TL;DR
This paper introduces a novel algorithm for reconstructing 3D surfaces from 2D cross-section curves, even when those curves lie on non-parallel planes and form complex networks. This advancement is crucial for applications like medical imaging, where anatomical structures need accurate 3D representation from varied scan data.
What Problem Does This Paper Address?
Reconstructing 3D surfaces from 2D cross-sections is a fundamental problem in computer graphics and has wide applications, particularly in bio-medical modeling. Traditionally, methods have focused on scenarios where cross-sections are parallel, making it easier to connect corresponding curves. However, real-world data, such as from freehand ultrasound or physician-selected MRI/CT slices, often involves non-parallel cross-sections. Furthermore, these cross-sections can contain complex "curve networks" that segment the object into multiple tissue types (e.g., skin, muscle, bone), not just a simple inside/outside boundary. Existing methods struggle with this combination of non-parallel planes and complex curve networks, often failing to produce a closed, interpolating surface that correctly partitions the 3D space into labeled regions.
Methods
The authors propose a "divide-and-conquer" approach, inspired by computational geometry. The core idea is to first partition the 3D space into convex "cells" using the input cross-section planes. Then, for each cell, they construct a closed surface network that interpolates the curve networks on the cell's boundary.
The key innovation lies in generalizing techniques for parallel planes to handle arbitrary convex cells. This involves:
- Medial Axes (MA): Instead of a simple medial line between parallel planes, they compute the 3D medial axis (a set of sheets) for each cell.
- Compartmentalization: The cell is then divided into smaller "compartments" using the MA sheets and projections of the curve networks onto the MA.
- Surface Extraction: The surface network within each cell is formed by the boundaries between compartments that have different labels. This process is analogous to how a curve network is extracted in 2D.
- Stitching: Finally, the surface pieces from all cells are stitched together to form the complete, closed surface network.
For mesh improvement, they employ standard techniques like surface diffusion flow for fairing (smoothing) and edge-swapping/triangle splitting for refinement, while ensuring the original curve constraints are maintained.
Datasets
The paper uses both synthetic data and real bio-medical data for evaluation.
- Bio-medical Data: Examples include contours of a human Parotid gland from CT scans, and anatomical structures from mouse brains.
- Synthetic Data: Examples include a synthetic "sapphire ring" and the Stanford Horse model.
No new benchmarks are introduced; the focus is on demonstrating the algorithm's capability on existing types of data.
Key Insights and Ideas
- Generalizing Medial Axes: The extension of medial axis concepts from 2D to 3D for partitioning complex cells is a crucial step in handling non-parallel cross-sections.
- Cell-based Reconstruction: The divide-and-conquer strategy of processing space cell-by-cell provides a robust framework for managing the complexity of arbitrary plane arrangements.
- Guaranteed Closed Surfaces: The algorithm is designed to inherently produce closed surface networks, which is a significant advantage for applications requiring volumetric representation.
- Interpolation of Curve Networks: The method successfully interpolates not just simple curves but also complex networks, preserving the segmentation information.
Key Results
The algorithm successfully reconstructs closed surface networks that interpolate input curve networks on both parallel and non-parallel cross-sections, even when dealing with complex curve networks. The paper demonstrates this on various examples, including anatomical structures and synthetic objects. The mesh improvement step significantly enhances the visual appearance of the reconstructed surfaces. Performance metrics (construction and improvement times) are provided in Table 1, showing reasonable computation times for the tested models.
Limitations and Future Work
The authors acknowledge several limitations:
- Topology Dependence: Similar to other projection-based methods, the topology of the resulting surface can be sensitive to the configuration of cross-section planes and the overlap of projections. Disconnected geometry might occur if projections don't overlap sufficiently.
- Mesh Improvement Artifacts: While fairing and refinement improve appearance, the improved surface might lose its guaranteed self-intersection-free property.
- Domain-Specific Improvement: The current mesh improvement is general. Future work could explore domain-specific techniques, such as using the original 3D volume data (if available) to guide fairing for medical applications.
Future work also includes investigating ways for users to adjust the surface topology and exploring more robust methods for handling disconnected geometry.
Summary updated