Fast deep learning based reconstruction for limited angle tomography

TL;DR

This work tackles limited-angle CT reconstruction by combining sinogram extrapolation based on moment (range) conditions with a fast, backprojection-centric neural network. The proposed FNOBP architecture employs a Fourier Neural Operator to learn corrections on the sinogram while retaining a single back-projection step, yielding runtimes comparable to classic FBP. Evaluations on Helsinki Tomography Challenge 2022 data show FNOBP consistently surpasses standard FBP methods and competitive performance relative to the HTC-2022 winner, particularly at larger angular spans, while maintaining computational efficiency. The approach highlights a practical path to high-quality, fast reconstructions in scenarios with incomplete data and suggests extensions to 3D and cross-domain transferability.

Abstract

A major challenge in computed tomography is reconstructing objects from incomplete data. An increasingly popular solution for these problems is to incorporate deep learning models into reconstruction algorithms. This study introduces a novel approach by integrating a Fourier neural operator (FNO) into the Filtered Backprojection (FBP) reconstruction method, yielding the FNO back projection (FNO-BP) network. We employ moment conditions for sinogram extrapolation to assist the model in mitigating artefacts from limited data. Notably, our deep learning architecture maintains a runtime comparable to classical filtered back projection (FBP) reconstructions, ensuring swift performance during both inference and training. We assess our reconstruction method in the context of the Helsinki Tomography Challenge 2022 and also compare it against regular FBP methods.

Figures (1)

• Figure 1: Architecture of the FNOBP network