3D microscopy segmentation stops at voxel masks: Nagini3D writes each cell as a parametric surface and edges past StarDist3D on boundary IoU

A segmentation result in 3D fluorescence microscopy can now be a continuous surface rather than only a discrete voxel mask, so curvature and deformation modes can be read directly from segmented cells. Earlier 3D segmentation methods output masks voxel by voxel, and shape analysis usually required a separate, time-consuming chain of heterogeneous tools after segmentation. On a CAPS volume, mean local curvature is displayed directly on the predicted surfaces, separating convex from concave regions; on zebrafish trunk neural crest cells in CREST, principal component analysis of the surface parameters gives a first principal direction corresponding to cell elongation, consistent with cells constrained to elongate while migrating through narrow channels.

One forward pass yields both object locations and the surface parameters describing each object, so multi-object segmentation and shape representation become a single step. Active surface methods were largely used for single-object segmentation and required per-object iterative optimization, while pure deep learning returns voxel labels without a continuous shape. During training, predicted surfaces are sampled into point clouds and compared with point clouds derived from ground-truth masks using the sliced Wasserstein distance, with a regularization term penalizing surface derivatives to suppress self-intersection; at inference, local maxima of the score map locate objects, surfaces are generated at those positions, and an optional refinement minimizes an image-gradient energy functional per surface.

The surface representation lets a segmentation be stored as locations and shape parameters instead of a full voxel mask. Voxel masks grow heavy with image size, and a mask for a large image can reach hundreds of megabytes. For larger BONES images, one segmentation mask weighs 90MB while storing surface locations and parameters weighs 0.2MB; the model reaches its performance threshold at M=18 parameters, and increasing to M=30 no longer improves Nagini3D+ quality.

The scope is multi-object segmentation and shape analysis in 3D fluorescence microscopy, evaluated on the synthetic CAPS set and the real BONES and CREST sets, with detection recall and precision, IoU, boundary IoU and mean inference time as metrics. The method targets closed surfaces of spherical topology, and the authors note the shape model struggles with very elongated objects, where an alternative such as tubular parametric surfaces may be needed. Training treats each patch as a 3D volume, with GPU memory requirements ranging from 40 to 80GB depending on the dataset, while inference runs on a standard laptop. The authors mention future exploration of shape priors in refinement or the training loss, and of using the surface representation for tracking and deformation estimation.

The main limitation reported is that the shape model struggles with very elongated objects and would need many control points, with tubular parametric surfaces as a possible alternative. Training requires substantial GPU memory, between 40 and 80GB depending on the dataset, forcing a trade-off among batch size, patch size and the number of surface parameters. In dense, low-contrast regions, per-surface refinement can degrade reliable predictions, so refinement strength is modulated automatically by neighborhood density and contrast. On CREST, the ground-truth masks used for training are sometimes shifted a few voxels inward, and the refinement step corrects such errors, though this behavior was observed on that dataset.