A new training-free computational method could give surgeons and medical engineers cleaner, more reliable three-dimensional models of organs and bones, without relying on the vast labeled datasets normally required by deep-learning systems. The framework, called GDLS-D, reconstructs anatomical surfaces from volumetric medical images in two stages: it first creates a topology-preserving template using a technique named geometric diffusion least squares, then repeatedly deforms that template until it closely matches the target anatomy. In tests involving liver, ilium and vertebral structures, the approach produced watertight meshes while suppressing the staircase-like artifacts that often appear when voxel-based scans are converted into polygonal surfaces. The researchers report that, across three evaluation metrics, their method improved average accuracy by 33.11 percent compared with two leading deep-learning approaches, Voxel2Mesh and MeshDeformNet.
The advance addresses a problem hidden inside nearly every medical scan. Computed tomography and magnetic resonance imaging represent anatomy as a three-dimensional grid of voxels, each carrying an intensity or tissue-related measurement. For diagnosis, those voxels may be enough. But applications such as surgical navigation, biomechanical simulation, patient-specific implants and three-dimensional printing require a continuous surface—a mesh made from vertices, edges and triangular faces. Turning a segmented volume into such a surface is not a neutral translation. A coarse or irregular boundary can produce jagged contours, holes, self-intersections or thin, distorted triangles. These imperfections can interfere with calculations of volume, curvature, stress and fluid flow, and can also make a virtual organ look visibly blocky when displayed in augmented or virtual reality.
Many conventional reconstruction pipelines use algorithms such as marching cubes, which estimate where a tissue boundary crosses each voxel cell and connect those intersections into triangles. The method is fast and remains widely used, but its output can inherit the resolution and geometry of the underlying grid. The result is a “staircase” pattern, particularly along oblique or curved surfaces. Smoothing can make the mesh appear more natural, yet excessive smoothing may erase clinically meaningful details or shift the surface away from the original segmentation. Other techniques can generate watertight surfaces, but may struggle to preserve anatomical topology—the number of connected components, holes and enclosed regions that define how a structure is shaped. A liver, pelvic bone or vertebra must not merely resemble a generic object; it must remain geometrically consistent with the patient-specific anatomy represented in the scan.
GDLS-D separates the reconstruction problem into template generation and geometric refinement. In the first stage, geometric diffusion least squares uses information from the volumetric mask to construct an initial surface that is homeomorphic to the anatomical structure. In practical terms, homeomorphism means that the template has the same essential topology as the target: it can be continuously transformed into the desired shape without tearing, joining separate regions or creating new holes. The diffusion component helps distribute geometric information across the surface, while the least-squares formulation provides an optimization framework for fitting that information. This gives the second stage a stable starting point rather than forcing it to discover both the overall structure and the fine boundary simultaneously.
The second stage treats the template as a deformable surface. Its vertices are moved iteratively so that the mesh approaches the target anatomy while remaining geometrically well behaved. Such optimization must balance competing goals. A surface that follows every small fluctuation in a mask may achieve a low boundary error but become wrinkled, fragmented or filled with poorly shaped triangles. A heavily regularized surface may be smooth and computationally stable but miss narrow anatomical features. The iterative deformation strategy is designed to maintain continuity and improve mesh quality during fitting. The researchers specifically report significantly lower aspect ratios than those of competing outputs. Triangle aspect ratio measures how elongated a triangle is; lower, more balanced values generally indicate a better-conditioned mesh for visualization and numerical simulation.
The method was evaluated on anatomical structures with very different geometries, including the liver, ilium and vertebrae. The reported bidirectional Chamfer distances using the L1 norm were 3.675 millimeters for the liver, 2.654 millimeters for the ilium and 2.508 millimeters for vertebrae. Chamfer distance compares two point sets by measuring how far points on one surface are from their nearest counterparts on the other, and then performs the comparison in both directions. A bidirectional score therefore penalizes both missing regions and unwanted protrusions. The values indicate millimeter-scale geometric agreement, although they should not be interpreted as a direct measure of clinical safety or surgical accuracy. The study was a reconstruction benchmark, not a trial showing that the method improves patient outcomes.
The researchers also tested the framework against deep-learning systems, including Voxel2Mesh and MeshDeformNet. Neural methods can learn a direct relationship between volumetric images and surface meshes, often producing impressive results when training and test data are closely matched. Their weakness is that performance can deteriorate when scans differ in resolution, anatomy, acquisition protocol or segmentation style from the data used during training. GDLS-D is described as training-free, meaning that it does not require a task-specific neural network to learn from a large collection of annotated examples before reconstructing a new surface. Instead, it optimizes each case directly. That design may improve generalization across anatomical structures and reduce dependence on curated labels, though it may also shift computational demands toward per-patient optimization and does not eliminate the need for accurate input segmentation.
A notable tension emerged in the comparison. Traditional or learned methods that closely overfit mask boundaries can achieve better numerical accuracy under some metrics because they reproduce every detail of the reference boundary, including irregularities caused by voxelization or annotation. GDLS-D sometimes gives up that kind of boundary mimicry in favor of a smoother, closed and usable surface. The authors say it effectively suppressed staircase artifacts and produced watertight meshes with lower aspect ratios. This distinction matters because the best mesh for a clinical simulation is not always the one with the smallest point-to-point error. A mesh used to model deformation, plan a resection or guide augmented-reality registration must also be continuous, stable and free of topological defects. Yet smoothing introduces its own risk: if a small protrusion, indentation or vessel-related feature is clinically important, a visually cleaner reconstruction could still be misleading.
The work used publicly available, fully de-identified imaging benchmarks rather than newly recruited patients. The WORD dataset supplied liver, ilium and lumbar-spine cases for the main reconstruction benchmark, while the CHAOS dataset was used for a smaller liver evaluation. Supplementary annotations for the WORD test set were generated with TotalSegmentator, an automated tool for labeling anatomical structures in CT images. Because of licensing restrictions, those derived annotations cannot simply be redistributed, but qualified researchers can reproduce the evaluation by running TotalSegmentator on the public WORD data and applying the reconstruction code. The use of open datasets supports independent testing, although broader validation will require scans from multiple institutions, different imaging protocols and real clinical workflows. The study also does not establish how quickly GDLS-D runs, how much manual correction is needed, or whether its meshes consistently preserve subtle structures that are absent or ambiguous in the original images.
The potential payoff is a bridge between medical imaging and the growing ecosystem of digital twins, surgical simulators and image-guided procedures. A reliable surface mesh can serve as the geometric foundation for simulating tissue mechanics, planning the safest path through an organ, designing patient-specific devices or overlaying virtual anatomy during minimally invasive surgery. Because GDLS-D does not depend on a trained model specialized for one organ, it could be adapted across anatomical structures with less retraining than a neural pipeline. That promise remains preliminary: reconstruction accuracy depends on the quality of the segmentation, optimization settings and target anatomy, and benchmark performance alone cannot guarantee robustness in the operating room. Still, by combining topology-aware initialization with iterative deformation, the study offers a different route to 3D medical reconstruction—one that prioritizes smooth, closed and computationally usable anatomy as well as numerical closeness. The authors present GDLS-D not as a replacement for clinical validation, but as a potentially generalizable foundation for turning volumetric scans into more trustworthy three-dimensional models.
