Muon scattering tomography utilises muons, typically originating from cosmic
rays to image the interiors of dense objects. However, due to the low flux of
cosmic ray muons at sea-level and the highly complex interactions that muons
display when travelling through matter, existing reconstruction algorithms
often suffer from low resolution and high noise. In this work, we develop a
novel two-stage deep learning algorithm, $\mu$-Net, consisting of an MLP to
predict the muon trajectory and a ConvNeXt-based U-Net to convert the
scattering points into voxels. $\mu$-Net achieves a state-of-the-art
performance of 17.14 PSNR at the dosage of 1024 muons, outperforming
traditional reconstruction algorithms such as the point of closest approach
algorithm and maximum likelihood and expectation maximisation algorithm.
Furthermore, we find that our method is robust to various corruptions such as
inaccuracies in the muon momentum or a limited detector resolution. We also
generate and publicly release the first large-scale dataset that maps muon
detections to voxels. We hope that our research will spark further
investigations into the potential of deep learning to revolutionise this field.

通过开发一种新的双阶段深度学习算法，μ-Net，结合了 MLP 预测子中子轨迹和 ConvNeXt-based U-Net 将散射点转换为体素，实现了在 1024 个子中子剂量下达到 17.14 PSNR 的最新性能，超过传统的重建算法，如最近距点算法和最大似然和期望最大化算法，并发布了首个将子中子探测映射到体素的大规模数据集，我们希望这项研究能进一步探索深度学习在这一领域的潜力。