CelloCut: Cross-Domain Cut Methods

• CelloCut is a cross-domain label encompassing methods that use cuts, partitions, or local modifications to ensure global consistency across applications.
• It spans diverse applications, from biomedical segmentation with candidate multi-cut and random ferns to geometry processing, cut-cell PDE solvers, and structural optimization.
• The various frameworks leverage techniques like graph-cuts, ILP formulations, and finite-volume discretizations to overcome local ambiguities and preserve global objectives.

CelloCut is not a single standardized method. In the literature summarized here, the label denotes several unrelated computational frameworks: a Candidate Multi-Cut formulation for cell and neuron segmentation (Funke et al., 2017), a cell instance segmentation pipeline based on random ferns and graph-cuts (Browet et al., 2016), a constructive watertight remeshing method via tetrahedral cell cuts (Yang et al., 18 May 2026), cut-cell solvers for sharp-interface diffusion (Libat et al., 22 Dec 2025) and incompressible flow (Bouchon et al., 2019), a cut-cell microstructure extension for two-scale structural optimization (Tozoni et al., 2023), and a wolf-suppressor design protocol for large bowed string instruments (Cacace et al., 15 May 2026). This suggests that the term is best understood as a family of domain-specific methods linked by the use of cuts, partitions, or constrained local modifications to obtain globally consistent behavior.

1. Nomenclature and scope

In the surveyed usage, “CelloCut” functions as a cross-domain label rather than a single canonical algorithm. The most explicit authorial use appears in geometry processing, where it names a framework for watertight remeshing via tetrahedral cell cuts. In other cases, the label is used interpretively for methods centered on candidate cuts, graph-cuts, or cut-cell constructions.

Domain	Method associated with “CelloCut”	Core formulation
Biomedical segmentation	Candidate Multi-Cut	ILP over candidate selection and merges
Biomedical segmentation	Random ferns and graph-cuts	Pixel classifier plus multi-label energy
Geometry processing	Constructive watertight remeshing	Binary tetrahedral labeling by min-cut
PDE discretization	Two-fluid diffusion / immersed flow	Cartesian cut-cell finite-volume schemes
Structural optimization / acoustics	Cut-cell tiles / wolf suppressor design	Two-scale optimization / coupled TMD model

A common misconception is that CelloCut always refers to a segmentation algorithm. The surveyed record does not support that reading. It also does not support the converse claim that CelloCut is only a geometry-processing method. In particular, the cut-cell microstructure paper explicitly does not name its method “CelloCut”; the label is only an apt shorthand in the synthesis provided for that work (Tozoni et al., 2023).

2. Candidate Multi-Cut for cell and neuron segmentation

In biomedical image analysis, CelloCut refers to the Candidate Multi-Cut (CMC) model introduced for 2D light microscopy of cell populations and 3D electron microscopy of neurons (Funke et al., 2017). The central motivation is to unify two previously successful but individually limited strategies: merge-tree candidate selection and Multi-Cut clustering of superpixels. Pure merge-tree selection can use rich, large-context candidate features but is constrained by hierarchy bias; pure Multi-Cut is fully expressive over an oversegmentation but operates at the superpixel scale, where many object-level cues are unavailable, and foreground-vs-background selection is not explicit.

The model is defined on a Candidate Region Adjacency Graph, $G = (V, E, S, f, g)$. Nodes $V$ are candidate regions, including leaf superpixels and retained merged candidates from a merge-tree; $E$ contains undirected adjacency edges between touching candidates across hierarchy levels; $S$ encodes directed parent–child hierarchy relations; $f$ assigns learned selection costs; and $g$ assigns learned merge costs. Binary variables $y_i \in \{0,1\}$ indicate candidate selection as foreground, and $m_{(i,j)} \in \{0,1\}$ indicate whether adjacent candidates are merged into the same object. Final segments are the connected components of selected candidates under selected merges, while non-selected candidates are background.

The optimization minimizes

$$\min_{\mathbf{y}, \mathbf{m}} \sum_{i \in V} f_i y_i + \sum_{(i,j)\in E} g_{(i,j)} m_{(i,j)},$$

subject to three constraint families. Conflict cliques enforce non-overlap along each leaf-to-root path,

$$\sum_{i \in C} y_i \le 1 \quad \forall C \in \mathcal{C}.$$

Incidence constraints enforce that a selected merge can only connect selected candidates,

$V$0

Path constraints enforce transitive clustering consistency,

$V$1

Because there can be exponentially many such paths, the ILP is solved with a cutting-plane strategy in which violated path constraints are added lazily.

Feature extraction is split between candidate features and edge features. Candidate features include size, circularity, eccentricity, a contour angle histogram with 16 bins, and intensity statistics over the candidate region and contour pixels for raw and boundary prediction images: sum, mean, variance, skewness, kurtosis, histogram with 20 bins, and 7 quantiles. Edge features include contact area size, intensity statistics across separating boundary pixels, and pairwise combinations of node features using absolute differences, min, max, and sum. Costs are derived from supervised learning with a random forest classifier.

The reported evaluation covers three 2D datasets—HeLa cells, diploid yeast cells, and fission yeast cells—and one 3D serial-section EM dataset, SNEMI3D. On the 2D datasets, CMC improves VOI and DS on two of the three datasets and matches the merge-tree on the third. On the 3D EM dataset, CMC improves VOI and RAND over both Multi-Cut and the assignment model. Oracle-like “best-effort” upper bounds also show higher expressiveness for CMC than for the corresponding merge-tree baselines with the same leaves. The practical significance of the formulation is that it expands the search space beyond the hierarchy while retaining access to larger-than-superpixel features and explicit foreground selection.

3. Random ferns and graph-cuts for cell instance segmentation

A second biomedical usage denotes a two-stage framework for instance-level cell segmentation in live fluorescence microscopy images of embryonic tissue, where membranes are weakly contrasted, fluorophores do not concentrate exclusively on membranes, and adjacent cells often touch (Browet et al., 2016). The stated goal is to assign every pixel to exactly one label among a set of cell instances and a background label, producing pixel-wise labels and delineated instance boundaries. The method addresses weak or blurred edges, high noise, limited spatial resolution, and the tendency of purely region-based clustering to merge neighboring cells.

The first stage is a random ferns classifier that estimates per-pixel interior, border, and exterior evidence. For each pixel, binary tests compare intensities at two random locations in a local window. A fern is a flat ensemble of $V$2 binary tests; the reported experiments use $V$3 ferns with $V$4 binary tests per fern and $V$5 classes: interior, border, and exterior. Training labels are generated from manual contours using morphology: interior from eroded masks, exterior from dilated background, and border from pixels lying on the exterior region of at least two different cells. Approximately 1500 training pixels per class are used, with 10 different orientations for rotational robustness. A semi-Naive Bayes approximation assumes independence between ferns but not between tests within a fern, yielding class log-scores

$V$6

The second stage is a multi-label graph-cut energy over one label per seed plus one background label,

$V$7

Seeds are extracted from connected components in the interior-score map over a decreasing sequence of thresholds, retaining the largest components whose size remains well below the expected cell size. The data term for seed labels is a line integral from a pixel to the seed of $V$8, so paths crossing exterior- or border-like regions are penalized. The background term is a directional line-search of length $V$9 seeking minimal accumulated exterior score; the reported value is $E$0. The pairwise Potts term is modulated by border evidence through a sigmoid weight, with an 8-connected neighborhood and reported parameters $E$1 and $E$2. Label costs discourage overuse of labels and allow redundant seeds to be merged away. Approximate minimization uses $E$3-expansion with label-cost extensions.

The reported validation is on live confocal microscopy of mouse embryos labeled with a membrane marker. A 10-fold cross-validation on the fern classifier yields $E$4 pixel classification accuracy across the three classes. The segmentation assessment is primarily qualitative: the method is reported to separate touching cells, adhere to weak membranes, merge multiple seeds within a single cell because of label costs, and reject seeds in background regions. The paper compares favorably against Mean Shift clustering in the joint spatial-intensity domain and graph-based segmentation by Felzenszwalb and Huttenlocher, but does not report IoU, Dice, or boundary F1.

4. Constructive watertight remeshing via tetrahedral cell cuts

In geometry processing, CelloCut is the name of a constructive framework for watertight remeshing that treats watertight conversion as a volumetric partitioning problem rather than a surface-level repair task (Yang et al., 18 May 2026). The target is a surface that induces a globally consistent interior–exterior partition of 3D space. The motivation is that meshes with complex topology, single-layer structures, or large missing regions often admit locally plausible but volumetrically inconsistent reconstructions, including pseudo-watertight artifacts such as closely spaced double shells.

The pipeline begins with a conservative thickened proxy. An unsigned distance field $E$5 is computed on a $E$6 grid, and the thickening field is defined as

$E$7

with default $E$8. The isosurface $E$9 is extracted by Marching Cubes to obtain $S$0, then decimated to $S$1. A Delaunay tetrahedralization $S$2 is built over the vertices of $S$3, and each tetrahedron $S$4 receives an initial label

$S$5

where $S$6 denotes interior and $S$7 denotes exterior.

The optimization is a constrained binary labeling solved by max-flow/min-cut: $S$8 One-sided constraints preserve proxy-supported interior evidence through infinite unary costs: $S$9 Pairwise penalties are area-weighted and fill-aware: $f$0 with default $f$1. This biases the solution toward compact completions and discourages unsupported newly introduced boundaries. The final mesh is extracted from the interface faces $f$2 where optimal tetrahedron labels differ, followed by SDF conversion and Marching Cubes re-extraction for geometric quality.

The method is explicitly claimed to guarantee strict watertightness by construction, since the min-cut partitions the tetrahedral complex into interior and exterior sets and the output boundary is the interface between them. On ModelNet10, CelloCut, ManifoldPlus, and Dora achieve zero violations across Boundary, Non-manifold Edges, Non-manifold Vertices, and Failure. On CelloScan, CelloCut reports $f$3, $f$4, $f$5, and $f$6, which are the best overall CD, ANC, and F1 among the compared methods. On CelloFill, it reports $f$7, $f$8, and $f$9, giving the lowest FID with competitive LPIPS and strong CLIP. The ablations are diagnostic: removing thickening reduces $g$0 to $g$1 and raises $g$2 to $g$3, while removing the unary one-sided constraint causes degeneration to empty interior and no extractable surface.

5. Cut-cell formulations for diffusion and incompressible flow

In numerical PDE discretization, CelloCut is also used for sharp-interface Cartesian cut-cell methods. One formulation is a two-fluid finite-volume method for static two-phase diffusion problems in which independent diffusion equations are discretized in each phase on a fixed staggered Cartesian grid, with coupling through embedded interface conditions enforcing continuity of normal flux and a general jump law (Libat et al., 22 Dec 2025). Another is an immersed cut-cell method for 2D incompressible flow past solid structures based on a MAC scheme, finite-volume discretization of viscous and convecting terms, and Dirichlet boundary conditions imposed by a Shortley–Weller formulation (Bouchon et al., 2019).

The two-fluid method is built around reduced geometric moments rather than explicit cut-cell polytope construction. Its geometry layer uses trimmed volumes $g$4, face apertures $g$5, interface measures $g$6 and centroids, centroidal face measures $g$7, and staggered volumes $g$8. In each cut cell, two interfacial unknowns are introduced,

$g$9

to enforce flux continuity and the weighted jump law $y_i \in \{0,1\}$0 sharply. The method supports steady Poisson and unsteady diffusion regimes, Dirichlet, Neumann, Robin boundary conditions, and general jumps. Reported convergence behavior is near second-order in bulk $y_i \in \{0,1\}$1, about $y_i \in \{0,1\}$2–$y_i \in \{0,1\}$3 in cut cells, and about first order in $y_i \in \{0,1\}$4 near the interface. The validation includes 1D, 2D, and 3D mono- and diphasic benchmarks, with Neumann tests conserving mass to machine precision, $y_i \in \{0,1\}$5.

The incompressible-flow method considers

$y_i \in \{0,1\}$6

on a Cartesian grid with pressure at cell centers and velocity components on MAC faces. Cut-cell geometry is derived from a level-set representation of the obstacle, and local conservation is maintained by finite-volume evaluation of convective and viscous fluxes on trimmed control volumes. Dirichlet velocity conditions on the solid boundary are imposed sharply with Shortley–Weller stencils. Time integration uses a BDF2–AB2 projection scheme, and linear systems are solved by a capacitance matrix method for fixed obstacles or PETSc with HYPRE BoomerAMG for moving bodies. Numerical results are reported for impulsively started flow around a circular cylinder at Reynolds numbers $y_i \in \{0,1\}$7 and $y_i \in \{0,1\}$8, including agreement of the early-time drag coefficient with the $y_i \in \{0,1\}$9 law of Bar-Lev and Yang and agreement over $m_{(i,j)} \in \{0,1\}$0 with Koumoutsakos and Leonard. A moving-cylinder example at Reynolds number $m_{(i,j)} \in \{0,1\}$1 is also shown.

Taken together, these two cut-cell usages share a design logic: Cartesian regularity is retained in the background grid, while geometric irregularity is isolated in trimmed measures, modified stencils, or interfacial unknowns. The consequence is sharp boundary or interface enforcement without abandoning finite-volume conservation.

6. Other usages: cut-cell microstructures and wolf suppressor design

A further usage appears in two-scale structural optimization, where the term is applied as shorthand to a cut-cell microstructure approach for extending a regular microstructure family to complex 2D and 3D shapes while preserving the shape’s surface (Tozoni et al., 2023). The paper’s own title is "Cut-Cell Microstructures for Two-scale Structural Optimization," and the synthesis explicitly notes that the method is not named “CelloCut” by the authors. Its workflow is nonetheless aligned with that label in the sense that boundary cells are replaced by individually optimized cut-cell tiles conforming to the object geometry, while interior cells retain the original regular family. The reported pipeline performs an initial macro optimization, inserts cut-cell microstructures in boundary cells, re-optimizes only interior cells to compensate for the boundary mismatch, and finally inserts full-cell microstructures in the interior. The abstract states that the performance of the regular microstructure family is only minimally affected by this extension while allowing use on complex 2D and 3D shapes.

In acoustics, the label is also used for a design and tuning protocol for wolf suppressors on large bowed string instruments based on a coupled string–body–bridge–TMD model (Cacace et al., 15 May 2026). The method models the wolf note as an acoustic instability caused by strong coupling between a string and a low-damped body resonance, and studies one or two tuned mass dampers attached to the body or string. Three performance indicators are defined exactly: $m_{(i,j)} \in \{0,1\}$2, which measures low-frequency envelope modulation associated with beating; $m_{(i,j)} \in \{0,1\}$3, which quantifies late-time attenuation; and $m_{(i,j)} \in \{0,1\}$4, which measures spectral deviation from the unsuppressed instrument. In the cello-scale numerical tests, the wolf occurs at $m_{(i,j)} \in \{0,1\}$5 Hz, corresponding to B3, with $m_{(i,j)} \in \{0,1\}$6 above $m_{(i,j)} \in \{0,1\}$7 without suppressors. A representative single-suppressor configuration uses $m_{(i,j)} \in \{0,1\}$8 g, $m_{(i,j)} \in \{0,1\}$9 kg/s, and $$\min_{\mathbf{y}, \mathbf{m}} \sum_{i \in V} f_i y_i + \sum_{(i,j)\in E} g_{(i,j)} m_{(i,j)},$$0 Hz, implying $$\min_{\mathbf{y}, \mathbf{m}} \sum_{i \in V} f_i y_i + \sum_{(i,j)\in E} g_{(i,j)} m_{(i,j)},$$1 N/m. The reported heat maps show different best-compromise placements for plucked and bowed excitation, and the study finds only marginal, not perceptually meaningful, gains from splitting the same total mass into two suppressors for the same wolf frequency.

Across these additional usages, the term does not identify a single mathematical primitive. In one case it refers to boundary-adapted tiles preserving surface geometry under homogenization constraints; in the other it refers to suppressors that split a problematic resonance and create an antiresonance near the wolf frequency. What unifies them is methodological rather than disciplinary: local modifications are introduced precisely where naive regular constructions fail, and those modifications are tuned to preserve a global objective such as deformation behavior, tonal balance, or stability.

7. Cross-cutting interpretation

The surveyed literature shows that CelloCut consistently appears where local evidence is insufficient to guarantee a valid global structure. In Candidate Multi-Cut, the issue is hierarchy bias versus superpixel-scale ambiguity; the response is an ILP that jointly selects and merges candidates across levels. In the random-ferns pipeline, the issue is weak membrane evidence; the response is to combine local interior/border/exterior probabilities with a global multi-label energy. In watertight remeshing, the issue is that local surface geometry does not determine a unique solid interpretation; the response is a volumetric partition over tetrahedra with one-sided constraints. In diffusion and flow solvers, the issue is geometric irregularity on a Cartesian grid; the response is a cut-cell discretization with conservative operators and sharp interface treatment. In structural optimization and acoustics, the issue is that regular interior models fail near boundaries or resonances; the response is a boundary-adapted tile or a tuned suppressor.

This broad usage also clarifies a second misconception: the word “cut” does not always mean graph-cut. In the biomedical random-ferns method and in the watertight remeshing framework, graph-cut optimization is central. In Candidate Multi-Cut, the formalism is an ILP related to Multi-Cut but augmented with candidate selection and path constraints. In the PDE and microstructure settings, “cut” refers instead to cut cells, cut volumes, or cut-cell tiles. In the acoustics setting, the operative mechanism is not a cut in the optimization sense at all, but the insertion of a tuned local resonator that alters the effective admittance seen by the string.

For technical work, the most reliable reading is therefore contextual. “CelloCut” in a segmentation paper most often denotes a structured partitioning method; in geometry processing it denotes a constructive solid reconstruction framework; in numerical analysis it denotes a cut-cell discretization; and in some syntheses it functions as a convenient umbrella label for boundary- or resonance-aware local correction schemes.