Adversarially Informed Neural Fields for Computed Tomography Reconstruction

Rasmus Juul Pedersen, Luke Besley, Jakob Sauer Jørgensen, Jens Wenzel Andreasen, Anders Bjorholm Dahl, Vedrana Andersen Dahl

Abstract: We present an adversarially informed neural field (AI-NF) framework for tomographic reconstruction from a limited set of projections—a scenario common in medical imaging, materials science, and other applications where full-angle scans are impractical. Limited projection data renders the reconstruction problem severely ill-posed, often leading to ambiguous and noisy solutions requiring appropriately strong regularization. To address these challenges, our method leverages a continuous neural field representation of the attenuation volume, augmented with data-driven regularization through adversarial learning. Specifically, we integrate an adversarial loss and feature matching into a multilayer perceptron architecture with hash encoding, guiding the reconstruction process by learning complex structural priors from a dataset of high-quality volumes. In our ray-based formulation, the neural attenuation field is integrated along discrete sampling points to simulate projection data, while additional regularization terms (smoothness, consistency, and curvature losses) further stabilize the reconstruction. Experiments on human organs from the Medical segmentation decathlon dataset demonstrate that our approach performs comparably to established methods such as SIRT under noise-free conditions, and it significantly outperforms both traditional iterative and baseline neural field methods in the presence of sinogram noise. Our results highlight the potential of adversarially informed regularization to enhance reconstruction fidelity from sparse and noisy measurements, paving the way for more robust imaging in resource-constrained scenarios. Code available at https://github.com/RasmusJuul/CT-Reconstruction-Neural-Representation.

External IDs:doi:10.1007/978-3-031-95911-0_10