Pericardium Based Model Fusion of CT and Non-contrasted C-arm CT for Visual Guidance in Cardiac Interventions

Yefeng ZhengMICCAI 2014 · DOI

TL;DR

This paper introduces a novel method for fusing pre-operative CT scans with intra-operative C-arm CT scans to create accurate 3D cardiac models for guiding minimally invasive cardiac interventions. By using the pericardium as a robust anatomical landmark, the method achieves automatic and efficient model alignment, significantly improving visual guidance during procedures.

What Problem Does This Paper Address?

Minimally invasive cardiac interventions are becoming increasingly common, and precise visual guidance during these procedures is crucial. Pre-operative CT scans provide detailed anatomical information for planning, but directly overlaying these 3D models onto real-time 2D fluoroscopic images (3D-to-2D registration) is challenging. This direct approach often requires manual interaction, contrast agent injection (which carries risks), and can be difficult due to differences in imaging times, cardiac phases, and machine variations. Intra-operative C-arm CT offers a solution by providing 3D imaging on the same system as fluoroscopy, simplifying 3D-to-2D registration. However, obtaining high-quality, contrasted C-arm CT scans is often not feasible. This paper addresses the need for a robust and automatic method to bridge the gap between pre-operative CT and intra-operative, potentially non-contrasted, C-arm CT for accurate visual guidance.

Methods

The core of this paper's contribution lies in its pericardium-based model fusion approach. The key steps are:

  • Anchor Structure Identification: The pericardium, the outer sac of the heart, is identified as an ideal "anchor structure." It's consistently visible in both contrasted CT and non-contrasted C-arm CT, and its shape is highly correlated with inner cardiac anatomy.
  • Automatic Pericardium Segmentation: The authors adapt a Marginal Space Learning (MSL) method for efficient and fully automatic segmentation of the pericardium. This method is significantly faster (around 1.5s/volume) and more robust than previous approaches, working on both contrasted and non-contrasted scans. The segmented mesh points provide built-in anatomical correspondence.
  • Deformation Field Estimation: A thin-plate-spline (TPS) model is used to estimate the deformation field between the segmented pericardium meshes from the pre-operative CT and the intra-operative C-arm CT. TPS is chosen for its smooth interpolation, lack of manual tuning, closed-form solutions, and physical interpretation.
  • Cardiac Model Warping: The estimated deformation field is then applied to warp the 3D cardiac model (extracted from the pre-operative CT) to align with the C-arm CT space.
  • Intelligently Weighted Average Fusion: To further improve accuracy, an "intelligently weighted average" method is proposed. This method leverages a library of pre-operative models from multiple patients. The weight assigned to each model's prediction is based on the shape similarity (pericardium distance) between its pericardium and the target C-arm CT's pericardium. Models with more similar pericardial shapes receive higher weights.

Datasets

The paper mentions collecting sixteen CT and C-arm CT pairs from a clinical site for the TAVI (Transcatheter Aortic Valve Implantation) application. These datasets were used for quantitative evaluation. While the primary application is for non-ECG-gated, non-contrasted C-arm CT, the evaluation was performed on contrasted C-arm CT data where the aortic root was clearly visible, allowing for quantitative assessment.

Key Insights and Ideas

  • The Pericardium as a Robust Anchor: The pericardium's consistent visibility and anatomical correlation make it an excellent, reliable landmark for cross-modality registration, even with significant differences in image characteristics.
  • Efficiency of MSL for Segmentation: Adapting MSL for pericardium segmentation offers a substantial speedup (at least 10x faster) and robustness compared to prior methods, making it practical for intra-operative use.
  • TPS for Non-Rigid Deformation: The thin-plate-spline model effectively captures the non-rigid deformations that can occur between pre-operative and intra-operative scans, leading to more accurate model warping than simpler affine transformations.
  • Leveraging Multi-Patient Data: The "intelligently weighted average" approach demonstrates that pooling information from multiple patients, weighted by anatomical similarity, can yield more accurate fused models than relying solely on a single patient's data.

Key Results

  • Quantitative Accuracy: For aortic root fusion, the TPS transformation with a patient-specific model achieved a mean mesh error of 4.18 mm, significantly better than affine transformation (5.60 mm). Using a generic model from a different patient resulted in a mean error of 5.00 mm.
  • Best Performance: The proposed Intelligently Weighted Fusion method achieved the best quantitative result, with a mean error of 3.62 mm.
  • Speed: The automatic pericardium segmentation runs at approximately 1.5 seconds per volume, which is considered very fast for medical imaging applications.
  • Qualitative Results: Visual inspection of warped models (four-chamber heart and aortic root) on C-arm CT showed visually accurate predictions, demonstrating the method's potential for guiding interventions.

Limitations and Future Work

  • Contrast Agent for Evaluation: The quantitative evaluation was performed on contrasted C-arm CT data, which is not the target scenario (non-contrasted C-arm CT). The authors note that the reported accuracy might be more optimistic than in the real application.
  • Manual Correction for Ground Truth: For quantitative evaluation, the ground-truth aortic root mesh in C-arm CT required "little manual correction," which deviates from a fully automatic pipeline. Similarly, minor errors in pericardium segmentation were not corrected to mimic real clinical application.
  • Future Work: The authors plan to investigate the accuracy of projecting the aligned 3D model onto real-time fluoroscopy images. They also suggest that minor local adjustments can further compensate for cardiac motion and respiratory motion observed in real-time fluoroscopy.

Summary updated

BibTeX
@inproceedings{zheng2014pericardium,
  author = {Yefeng Zheng},
  title = {{Pericardium Based Model Fusion of CT and Non-contrasted C-arm CT for Visual Guidance in Cardiac Interventions}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2014},
  doi = {10.1007/978-3-319-10470-6\_87},
}
Figure