Shower-Aware Dual-Stream Voxel Networks for Structural Defect Detection in Cosmic-Ray Muon Tomography

Palash Dutta; Parthiv Dasgupta; Raja Karmakar; Sambhav Agarwal; Sudeshna Goswami

arxiv: 2604.03741 · v1 · submitted 2026-04-04 · 💻 cs.CV · physics.comp-ph

Shower-Aware Dual-Stream Voxel Networks for Structural Defect Detection in Cosmic-Ray Muon Tomography

Parthiv Dasgupta , Sambhav Agarwal , Palash Dutta , Raja Karmakar , Sudeshna Goswami This is my paper

Pith reviewed 2026-05-13 17:35 UTC · model grok-4.3

classification 💻 cs.CV physics.comp-ph

keywords muonshowerscatteringchannelscosmic-raydefectdetectiondice

0 comments

The pith

SA-DSVN achieves 96.3% voxel accuracy and 0.59-0.81 per-defect Dice scores on simulated muon tomography data by fusing scattering and shower-multiplicity streams, with the shower stream alone driving most of the gain over scattering-only baselines.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Cosmic-ray muons pass through concrete and scatter based on material density while also creating secondary electromagnetic showers. Traditional imaging uses only the scattering angles. This work trains a neural network with two separate processing streams: one for scattering data and one for shower counts. The streams exchange information through cross-attention before producing a 3D map of defect locations. On large simulated datasets containing honeycombing, fractures, corrosion voids, and delamination inside rebar cages, adding the shower stream raised average detection quality from 0.535 to 0.685 Dice and reached 100% volume-level detection at fast inference speed.

Core claim

A five-variant ablation study demonstrates that the shower multiplicity stream alone accounts for the majority of discriminative power, raising defect-mean Dice from 0.535 (scattering only) to 0.685 (shower only).

Load-bearing premise

The Geant4-based Vega simulations faithfully reproduce real muon scattering angles, shower multiplicities, and defect geometries inside dense rebar cages, and that the 60 validation volumes are statistically representative of field conditions.

read the original abstract

We present SA-DSVN, a 3D convolutional architecture for voxel-level segmentation of structural defects in reinforced concrete using cosmic-ray muon tomography. Unlike conventional reconstruction methods (POCA, MLSD) that rely solely on muon scattering angles, our approach jointly processes scattering kinematics (9 channels) and secondary electromagnetic shower multiplicities (40 channels) through independent encoder streams fused via cross-attention. Training data were generated using Vega, a cloud-native Geant4 simulation framework, producing 4.5 million muon events across 900 volumes containing four defect types - honeycombing, shear fracture, corrosion voids, and delamination - embedded within a dense 7x7 rebar cage. A five-variant ablation study demonstrates that the shower multiplicity stream alone accounts for the majority of discriminative power, raising defect-mean Dice from 0.535 (scattering only) to 0.685 (shower only). On 60 independently simulated validation volumes, the model achieves 96.3% voxel accuracy, per-defect Dice scores of 0.59-0.81, and 100% volume-level detection sensitivity at 10 ms inference per volume. These results establish secondary shower multiplicity as a previously unexploited but highly effective feature for learned muon tomographic reconstruction.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper's main contribution is a dual-stream voxel network that treats secondary shower multiplicity as the dominant signal for defect segmentation in simulated muon tomography, with an ablation showing clear gains over scattering-only baselines.

read the letter

The central point is straightforward: this dual-stream architecture processes scattering kinematics in one encoder and shower multiplicities in another, then fuses them with cross-attention for voxel-level defect segmentation in reinforced concrete. On their Geant4/Vega simulations the shower stream alone lifts mean Dice from 0.535 to 0.685, and the full model reaches 96.3% voxel accuracy plus 100% volume-level detection on 60 held-out volumes. That ablation is the clearest part of the work; it isolates the contribution without obvious circularity and gives concrete per-defect Dice ranges of 0.59-0.81. The inference speed (10 ms per volume) is also practical for the intended civil-engineering use case. The simulation setup itself is described in enough detail to be reproducible if someone has access to Vega. The soft spots are mostly around validation. All numbers come from simulated data with no real muon-detector measurements or direct comparison of simulated versus observed shower multiplicities and scattering tails inside dense 7x7 rebar. If the Geant4 physics lists over- or under-produce electromagnetic showers in high-Z material, the reported advantage will not carry over. The abstract also omits error bars, training variance, or any sensitivity checks on the simulation parameters. The 60 validation volumes are modest for claiming robustness across four defect types. This paper is mainly useful to groups already working on learned reconstruction for muon tomography or non-destructive testing of concrete. A reader who wants to see how shower information can be turned into a trainable feature will find the ablation helpful, but anyone needing field-ready performance should treat the numbers as preliminary. I would send it to peer review so referees can press on the simulation fidelity and ask for at least one real-data sanity check or physics-list variation study. The core idea is clean enough that the review process would likely strengthen it rather than sink it.

Referee Report

3 major / 1 minor

Summary. The manuscript presents SA-DSVN, a dual-stream 3D convolutional architecture for voxel-level segmentation of structural defects (honeycombing, shear fracture, corrosion voids, delamination) in reinforced concrete using cosmic-ray muon tomography. It processes 9-channel scattering kinematics and 40-channel secondary electromagnetic shower multiplicities through independent encoders fused by cross-attention. Data are generated via the Vega Geant4 framework (4.5 million events across 900 training volumes with a dense 7x7 rebar cage); a five-variant ablation shows the shower stream alone raises defect-mean Dice from 0.535 (scattering-only) to 0.685 (shower-only). On 60 held-out simulated validation volumes the model reports 96.3% voxel accuracy, per-defect Dice 0.59-0.81, 100% volume-level sensitivity, and 10 ms inference.

Significance. If the Geant4/Vega simulations faithfully reproduce real shower multiplicities and scattering tails inside dense rebar, the work would be significant: it isolates shower multiplicity as a previously unexploited, high-impact feature that substantially outperforms conventional scattering-only methods (POCA, MLSD) in a learned reconstruction setting. The ablation design cleanly attributes the Dice gain to the shower stream and the reported inference speed is attractive for practical deployment.

major comments (3)

[Abstract / Results] Abstract and Results section: all quantitative claims (Dice 0.685 vs 0.535, 96.3% accuracy, 100% sensitivity) are derived exclusively from Vega/Geant4 simulations; no real muon-detector data, no direct comparison of simulated versus measured shower-multiplicity or scattering-angle distributions inside a 7x7 rebar cage, and no sensitivity analysis to Geant4 physics-list variations are reported. This is load-bearing for the central claim that shower multiplicity is “highly effective” under field conditions.
[Results] Results (ablation study): the reported Dice improvement of 0.15 is presented without error bars, standard deviations across random seeds, or training-stability metrics, making it impossible to judge whether the shower-stream advantage is statistically robust or sensitive to hyper-parameter choices.
[Methods] Methods: the simulation protocol (Vega/Geant4) is described but no quantitative validation metrics (e.g., Kolmogorov-Smirnov tests on multiplicity or scattering-angle histograms) against experimental reference data are supplied, leaving the weakest assumption—that the simulated defect geometries and shower physics match reality—unexamined.

minor comments (1)

[Abstract] Abstract: the per-defect Dice range 0.59-0.81 is given without identifying which defect achieves which score; a small table would improve readability.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for the thorough and constructive review. We acknowledge that the work is entirely simulation-based and will revise the manuscript to clarify limitations and strengthen statistical reporting. Point-by-point responses to the major comments follow.

read point-by-point responses

Referee: [Abstract / Results] Abstract and Results section: all quantitative claims (Dice 0.685 vs 0.535, 96.3% accuracy, 100% sensitivity) are derived exclusively from Vega/Geant4 simulations; no real muon-detector data, no direct comparison of simulated versus measured shower-multiplicity or scattering-angle distributions inside a 7x7 rebar cage, and no sensitivity analysis to Geant4 physics-list variations are reported. This is load-bearing for the central claim that shower multiplicity is “highly effective” under field conditions.

Authors: We agree that all reported metrics are derived from simulations. The study is intentionally scoped as a controlled simulation experiment to isolate the contribution of the shower-multiplicity stream with perfect ground truth. In the revision we will add an explicit Limitations section stating that results are simulation-derived, discuss Geant4 physics-list assumptions, and note the absence of real-detector validation. We will also include a sensitivity study to selected Geant4 physics lists. revision: partial
Referee: [Results] Results (ablation study): the reported Dice improvement of 0.15 is presented without error bars, standard deviations across random seeds, or training-stability metrics, making it impossible to judge whether the shower-stream advantage is statistically robust or sensitive to hyper-parameter choices.

Authors: We accept this criticism. The revised manuscript will report the ablation results as means and standard deviations over five independent training runs with different random seeds, and will include error bars on the Dice scores to demonstrate robustness. revision: yes
Referee: [Methods] Methods: the simulation protocol (Vega/Geant4) is described but no quantitative validation metrics (e.g., Kolmogorov-Smirnov tests on multiplicity or scattering-angle histograms) against experimental reference data are supplied, leaving the weakest assumption—that the simulated defect geometries and shower physics match reality—unexamined.

Authors: We will add quantitative validation in the revised Methods section by comparing simulated scattering-angle and shower-multiplicity distributions against published experimental references (where available) and will report Kolmogorov-Smirnov statistics. We note, however, that high-statistics experimental data for the precise 7x7 rebar cage geometry are not publicly available, so the comparison will be limited to general muon-scattering and shower-multiplicity benchmarks. revision: partial

standing simulated objections not resolved

Absence of real muon-detector data and direct simulated-versus-measured comparisons inside a dense 7x7 rebar cage; these cannot be supplied without new experimental campaigns.

Circularity Check

0 steps flagged

No significant circularity in derivation or ablation results

full rationale

The paper presents an empirical 3D CNN architecture (SA-DSVN) trained via supervised learning on 4.5M Geant4/Vega-simulated muon events across 900 volumes, with performance reported on a separate set of 60 independently simulated validation volumes. The ablation (scattering-only vs. shower-only streams) computes Dice scores post-training on held-out data; these metrics are not equivalent by construction to any fitted parameters or inputs. No equations, self-citations, uniqueness theorems, or ansatzes appear in the provided text that would reduce the central claims to definitional loops. The chain is standard data-driven evaluation and remains self-contained against the simulation benchmark.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central performance claims rest on the assumption that Geant4 accurately models both scattering and shower production in the specific geometry of rebar-reinforced concrete with embedded defects; no independent experimental calibration of these shower multiplicities is provided.

free parameters (1)

network hyperparameters and training schedule
Standard deep-learning training involves many tunable values (learning rate, batch size, attention dimensions) that are chosen or optimized on the training set.

axioms (1)

domain assumption Geant4 accurately reproduces muon scattering angles and electromagnetic shower multiplicities inside dense rebar cages with the four defect types
All 4.5 million training events and the 60 validation volumes are generated under this physics model.

pith-pipeline@v0.9.0 · 5547 in / 1344 out tokens · 39330 ms · 2026-05-13T17:35:17.926027+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

dual-stream architecture that separates scattering and shower features... cross-attention fusion at the bottleneck... ablation shows shower multiplicity stream alone accounts for majority of discriminative power (0.685 vs 0.535 Dice)
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Vega... Geant4 v11.x with QGSP_BERT physics list... Coulomb scattering, bremsstrahlung, pair production

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.