Physics-Informed Deep Learning Enabled Ultra-Sparse Projection Neutron Computed Tomography

Technical Summary: Physics-Informed Deep Learning Enabled Ultra-Sparse Projection Neutron Computed Tomography

Problem Statement
Neutron Computed Tomography (NCT) is a critical tool for visualizing light elements within dense materials and performing non-destructive inspection of engineered components. However, its application is severely limited by the low flux of neutron sources compared to laboratory X-ray sources. A standard 3D NCT scan typically requires 12–14 hours to acquire the necessary 360 projections, with individual exposures lasting 30–80 seconds. This duration renders time-resolved (4D) imaging practically inaccessible for dynamic processes. While deep learning has been applied to post-processing, existing methods have failed to address ultra-sparse acquisition without introducing "hallucinations"—plausible but physically non-existent features that compromise quantitative analysis. Furthermore, previous approaches have not integrated temporal conditioning to ensure physical validity across projection sequences.

Methodology
The authors propose a physics-informed deep learning framework designed to reconstruct high-fidelity tomographic projections from ultra-sparse angular sampling. The core of the methodology is a spatiotemporal generator–discriminator architecture that operates directly on projection acquisition sequences.

• Architecture: The model utilizes a temporally-conditioned encoder–bottleneck–decoder generator. It employs ConvLSTM units to synthesize intermediate projections based on two surrounding measured frames and a time condition ($t \in [0,1]$). The bottleneck representation is concatenated with a conditioning tensor to enforce temporal context.
• Discriminator: A multi-scale PatchGAN discriminator evaluates three-frame sequences (measured–synthesized–measured) at full, half, and quarter resolutions. This design penalizes inconsistencies in local edges and broader attenuation patterns, encouraging temporal smoothness rather than isolated image realism.
• Training Strategy: The system was trained on 361 projections of an IAEA/PSI contrast phantom acquired at 1° increments. Supervised interpolation pairs were constructed in the form $(i, i+2 \rightarrow i+1)$, where the outer projections served as inputs and the intermediate projection as ground truth. The model was trained using a loss function combining adversarial, L1, and SSIM terms.
• Inference: For sparse-view inference, the trained generator is applied recursively or sequentially between measured angular positions to synthesize missing views required for reconstruction. The framework was tested for zero-shot generalization across three facilities (BRIN, PSI, and IMAT) using diverse samples, including resolution phantoms, a motorcycle engine block, a computer mouse, and Berea sandstone.

Key Contributions

1. Ultra-Sparse Reconstruction: The framework successfully reconstructs 90% of projection data using only 10% of the acquisition time, reducing the number of required projections from 360 to as few as 37.
2. Temporal Conditioning: By integrating ConvLSTM units, the model explicitly conditions on temporal context, learning the manifold of physically plausible projections to avoid hallucinations.
3. Quantitative Validity: Unlike generative approaches that risk introducing non-physical features, this method preserves quantitative structural integrity, validated through grey-level measurements and nominal/actual comparisons.
4. Cross-Facility Generalization: The model demonstrates robust generalization across different beamline geometries and facilities without retraining.

Results
The framework was validated on 16-bit, 1024×1024 neutron radiography data. Key performance metrics include:

• Image Quality: Reconstructions from ultra-sparse sampling (37 projections) achieved a Peak Signal-to-Noise Ratio (PSNR) of $41.62 \pm 2.78$ dB and a Structural Similarity Index (SSIM) of $0.9799 \pm 0.0095$.
• Material Detection: In a contrast phantom containing six materials, the ultra-sparse AI-enhanced reconstructions preserved material-specific grey-level contrasts consistent with theoretical macroscopic cross-sections. No detectable hallucinations were observed in the material distribution.
• Geometric Accuracy:
  • Resolution Phantoms: Measurements of slit thickness in FeAl and AlCu phantoms showed strong correlation with ground truth ($R^2 = 0.981$ and $0.984$, respectively), with deviations within acceptable limits.
  • Engine Block: Geometry measurements (length, width, height) showed no statistically significant differences between ultra-sparse reconstructions and ground truth.
  • Nominal/Actual Comparison: Using VG Studio Max 3.0, the analysis revealed that over 92% of surfaces (averaged across samples) lay within a tolerance of $\pm 0.089$ mm. The contrast phantom achieved 99.9% area within tolerance, while the more complex engine block achieved 56.2%, with specific features still matching perfectly.
• Time Efficiency: The acquisition time was reduced from 12–14 hours to 4–5 hours at the BRIN facility, with potential for greater speedups at higher-flux facilities.

Significance
The paper claims that this framework enables operando studies of fast dynamic processes that were previously accessible only with synchrotron X-ray sources. By reducing acquisition time by over 90% while maintaining quantitative structural integrity and avoiding hallucinations, the method overcomes the primary bottleneck of neutron imaging: low flux. The ability to generalize across different facilities and sample types suggests a pathway toward practical, high-speed neutron tomography for industrial and scientific applications where time-resolved data is critical.