CrossSDF: 3D Reconstruction of Thin Structures From Cross-Sections

CrossSDF introduces a novel approach for reconstructing 3D signed-distance fields from 2D cross-sections, addressing a fundamental challenge in medical imaging, manufacturing, and topography. While existing point cloud reconstruction methods struggle with sparse data between slicing planes, and current specialized approaches have difficulty with thin structures and topological consistency, CrossSDF excels at reconstructing complex geometries from planar cross-sections.

Our method uniquely combines contour-aware neural SDF training with specialized losses designed for known 2D slice geometry. This results in significantly improved reconstruction quality, particularly for thin structures, while avoiding the interpolation artifacts and over-smoothing common in previous approaches. The method is particularly valuable for medical applications where accurate reconstruction of thin vessel structures from CT and MRI scans is crucial.

CrossSDF transforms 2D cross-sectional data into complete 3D neural SDF representations through a sophisticated pipeline that ensures accurate reconstruction of thin structures. Our approach integrates three key innovations: adaptive sampling focused on thin structures, a hybrid encoding combining hash-grid and Fourier features for detail preservation, and a novel symmetric difference loss that maintains topological consistency between slices. This comprehensive approach enables robust reconstruction of complex geometries while preserving fine structural details.

Reconstruction results on thin structures featuring the Heart (top row) and Pulmonary (bottom row) for various methods. Results are presented using both input-aligned and non-aligned planes, displayed on the ground truth meshes.