Mesh-Cast: A Multidomain Overview

• Mesh-Cast is a multifaceted term applied to various domain-specific techniques that harness mesh structures for data propagation, rendering, simulation, and segmentation.
• Applications include wireless multicast algorithms with channel and interference management, hybrid mesh–NeRF rendering pipelines, and synthetic defect simulations for NDT.
• In medical imaging, Mesh-Cast underpins sequential processing across slices and channels, enhancing MRI tumor segmentation by preserving volumetric context with lower computational cost.

Searching arXiv for "Mesh-Cast" and closely related entries to ground the article. Mesh-Cast is not a single canonical method in the arXiv literature. The label is used for several technically distinct constructs: a multicast design for multichannel wireless mesh networks with Pulse Coupled Neural Network shortest-path computation (Sobana et al., 2012), a hybrid mesh-aware radiance field pipeline that couples polygonal meshes with NeRF rendering and GPU physics (Qiao et al., 2023), a mesh-centric workflow for parametric synthetic casting defects and non-destructive testing data generation (Jeziorski et al., 5 Feb 2026), and the core sequential mechanism in M-Net for MRI tumor segmentation (Lu et al., 28 Jul 2025). Closely related usage also appears in a mesh-centered interpretation of CAST for single-image 3D scene reconstruction (Yao et al., 18 Feb 2025) and in a hyphenated description of Micromegas integration into the CAST helioscope (García, 2015).

1. Nomenclature and research scope

The term appears in multiple domains and does not denote a field-wide standardized framework. A plausible implication is that “Mesh-Cast” functions as a reusable label for methods that either operate on mesh-structured entities, cast information across coupled axes or representations, or link mesh-based reasoning to another computational substrate.

Domain	Meaning of Mesh-Cast	Representative paper
Wireless networking	Multicast tree construction, channel assignment, and dual-source PCNN shortest-path search in WMNs	(Sobana et al., 2012)
Graphics and simulation	Two-way coupling of polygonal meshes, NeRF rendering, lighting, shadows, and XPBD physics	(Qiao et al., 2023)
Manufacturing and NDT	Parametric synthetic defect meshes for cast-metal inspection data generation	(Jeziorski et al., 5 Feb 2026)
Medical imaging	Sequential propagation across slice and channel axes in MRI segmentation	(Lu et al., 28 Jul 2025)
Single-image 3D reconstruction	Mesh-centered view of CAST with SDF-based object generation, alignment, and physics-aware correction	(Yao et al., 18 Feb 2025)
Axion instrumentation	Hyphenated descriptor for Micromegas integration in the CAST helioscope	(García, 2015)

Across these usages, “mesh” can mean a wireless mesh topology, a polygonal surface mesh, a Voronoi-derived defect mesh, or a grid of interactions over time and channel indices. Likewise, “cast” can denote multicast dissemination, ray casting and path transport, casting synthetic defects into a host geometry, or propagating sequential information by tensor transposition. This semantic spread is central to understanding the literature: identical terminology does not imply methodological continuity across the cited works.

2. Wireless mesh-network Mesh-Cast: multicast structure, channel assignment, and DSPCNN

In wireless mesh networks, the relevant formulation targets multicast under assumptions that mesh routers are largely stationary, have rechargeable power, and may be equipped with multiple radio interfaces, each tunable to one of multiple channels, including partially overlapping channels (Sobana et al., 2012). The objective is to maximize system throughput rather than conserve energy. The proposed mechanisms are the Level Channel Assignment (LCA) algorithm, the Multi-Channel Multicast (MCM) algorithm, and the Dual-Source PCNN (DSPCNN) method for shortest-path computation.

The LCA algorithm begins from a BFS partition of the undirected mesh graph $G=(V,E)$ into levels rooted at the source $s$. Parent–child relations are defined only across adjacent levels. The multicast-tree heuristic connects receivers by reusing an existing tree parent whenever possible; otherwise, it randomly chooses a parent, adds it as a relay, and recursively climbs upward until reaching a node whose parent is already in the tree. The stated objective is to minimize the number of relay nodes and total hop count between source and receivers. Channel assignment is level-based: the source uses channel $0$ to send to level $1$; an internal node at level $i\ge 1$ uses receive interface $\mathrm{RI}=i-1$ and send interface $\mathrm{SI}=i$; a leaf similarly uses $\mathrm{RI}=i-1$ and $\mathrm{SI}=i$ to serve local mesh clients. This per-level separation yields a layered forwarding pipeline, but the authors note three limitations: same-channel reuse within a level increases intra-level interference, channels may be wasted when the number of available channels exceeds the number of levels, and adjacent channels remain partially overlapping.

MCM replaces this fixed layering with a slimmer multicast backbone. It first constructs a “tree mesh” by deleting same-level edges and keeping only inter-level edges. Relay selection then proceeds bottom-up: beginning from the lowest level containing receivers, relay nodes are selected greedily to cover receivers or previously selected relays in the next lower level, continuing upward until the source is connected. The stated objective remains the minimization of relay nodes and hop count, now on a reduced inter-level structure.

MCM also introduces two channel-assignment strategies. In ascending channel allocation, channels are assigned top-down in cyclic order $0,1,\dots,C-1,0,\dots$, while maintaining the connectivity constraint $s$0. In the interference-aware heuristic, the interference range between nearby senders $s$1 and $s$2 is modeled as

$s$3

where $s$4 is the common transmission range and $s$5 decreases as channel separation grows. The send-interface channel for relay $s$6 is chosen to minimize

$s$7

The rationale is that squared interference range approximates interference area, so minimizing the sum favors larger channel separations and lower collision probability.

The DSPCNN component addresses shortest-path search. Canonical PCNN dynamics are defined by feeding, linking, internal activity, binary firing, and dynamic threshold variables:

$s$8

$s$9

$0$0

$0$1

DSPCNN modifies this by firing both source neuron $0$2 and goal neuron $0$3 at initialization. Two auto-waves, $0$4 and $0$5, propagate in parallel; when a neuron is excited by both influences, the meeting neuron $0$6 is identified, evolution stops, and the shortest path is reconstructed by backtracking toward $0$7 and $0$8. The paper qualitatively claims higher efficiency and lower search space for very long paths, but it does not report quantitative bounds against Dijkstra, Floyd, or standard PCNN.

The performance evidence is largely qualitative. No explicit simulation environment, node counts, channel counts, traffic model, throughput gain, latency reduction, packet delivery ratio, or statistical significance are reported. The paper instead argues conceptually that fewer relay nodes reduce transmissions, shorter hop counts reduce delay and contention, and multi-channel separation increases spatial reuse.

3. Mesh-Cast as a hybrid mesh–NeRF rendering and simulation system

In graphics, Mesh-Cast denotes a two-way coupling of polygonal surface rendering, volumetric NeRF rendering, light transport, and physics simulation (Qiao et al., 2023). Inserted meshes receive illumination from the NeRF volume, cast shadows into the NeRF, and participate in interreflection and refraction, while NeRF volume segments attenuate and color the light arriving at mesh surfaces. The coupling is implemented by a unified path-tracing integrator that alternates between mesh ray tracing and NeRF ray marching, and by integration with a GPU physics simulator based on XPBD.

The surface component follows the standard rendering equation

$0$9

where $1$0 is emission and $1$1 is the BSDF. The volumetric component uses the NeRF integral

$1$2

The more general radiative-transfer form is stated as

$1$3

but the implementation sets $1$4 for efficiency and relies on the NeRF’s baked radiance $1$5 and density $1$6.

Along a ray with multiple bounces, Mesh-Cast maintains accumulated throughput $1$7 and radiance $1$8. At a surface event $1$9 at position $i\ge 1$0, throughput is updated as

$i\ge 1$1

and emitted radiance is added through $i\ge 1$2. Between surface hits, the system marches through the NeRF volume with step size $i\ge 1$3, applying

$i\ge 1$4

and accumulating radiance with a shadow-dependent visibility factor:

$i\ge 1$5

Shadow rays are traced from NeRF samples to estimated area lights. Mesh occlusion is tested by BVH, and volumetric attenuation is accumulated as

$i\ge 1$6

The point-wise visibility is then the mesh-visibility term multiplied by this transmittance.

A central implementation detail is color-space consistency. Because path tracers assume linear radiance while standard NeRF training usually uses 8-bit sRGB LDR images, Mesh-Cast trains NeRF with HDR images and renders in linear color. The conversion from sRGB to linear is specified as

$i\ge 1$7

The implementation uses Instant-NGP with is_hdr and linear_colors enabled, and applies optional tone mapping only at the display stage.

Emitter estimation is treated as an inverse-rendering problem. The pipeline reconstructs a neural SDF, extracts an explicit mesh by Marching Cubes, unwraps UVs, and optimizes an emission texture in Mitsuba3 with the PRB integrator. The objective is

$i\ge 1$8

Faces with low emission are pruned to form a compact area-light mesh, and discrete light-selection probabilities are defined by emitted energy over emitter areas.

The system is integrated with XPBD for rigid bodies, cloth, and soft bodies. Mesh–NeRF collisions are handled through SDF queries, and moving NeRF objects are updated by inverse transformation of sample coordinates before Instant-NGP evaluation. To preserve interactivity, the implementation uses bounce limits, early ray termination, occupancy-grid skipping, limited shadow-ray budgets, and no volumetric scattering. Reported performance is interactive GPU execution at $i\ge 1$9–$\mathrm{RI}=i-1$0 FPS depending on resolution and scene complexity, with approximately $\mathrm{RI}=i-1$1 FPS at $\mathrm{RI}=i-1$2 on an RTX 2070 Max-Q. The paper also states that rendering time scales linearly with pixels.

The limitations are explicit. Light-source estimation may miss small or indirect emitters; highly specular or transmissive media remain approximate without true volumetric scattering and participating-media MIS; fast dynamics introduce path incoherence and BVH refit overhead; and large NeRF hash grids and BVH structures increase memory footprint.

4. Mesh-Cast for synthetic casting defects and NDT simulation

In manufacturing and inspection, Mesh-Cast refers to a mesh-centric workflow for controllable, parametric synthetic defect geometries embedded into a digital twin and rendered by physically based Monte Carlo simulation for non-destructive testing (Jeziorski et al., 5 Feb 2026). The motivation is that real defect data are scarce and biased by prevention efforts, while rare, safety-critical faults are underrepresented. The method therefore provides parametric 3D mesh defect models, boolean embedding into host geometry, and automatic pixel-perfect annotations.

The geometric foundation is a family of 2D Voronoi tessellations over a window $\mathrm{RI}=i-1$3. For standard Voronoi cells,

$\mathrm{RI}=i-1$4

while weighted power diagrams use

$\mathrm{RI}=i-1$5

and anisotropic variants replace the Euclidean norm by

$\mathrm{RI}=i-1$6

Seeds can be sampled by homogeneous Poisson processes, Poisson disk sampling, stratified sampling, or importance sampling near regions of interest. Example size distributions include Weibull, Gamma, and log-normal laws.

Defect taxonomy is broad. Shrinkage porosity is represented as a tubular or lobed void network thickened from a 2D Voronoi edge graph and embedded in 3D. Gas porosity uses isolated spherical or ellipsoidal SDFs. Inclusions are modeled as volumetric solids generated by a 2D Voronoi lifted into 3D slabs. Cracks and hot tears use a jagged path along Voronoi edges with segment-wise dilation and depressed height profiles. Cold shuts and rat tails use natural cubic splines. Buckles, bulges, and coat lifts are positive or mixed-elevation path-based structures. Surface cavities and scabs are assembled from coarse Voronoi cells refined by a finer tessellation, with outer-region elevation and local texture bulges.

The 2D-to-3D mapping relies on edge thickening, extrusion, and SDF-based boolean operations. For an edge set $\mathrm{RI}=i-1$7 and local thickness $\mathrm{RI}=i-1$8,

$\mathrm{RI}=i-1$9

where $\mathrm{SI}=i$0 is the 2D dilation of edge polylines by segment-wise widths. For cavities or scabs over $\mathrm{SI}=i$1,

$\mathrm{SI}=i$2

Boolean embedding with host geometry $\mathrm{SI}=i$3 can be expressed through

$\mathrm{SI}=i$4

Roughness is added by $\mathrm{SI}=i$5 with Perlin or simplex noise, and anisotropy can be introduced either in the tessellation metric or in anisotropic dilation.

The end-to-end workflow is explicitly staged: seed sampling, tessellation construction, path or cell selection, dilation and height assignment, triangulation into watertight defect meshes, boolean embedding into the host CAD or mesh, remeshing and refinement, physically based rendering or simulation, and automatic annotation generation. Output annotations include semantic and instance masks, depth maps, normal maps, class labels, unique IDs, and parameter vectors linked to each instance.

Physically based simulation supports several NDT modalities. Radiography uses the Beer–Lambert law

$\mathrm{SI}=i$6

with material-dependent attenuation. Visual surface inspection uses a microfacet GGX BRDF,

$\mathrm{SI}=i$7

with additional sensor noise, lens PSF blur, and quantization. Ultrasonics are described as optional, with attenuation and scattering dependent on defect size relative to wavelength.

The framework emphasizes controllability. Reported global parameter ranges include seed density $\mathrm{SI}=i$8 of $\mathrm{SI}=i$9–$\mathrm{RI}=i-1$0 seeds per $\mathrm{RI}=i-1$1 mm, dilation width $\mathrm{RI}=i-1$2, thickness $\mathrm{RI}=i-1$3, sheet thickness $\mathrm{RI}=i-1$4, roughness $\mathrm{RI}=i-1$5, and spline control-point count $\mathrm{RI}=i-1$6. Computationally, 2D Voronoi construction is stated as $\mathrm{RI}=i-1$7, Dijkstra search on the Voronoi graph as $\mathrm{RI}=i-1$8, and local refinement near defects may add $\mathrm{RI}=i-1$9–$\mathrm{SI}=i$0 triangles. Visual PBR with tens of samples per pixel yields $\mathrm{SI}=i$1–$\mathrm{SI}=i$2 s per $\mathrm{SI}=i$3 MP image on a modern GPU. Limitations include thin slivers or self-intersections at sharp turns, possible over-smoothing by SDF blending, simplified physical models for radiography and ultrasonics, and parameter identifiability.

5. Mesh-Cast in M-Net: sequential propagation across slice and channel axes

In medical imaging, Mesh-Cast is the core sequential mechanism in M-Net, a framework for MRI brain-tumor segmentation that treats adjacent 2D slices as a “temporal-like” sequence and propagates information along both the slice axis and the feature-channel axis (Lu et al., 28 Jul 2025). The central problem is that 2D slice-based segmentation loses volumetric continuity, while full 3D convolutions incur substantial computational and memory cost. Mesh-Cast preserves volumetric context without using volumetric kernels.

The model operates on a mini-batch

$\mathrm{SI}=i$4

where $\mathrm{SI}=i$5 is the number of slices in the input window and $\mathrm{SI}=i$6 is the number of image channels or modalities. Spatial dimensions are flattened to $\mathrm{SI}=i$7, yielding $\mathrm{SI}=i$8. A temporal sequential unit processes along the $\mathrm{SI}=i$9 dimension first, producing $0,1,\dots,C-1,0,\dots$0. Mesh-Cast then swaps the temporal and channel axes,

$0,1,\dots,C-1,0,\dots$1

runs a second sequential pass across channels, and transposes back to obtain

$0,1,\dots,C-1,0,\dots$2

At the level of abstract state updates, the temporal pass is written as

$0,1,\dots,C-1,0,\dots$3

and the channel pass as

$0,1,\dots,C-1,0,\dots$4

The paper describes this as a mesh over nodes $0,1,\dots,C-1,0,\dots$5, where messages are first propagated along time and then across channels.

Mesh-Cast is paired with a Vision Sequential Module at every encoder–decoder level. The Vision Sequential Module serializes intra-image structure by a Cross-Scan over four directions: left-to-right, right-to-left, top-to-bottom, and bottom-to-top. Both modules expose a common sequential interface, so the same design can host ConvLSTM, LSTM, xLSTM, GRU, Transformer, Mamba/SSM, or TCN units.

When multiple Mesh-Cast layers are stacked, the paper specifies an SE-style layer attention. If the output of the $0,1,\dots,C-1,0,\dots$6-th Mesh-Cast layer is $0,1,\dots,C-1,0,\dots$7, aggregation may be written as

$0,1,\dots,C-1,0,\dots$8

with small $0,1,\dots,C-1,0,\dots$9 recommended to avoid over-amplification.

Training uses an MRI sequential input pattern and a Two-Phase Sequential (TPS) strategy. Each BraTS case is cropped to $s$00 and split into $s$01 axial slices of $s$02; the four modalities T1, T1c, T2, and FLAIR define $s$03. TPS first trains on shuffled sequences drawn from arbitrary positions and subjects, then fine-tunes on ordered sequences. The paper reports that Shuffle$s$04Order outperforms Order$s$05Shuffle and single-phase alternatives. Loss is a joint BCE and Dice objective over three binary heads for WT, TC, and ET:

$s$06

The computational argument is explicit. A 3D convolution with kernel size $s$07 has complexity

$s$08

whereas Mesh-Cast uses two 1D sequential passes with total cost approximately

$s$09

plus transposes and reshapes. Reported empirical numbers are: M-Net with Mesh-Cast and Mamba SSM at $s$10G FLOPs versus U-Net at $s$11G and nnUNet at $s$12G, and inference time on BraTS 2023 with RTX 2080Ti of $s$13 min for M-Net versus $s$14 min for nnUNet.

The paper reports strong segmentation performance. On BraTS 2019, M-Net with Mesh-Cast, Mamba SSM, and TPS achieves Dice scores of WT $s$15, TC $s$16, and ET $s$17, with Hausdorff95 of WT $s$18, TC $s$19, and ET $s$20. The corresponding backbone using slices only reports WT $s$21, TC $s$22, ET $s$23, and Hausdorff95 of WT $s$24, TC $s$25, ET $s$26. On BraTS 2023, M-Net reports WT $s$27, TC $s$28, ET $s$29, and Hausdorff95 of WT $s$30, TC $s$31, ET $s$32. The authors attribute improved boundary continuity and reduced slice-to-slice jitter to explicit inter-slice dependency modeling.

6. Related CAST-centered usages and terminological boundaries

Two further papers show that the label intersects with the acronym CAST in ways that are related in wording but distinct in substance. In "CAST: Component-Aligned 3D Scene Reconstruction from an RGB Image" (Yao et al., 18 Feb 2025), the paper explicitly states that it does not define a separate variant named Mesh-Cast; rather, the term naturally describes CAST’s mesh-centered instantiation. CAST reconstructs scenes from a single RGB image by combining open-vocabulary 2D instance discovery, monocular relative depth, GPT-4V relation inference, object-wise occlusion-aware 3D generation, alignment by a diffusion model plus Umeyama similarity estimation, and physics-aware pose correction with SDF-based penalties.

In this mesh-centered CAST formulation, each object is represented by an SDF decoded from a geometry latent and then extracted as a triangular mesh by marching cubes. Alignment uses the similarity transform

$s$33

with $s$34 obtained from pointwise correspondences between scene-space and canonical partial point clouds. Physics-aware correction optimizes object poses over a constraint graph derived from fine relations such as Stack, Lean, and Hang. Non-penetration is penalized by $s$35, and contact or support losses are written directly in terms of SDF queries over sampled surface points. Quantitatively, on 3D-Front, CAST reports scene-level Chamfer Distance $s$36, scene-level F-score $s$37, object-level Chamfer Distance $s$38, object-level F-score $s$39, and IoU-B $s$40, outperforming the baselines listed in the paper.

A different CAST-related usage appears in the axion-helioscope literature, where “Mesh‑Cast” is used as a hyphenated descriptor for integrating Micromegas detectors into the CERN Axion Solar Telescope (García, 2015). Here the “mesh” is the Micromegas micro-mesh, not a polygonal or graph mesh. The paper describes microbulk Micromegas detectors with a $s$41 mm drift region, a $s$42–$s$43 $s$44m thick micro-mesh with $s$45 $s$46m hole pitch, a $s$47–$s$48 $s$49m amplification gap, and a $s$50 mm$s$51 active area implemented as $s$52 strips of $s$53 $s$54m pitch in 2011. During the 2011 $s$55He campaign, three Micromegas delivered the bulk of CAST sensitivity, with reported background rates in the $s$56–$s$57 keV region of interest of approximately $s$58–$s$59 counts keV$s$60 cm$s$61 s$s$62 and a Micromegas-only limit of

$s$63

at $s$64 CL for

$s$65

This use is terminologically adjacent to Mesh-Cast but methodologically unrelated to the graphics, networking, manufacturing, and medical-imaging usages.

The main misconception to avoid is therefore straightforward: Mesh-Cast is not a single transferable architecture whose details carry unchanged across fields. In one setting it is a WMN multicast design, in another a hybrid path tracer and simulator, in another a Voronoi-driven defect-modeling workflow, and in another a tensor-transposition mechanism for sequential segmentation. The shared term is real, but the underlying mathematics, objectives, and evaluation protocols are domain-specific.