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.
Figures & tables
Fig. 1: Side-view schematic of the Geant4 detector geometry. A 1 m 3 concrete target containing a 7×7 rebar cage sits between two detector stations (3 planes each, spaced 100 mm apart). A 4 GeV μ− enters from above; scattering angle and secondary shower particles are recorded on all six planes.
Fig. 2: Ground-truth voxel label maps for the five target configurations on the 203 grid. Grey voxels represent the 7×7 rebar cage (present in all configurations). Coloured voxels indicate defect regions: orange = honeycombing (scattered bar removal), red = shear fracture (diagonal section removal), purple = corrosion void ( 3×3 corner removal), yellow = delamination (thin air layer). Concrete voxels are rendered transparent.
Fig. 3: SA-DSVN architecture. The scattering stream (9 channels) and shower stream (40 channels) pass through independent 3-stage encoders. At the 53 bottleneck, cross-attention fuses both representations. The decoder upsamples through attention-gated skip connections, with deep supervision at intermediate stages.
Variant
Params
Overall
Honey.
Shear
Corr.
Delam.
SA-DSVN (Full)
1.87M
0.945
0.594
0.790
0.705
0.641
No Attn Gate
1.85M
0.945
0.583
0.801
0.696
0.661
No Deep Sup.
1.87M
0.944
0.575
0.796
0.691
0.655
Scat. Only
1.26M
0.920
0.458
0.666
0.591
0.426
Shower Only
1.27M
0.945
0.586
0.801
0.701
0.654
TABLE I: Ablation study. All models trained for 100 epochs with augmentation. Best values per column in bold .
Fig. 4: Per-defect Dice scores across the five ablation variants. The scattering-only model falls substantially behind all configurations that include shower data.
Fig. 5: Per-defect Dice improvement when adding each stream to the baseline. The shower stream provides the bulk of discriminative power; scattering adds a smaller but consistent contribution, especially for honeycombing.
Fig. 6: Validation Dice and training loss curves for all five ablation variants over 100 epochs. Dual-stream models converge to higher ceilings than the scattering-only variant.
Class
Dice
IoU
Concrete
0.985
0.972
Honeycombing
0.594
0.423
Shear
0.790
0.653
Corrosion
0.705
0.545
Delamination
0.641
0.472
Rebar
0.977
0.958
TABLE II: Per-class Dice and IoU on the validation set (134 volumes, same distribution as training).
Class
Dice
Sens.
Spec.
Prec.
AUC
Concrete
0.972
—
—
—
—
Honeycombing
0.588
1.00
0.80
0.50
1.00
Shear
0.807
1.00
0.90
0.67
1.00
Corrosion
0.698
1.00
1.00
1.00
1.00
Delamination
0.681
1.00
1.00
1.00
1.00
Rebar
0.950
—
—
—
—
TABLE III: Fresh validation on 60 independently generated volumes (10 per class). No fine-tuning or adaptation applied. Volume-level metrics are one-vs-rest binary detection.
Fig. 7: Normalised confusion matrix on the 60-volume fresh validation set. Off-diagonal mass concentrates at defect–concrete boundaries, indicating correct localisation with boundary uncertainty.
Fig. 8: Volume-level ROC curves for defect detection on the fresh validation set. All four classes achieve AUC = 1.000.
Fig. 9: Representative 2D slices through three validation volumes. Left column: ground truth. Centre: model prediction. Right: error map (coloured pixels indicate misclassified voxels). Boundary errors are visible but defect regions are spatially correct.
Defect
No Aug
With Aug
Honeycombing
0.076
0.588
Shear
0.063
0.807
Corrosion
0.000
0.698
Delamination
0.000
0.681
TABLE IV: Effect of augmentation on generalisation. Dice scores on 60 fresh validation volumes.
Fig. 10: Effect of augmentation on generalisation. Test-set performance (same distribution as training) is comparable, but fresh-data performance collapses without augmentation—demonstrating that augmentation is essential for voxelised muon tomography.
Fault detection identifies abnormal behavior in machines, sensors, infrastructure, software, and autonomous systems before it causes harm or disruption.
Institute of High Energy Physics, Chinese Academy of Sciences Beijing 100049, China · Department of Radiology Beijing Jishuitan Hospital Beijing, China