Abstract
C-arm CT is an emerging imaging technique in transcatheter aortic valve implantation (TAVI) surgery. Automatic aorta segmentation and valve landmark detection in a C-arm CT volume has important applications in TAVI by providing valuable 3D measurements for surgery planning. Overlaying 3D segmentation onto 2D real-time fluoroscopic images also provides critical visual guidance during the surgery. In this paper, we present a part-based aorta segmentation approach, which can handle aorta structure variation in case that the aortic arch and descending aorta are missing in the volume. The whole aorta model is split into four parts: aortic root, ascending aorta, aortic arch, and descending aorta. Discriminative learning is applied to train a detector for each part separately to exploit the rich domain knowledge embedded in an expert-annotated dataset. Eight important aortic valve landmarks (three aortic hinge points, three commissure points, and two coronary ostia) are also detected automatically in our system. Under the guidance of the detected landmarks, the physicians can deploy the prosthetic valve properly. Our approach is robust under variations of contrast agent. Taking about 1.4 seconds to process one volume, it is also computationally efficient.
Introduction
Transcatheter Aortic Valve Implantation (TAVI) is a life-saving, minimally invasive procedure for patients with aortic valve disease who are too high-risk for open-chest surgery. To succeed, surgeons need precise 3D measurements of the aortic root to select the right prosthetic valve size and align it perfectly.
While C-arm CT (which can be performed right in the operating room) is an ideal tool, it presents two major technical hurdles for automation:
- Inconsistent Image Quality: The timing and amount of contrast agent vary wildly between scans, often resulting in "washed out" images or artifacts from catheters.
- Variable Field of View (FOV): Unlike standard CT, C-arm CT often captures only a portion of the aorta. Traditional "whole-organ" segmentation algorithms fail if the aortic arch or descending aorta is missing from the scan.
Methods
The proposed system utilizes a hybrid approach combining discriminative machine learning with geometric tracking.
Part-Based Aorta Model
Instead of treating the aorta as a single monolithic object, the system decomposes it into four distinct parts: the aortic root, ascending aorta, aortic arch, and descending aorta. This modularity allows the system to process whatever parts are visible in the scan, providing robustness against truncated views.
Detection and Tracking
- Marginal Space Learning (MSL): To find the aortic root and arch, the researchers use MSL, which breaks down the search for an object's position, orientation, and scale into separate stages, significantly reducing computational complexity.
- 2D Circle Tracking: For the tubular ascending and descending aortas, a 2D circle detector trained on Haar wavelets tracks the vessel slice-by-slice starting from the root.
- Hierarchical Landmark Detection: The system detects eight specific landmarks (three hinge points, three commissure points, and two coronary ostia). It first identifies a "global" object containing all points to capture their geometric relationships, then refines each point locally for precision.
Key Results
The study utilized a dataset of 192 C-arm CT volumes from 152 patients, including challenging cases with severe valve regurgitation and poor contrast timing.
- Segmentation Accuracy: The mean symmetric point-to-mesh error was 1.1 mm, outperforming previous state-of-the-art results on MR and standard CT datasets.
- Landmark Precision: After local refinement, the error for aortic hinges dropped to 2.41 mm.
- Efficiency: The entire pipeline processes a volume in approximately 1.4 seconds on a standard dual-core PC, making it roughly 10 times faster than previous semi-automatic methods.
Discussion and Conclusion
The "part-based" decomposition is a critical innovation, allowing the system to handle the variable fields of view common in clinical C-arm CT. By moving from simple image processing heuristics (like thresholding) to discriminative learning, the system achieves robustness against inconsistent contrast levels.
While the system is highly efficient, limitations remain. In approximately 14% of cases, image quality was so poor that even human experts could not identify landmarks. Additionally, the system does not yet explicitly account for patient motion between 3D acquisition and 2D real-time fluoroscopy. Future work may involve extending this approach to standard cardiac CT or other tubular anatomical structures.
Summary updated