PrIntMesh: Precise Intersection Surfaces for 3D Organ Mesh Reconstruction

TL;DR

PrIntMesh addresses the need for anatomically plausible 3D organ reconstructions by enforcing topology through a unified, multi-part template that is deformed by image features. A two-stream architecture combines voxel-based segmentation with a mesh-deformation decoder, while explicit supervision of shared interfaces and geometric regularization preserve internal walls and surface smoothness. The method optimizes a global loss that combines $L_{CE}$, $L_{Dice}$, and region-wise $L^{l}_{\text{match}}$ with mesh-regularization terms, yielding watertight, topologically correct meshes across heart, hippocampus, and lungs, even with limited data. This results in data-efficient, clinically applicable reconstructions with strong topology guarantees, reducing the need for post-processing and enabling downstream simulations.

Abstract

Human organs are composed of interconnected substructures whose geometry and spatial relationships constrain one another. Yet, most deep-learning approaches treat these parts independently, producing anatomically implausible reconstructions. We introduce PrIntMesh, a template-based, topology-preserving framework that reconstructs organs as unified systems. Starting from a connected template, PrIntMesh jointly deforms all substructures to match patient-specific anatomy, while explicitly preserving internal boundaries and enforcing smooth, artifact-free surfaces. We demonstrate its effectiveness on the heart, hippocampus, and lungs, achieving high geometric accuracy, correct topology, and robust performance even with limited or noisy training data. Compared to voxel- and surface-based methods, PrIntMesh better reconstructs shared interfaces, maintains structural consistency, and provides a data-efficient solution suitable for clinical use.

Figures (9)

• Figure 1: Voxel-based vs Mesh-based approaches. (a) Voxel-based approaches such as nnU-Net Isensee21 currently are the dominant ones for organ segmentation. However, they can yield gaps, floaters, and jagged boundaries, requiring post-processing to produce usable meshes. (b) Existing mesh-based methods generate smooth surfaces but model each class separately, leading to interpenetrations and misaligned boundaries between components. (c) PrIntMesh deforms a joint template for all components and precisely reconstructs shared surfaces, producing a smooth, watertight and topologically correct mesh, without post-processing.
• Figure 2: Building the heart template. (a) Initial rhombicuboctahedron. (b) Four rhombicuboctahedra glued together. (c) Their vertices are uniformly displaced to create four spheres.
• Figure 3: Templates for (a) the hippocampus and (b) the lungs.
• Figure 4: Network Architecture. The voxel encoder and decoder follow a U-Net architecture including skip-connections, and are trained with cross-entropy loss. The initial heart template goes through a simple alignment procedure based on the segmentation, and is then deformed and subdivided in multiple steps by the mesh decoder. The mesh decoder samples features from the voxel decoder to apply correct deformations, and the shape loss is applied to each intermediate mesh.
• Figure 5: Qualitative comparison of 4-chamber heart outputs. Left column describes the anatomical view. Each row compares methods visually. Our method produces smoother and more consistent geometry, especially in edge-sensitive regions.