Unsupervised Prior-Guided Implicit Neural Representation for Iterative Reconstruction in Neutron Coded Aperture Imaging

Abstract

Fast-neutron coded aperture imaging (NCAI) is a key diagnostic technique for resolving the spatial distribution of fusion hotspots in inertial confinement fusion (ICF) experiments. However, recovering the source distribution from the coded image is an illposed inverse problem. Existing supervised deep-learning methods face practical limitations, including limited labeled data, high transfer costs, and poor out-of-distribution (OOD) generalization. To address these limitations, we propose prior-guided INR (PGINR), a fully unsupervised, prior-guided implicit neural representation (INR) method for iterative reconstruction. PGINR embeds the iteratively updated prior image into the INR representation, enabling source intensity prediction to exploit both spatial coordinate information and image-domain prior information. During iterative reconstruction, the current prior image provides image-domain guidance for INR optimization, while the current INR reconstruction is used to update the prior image for subsequent optimization. In numerical experiments, PGINR consistently outperforms several representative methods on both binary-valued and continuous-valued sources. Further robustness experiments show that PGINR maintains its performance advantage under different Gaussian measurement noise levels and practical alignment deviations. Taken together, the qualitative and quantitative results indicate that PGINR enables high-fidelity reconstruction of hotspot geometry and intensity distribution without relying on paired, labeled training data, making it well-suited for data-scarce ICF neutron imaging diagnostics.