (PBSM). Our findings indicate that while a surface area of 16.5% is sufficient for high-quality alignment, the orientation of the captured surface relative to tissue deformation and the geometric integrity of the mesh are more critical factors than the sheer volume of data.

Introduction

Laparoscopic liver surgery presents a significant challenge: the loss of haptic feedback makes it difficult for surgeons to localize tumors or avoid critical vascular structures. Augmented Reality (AR) addresses this by overlaying pre-operative CT-derived 3D models onto the live surgical field. However, the liver is a highly deformable organ, shifting due to pneumoperitoneum and surgical manipulation. To maintain accuracy, "non-rigid registration" must be performed to warp the pre-operative volume to the intra-operative surface. This process is often hindered by noisy, incomplete intra-operative data. This paper quantifies the impact of surface area, orientation, and mesh noise on registration performance.

Methods and Evaluation Pipeline

The study utilizes Physics-Based Shape Matching (PBSM), an algorithm that treats the liver volume as an electrically charged object that aligns with an oppositely charged target surface, governed by Finite Element Method (FEM) physics.

The evaluation followed a three-step pipeline:

  1. Registration: Aligning the volume to the target using PBSM.
  2. Voxelization: Converting resulting meshes into a high-resolution $1024^3$ voxel grid. This step standardizes the comparison between meshes with varying vertex counts and topologies.
  3. Metric Computation: Assessing similarity via Hausdorff distance (maximum error), Jaccard index, and Mutual Information.

Experiments were designed to test three variables: Orientation (viewing angles), Mesh Quality (automatic cleaning of non-manifold geometry), and Surface Area Sensitivity (incremental reduction of visible surface).

Datasets

The evaluation relied on two distinct datasets to provide a comprehensive ground truth for comparison:

  • In Silico Dataset: A simulated deformed liver created through a non-linear biomechanical model.
  • Phantom Dataset: A physical silicone liver phantom with realistic elastic properties, where deformation was induced by a rigid sphere and tracked via CT and endoscopy.
Figure 1: Reference data sets used in the evaluation. The in silico data set (left) and phantom data set (right). The initial surface (blue) is shown in the same coordinate system as the deformed surface (white).

Key Insights

The "16.5% Rule" and Orientation

The study identified a critical threshold for data efficiency: once the captured surface reached approximately 16.5% of the total organ surface, additional data provided diminishing returns in accuracy. More importantly, the orientation of this surface is paramount. Surfaces aligned with the primary axis of deformation provided significantly better registration cues. If the camera view misses the area of maximum warping, registration accuracy drops precipitously.

Mesh Quality and Geometric Integrity

Intra-operative meshes frequently contain "non-manifold" errors—geometries that are physically impossible. The researchers demonstrated that automatically "cleaning" these errors significantly improved registration metrics. Specifically, advanced post-processing improved the Jaccard index by 6.9% and Mutual Information by 9.8%, proving that data "sanity" is as important as data quantity.

Performance Benchmarks

PBSM demonstrated superior performance compared to traditional methods. It achieved a mean error of 0.48 mm, significantly outperforming the Coherent Point Drift (CPD) algorithm, which showed an error of 1.96 mm on the same datasets. Statistical analysis confirmed that the specific portion of the liver visible to the camera has a significant impact on registration success ($P < 2e-16$).

Limitations and Future Directions

While providing a roadmap for surgical navigation, the study notes several constraints:

  • Computational Cost: The high-resolution voxelization used for evaluation is currently too intensive for real-time intra-operative use.
  • Dynamic Environments: The study focused on static deformation, whereas real surgery involves continuous motion from respiration and heartbeat.
  • Clinical Translation: Future work must validate these minimum surface requirements in clinical settings, accounting for occlusions like blood, smoke, and varying lighting conditions.

The authors suggest that future surgical endoscopes could incorporate guidance systems to help surgeons capture the most informative 16.5% of the liver surface, ensuring stable and accurate AR overlays.

Summary updated

Abstract

Purpose Soft tissue deformation severely impacts the registration of pre- and intra-operative image data during computer-assisted navigation in laparoscopic liver surgery. However, quantifying the impact of target surface size, surface orientation, and mesh quality on non-rigid registration performance remains an open research question. This paper aims to uncover how these affect volume-to-surface registration performance. Methods To find such evidence, we design three experiments that are evaluated using a three-step pipeline: (1) volume-to-surface registration using the physics-based shape matching method or PBSM, (2) voxelization of the deformed surface to a $$1024^3$$ 1024 3 voxel grid, and (3) computation of similarity (e.g., mutual information), distance (i.e., Hausdorff distance), and classical metrics (i.e., mean squared error or MSE). Results Using the Hausdorff distance, we report a statistical significance for the different partial surfaces. We found that removing non-manifold geometry and noise improved registration performance, and a target surface size of only 16.5% was necessary. Conclusion By investigating three different factors and improving registration results, we defined a generalizable evaluation pipeline and automatic post-processing strategies that were deemed helpful. All source code, reference data, models, and evaluation results are openly available for download: https://github.com/ghattab/EvalPBSM/ .
BibTeX
@article{hattab2020case,
  author = {Georges Hattab and Carina Riediger and Jürgen Weitz and Stefanie Speidel},
  title = {{A case study: impact of target surface mesh size and mesh quality on volume-to-surface registration performance in hepatic soft tissue navigation}},
  journal = {Int. J. Comput. Assist. Radiol. Surg},
  year = {2020},
  doi = {10.1007/s11548-020-02123-0},
}
Figure