Iterative Deep Learning for Muon Scattering Tomography

Muon scattering tomography (MST) allows to reconstruct internal structure of the matter based on imaging of muon rays, Coulomb scattering when passing near heavier nuclei of the penetrated matter. While point of closest approach (PoCA) and maximum-likelihood expectation maximization (MLEM) methods offer reasonable MST solutions, they require large number of muons or huge computational cost while still struggling to distinguish between similar materials. As an alternative we propose to utilize deep, fully convolutional neural networks with atrous convolutions designed to deliver robust semantic image segmentation (SIS). The information from muon rays is summarized as image pixels containing path length and momentum weighted statistics of per-voxel muon scattering angles and displacements encoded independently along image rgb channels. Reserving one channel for iterative self-improvement by feeding classification outputs back as inputs delivered significant material detection improvements with nearly perfect localization and improved material differentiation. The presented iterative SIS model has been tested within IEEE BigData 2023 Cup dedicated to muon scattering tomography and scored $2 ^{nd}$ place significantly outperforming PoCA/MLEM and Catboost with 99.9% localization accuracy and IoU score in excess of 0.35, on the very challenging task of detecting nested material layers composed of a mixture of up to 22 elements and compounds.