Abstract
Quantifying myocardial deformation is essential for diagnosing cardiac disease, yet traditional methods often require labor-intensive manual segmentation. This paper presents Generalized Robust Point Matching (G-RPM), a fully automated 3D algorithm for tracking complex, non-rigid motion of the left ventricle. By integrating Gaussian curvature calculated directly from image gradients and implementing a double-sided outlier handling mechanism, G-RPM achieves high-accuracy motion estimation without prior segmentation. Validation against physical markers in canine MRI sequences demonstrates significant improvements in tracking accuracy compared to standard point-matching frameworks.
Introduction
Accurate quantification of left ventricular (LV) motion and deformation is a cornerstone of cardiac diagnostics. While advanced imaging techniques like MRI tagging provide detailed information, the computational analysis of these images often faces a significant bottleneck: segmentation. Most existing motion-tracking algorithms require the manual or semi-automatic delineation of myocardial boundaries before tracking can commence, a process that is time-consuming and prone to inter-observer variability.
To address this, researchers have explored the Robust Point Matching (RPM) framework, which can align point sets without explicit segmentation. However, standard RPM suffers from two primary limitations in medical imaging: it relies solely on spatial distance, ignoring local shape geometry, and it lacks robustness against outliers in the reference template. This study introduces Generalized Robust Point Matching (G-RPM) to overcome these hurdles.
Methods
The G-RPM algorithm extends the standard RPM framework by incorporating local shape attributes and enhancing noise robustness. The key components of the methodology include:
- Attribute-Augmented Matching: Instead of matching points based only on Euclidean distance, G-RPM utilizes Gaussian curvature as a point attribute. This ensures that points with similar geometric signatures (e.g., high-curvature regions) are more likely to be matched.
- Direct Curvature Estimation: To avoid the noise inherent in surface extraction, Gaussian curvature is calculated directly from the gray-level gradients of the 3D MRI volume using area-based operators. This approach bypasses the need for initial segmentation.
- Robust Outlier Handling: The objective function is modified with a fitting-weight factor (inverse variance) to handle outliers in both the source and target point sets, preventing noise or false features from skewing the transformation.
- Non-Rigid Transformation: The algorithm employs Thin-Plate Splines (TPS) to model the warping of the heart muscle, allowing for the decomposition of motion into global affine and local non-affine components.
Results and Discussion
The performance of G-RPM was evaluated using synthetic 2D data and in-vivo 3D canine cardiac MRI sequences. In synthetic tests with 30% random outliers, G-RPM demonstrated superior stability compared to standard RPM.
Quantitative validation using canine sequences with implanted physical markers revealed that the inclusion of curvature information significantly improves tracking performance. The average displacement error was reduced from 1.61 pixels to 1.38 pixels when curvature attributes were utilized. Furthermore, the algorithm achieved a mean boundary error of 0.40 pixels when compared to expert manual segmentations. Interestingly, the benefit of shape-based matching was most pronounced during periods of small deformation, where it provided a 26% reduction in error.
Conclusion
G-RPM offers a robust, fully automated solution for cardiac motion analysis that eliminates the need for manual segmentation. By leveraging direct curvature calculation and advanced outlier rejection, the system provides accurate tracking even in noisy clinical data. Future work will focus on incorporating modality-specific features, such as corner detectors for MRI tagging, and integrating biomechanical constraints to ensure physiologically plausible motion estimations.
Summary updated