Abstract
Inicialmente centralizadas na solução de problemas científicos em Dinâmica dos Fluidos, as interfaces evolutivas, com o advento da modelagem mais eficiente e robusta provida pelo método Level Set, expandiram os seus limites originais de aplicabilidade para diversos campos de conhecimento. Especificamente à área de Modelagem Geométrica, os trabalhos até então apresentados, relacionando o Level Set à reconstrução de superfícies tridimensionais, concentram-se na reconstrução a partir de uma nuvem de dados dispersos no espaço; a abordagem baseada em fatias planas paralelas e transversais ao objeto a ser reconstruído evidencia-se ainda incipiente. Esse cenário fomenta, portanto, uma análise da viabilidade do Level Set para a reconstrução de superfícies tridimensionais. Fundamentando-se nessa constatação, este trabalho propõe-se a oferecer uma metodologia que agregue, simultaneamente, as ideias comprovadamente eficientes já publicadas sobre a aproximação em questão e as propostas para contornar as limitações inerentes ao método ainda não satisfatoriamente tratadas, em particular a suavização excessiva de características finas dos contornos em evolução sob o Level Set. Relativamente a esse ponto, o emprego da variante Particle Level Set é sugerido como uma solução, por sua intrínseca capacidade comprovada para a preservação de massa de fronteiras dinâmicas. Ao final, conjuntos de dados sintéticos e reais são utilizados para avaliar a metodologia de reconstrução de superfícies tridimensionais apresentada qualitativamente.
Introduction
Reconstructing a 3D object from a stack of 2D cross-sections is a classic challenge in medical imaging and geometric modeling. Traditional methods often rely on triangulation—connecting dots between slices—but this approach frequently struggles when slices are far apart or when the object's shape changes drastically between them, such as a bone branching into two parts.
This paper presents a robust methodology for reconstructing 3D surfaces from 2D parallel slices (like CT scans) using Level Set (LS) and Particle Level Set (PLS) methods. By treating the gap between slices as a "metamorphosis" process, the authors demonstrate how to preserve sharp geometric features and handle complex topological changes more effectively than traditional techniques.
Methodology
The authors employ two primary mathematical frameworks to facilitate planar metamorphosis:
- Level Set (LS) Method: Instead of tracking a curve directly, LS represents the boundary implicitly as the "zero level" of a higher-dimensional function. To fill the gaps between two slices (Source and Destination), the algorithm solves a Partial Differential Equation (PDE) that "evolves" the source contour until it matches the destination. While powerful, LS suffers from numerical diffusion, which causes it to over-smooth sharp corners.
- Particle Level Set (PLS): This is the core enhancement. The authors place "Lagrangian particles" (tiny markers) near the boundary. These particles move along with the evolving shape. If the Level Set function starts to over-smooth a corner, the particles—acting as a secondary memory of the boundary's position—are used to "pull" the function back into place.
Experimental Setup and Datasets
The study evaluated the methodology using both controlled synthetic data and complex real-world medical data. The process involved pre-processing slices (centering centroids and resizing) and then iteratively generating intermediate slices to create a dense, smooth 3D volume. The math was implemented in MATLAB, while the final 3D reconstruction and visualization were handled using Python and the VTK (Visualization Toolkit) library.
Key datasets included:
- Synthetic - "Pirâmide" (Pyramid): Seven rectangular slices of decreasing size, used specifically to test how well the method preserves sharp corners.
- Synthetic - "Castiçal" (Candelabrum): Four slices designed to simulate extreme topological changes (one circle splitting into two, then three).
- Real - Human Anatomy: CT scan data of a Clavicle (collarbone) and an Iliac bone (hip bone), providing a test for high-resolution, organic shapes.
Key Results and Discussion
The "Pyramid" test clearly demonstrated that standard LS rounded off the corners of the rectangles, whereas PLS used its particles to correct the boundary, keeping the corners significantly sharper. Furthermore, the LS framework handled topological metamorphosis—such as one circle splitting into multiple—automatically as part of the physics of the PDE, a significant advantage over mesh-based methods.
However, this fidelity comes with a trade-off in computational speed. Quantitative analysis showed a massive difference in execution time. For the "Iliac" dataset, generating an intermediate slice took 2.35 seconds with standard LS, but jumped to 33.84 seconds with PLS. Despite the time cost, the PLS method generated more refined intermediate slices (2,460 vs 2,448 for the Iliac bone), indicating superior convergence to the destination shape.
Conclusion
While the Particle Level Set method offers superior qualitative results for 3D reconstruction from slices, its primary limitation is computational speed, being roughly 10 to 15 times slower than standard LS. To make this methodology viable for real-time clinical use, future work should focus on hardware acceleration (GPU), parallelization, and the implementation of Narrow Band LS to restrict calculations to the immediate vicinity of the boundary.
Summary updated