U-Net Based Image Enhancement for Short-time Muon Scattering Tomography

Muon Scattering Tomography (MST) is a promising non-invasive inspection technique, yet the practical application of short-time MST is hindered by poor image quality due to limited muon flux. To address this limitation, we propose a U-Net-based framework trained on Point of Closest Approach (PoCA) images reconstructed with simulation MST data to enhance image quality. When applied to experimental MST data, the framework significantly improves image quality, increasing the Structural Similarity Index Measure (SSIM) from 0.7232 to 0.9699 and decreasing the Learned Perceptual Image Patch Similarity (LPIPS) from 0.3604 to 0.0270. These results demonstrate that our method can effectively enhance low-statistics MST images, thereby paving the way for the practical deployment of short-time MST.

💡 Research Summary

Muon Scattering Tomography (MST) offers a non‑invasive way to image high‑Z materials, but its practical deployment is hampered by the low flux of cosmic‑ray muons. Short‑duration measurements therefore produce PoCA (Point of Closest Approach) reconstructions that are noisy and lack structural detail. In this work the authors propose a deep‑learning based post‑processing pipeline that dramatically improves the quality of such low‑statistics images.

The core of the method is a U‑Net architecture customized for 2‑D PoCA images (300 × 300 pixels). The network follows the classic encoder‑decoder design with skip connections, double‑convolution blocks, batch normalization, and ReLU activations. To balance pixel‑wise fidelity and perceptual quality, the loss function combines an L1 term with an Image Quality Assessment (IQA) term (α = 0.7). Training is performed with the Adam optimizer (learning rate = 1e‑4), batch size = 4, for 300 epochs on a workstation equipped with an Intel i9‑14900K CPU and an RTX 3050 GPU (6 GB VRAM).

A major challenge is the domain gap between simulated training data and real experimental data. The authors address this by creating a hybrid dataset through a novel “Stamping” technique. First, 1,000 small (5 × 5 pixel) patches are randomly extracted from a low‑event experimental PoCA image, forming a “Real Style Patch Library.” For each simulated PoCA image, 500 patches are randomly selected and pasted at random locations, thereby injecting the complex noise characteristics of the real detector (electronic noise, background fluctuations, systematic artifacts) into otherwise clean simulated images. This double‑randomness strategy prevents over‑fitting to a specific noise pattern while preserving the underlying structural content.

The training dataset itself is built from extensive Geant4‑based Monte‑Carlo simulations. Six hundred thousand muon events are generated for a C‑shaped tungsten target identical to the experimental setup. To emulate detector resolution, isotropic Gaussian positional noise with standard deviations ranging from 0.1 mm to 1.0 mm (in 0.1 mm steps) is added, expanding the dataset by an order of magnitude across event‑level and resolution variations. High‑quality PoCA images (noise‑free, 600 k events) serve as ground‑truth labels, while the noisy, resolution‑varied images become inputs.

Experimentally, the authors collected 18,417 valid muon events over 24 hours using a four‑plane micromegas detector system (spatial resolution ≈ 100 µm). The data were split into eleven independent subsets of 10,000 events each, producing eleven PoCA reconstructions (Image‑1 to Image‑11) for evaluation.

Quantitative results on these experimental images show a striking improvement: SSIM rises from 0.7232 to 0.9699, LPIPS drops from 0.3604 to 0.0270, and PSNR/MSE also improve substantially. Visual inspection confirms that internal voids and high‑Z regions of the tungsten target become sharply defined, eliminating the blur and speckle typical of raw PoCA reconstructions.

The study demonstrates that a U‑Net trained on a carefully engineered hybrid dataset can act as an efficient post‑processor, delivering near‑ground‑truth image quality from short‑time MST measurements. This reduces the required acquisition time and relaxes detector performance constraints, paving the way for practical applications such as border security screening, structural health monitoring, and nuclear material detection. Future work may extend the approach to full 3‑D reconstructions, incorporate a broader range of materials, and integrate the model into real‑time processing pipelines.