A representation-based method for continuous CT image reconstruction

In recent years, computed tomography (CT) imaging requires high-resolution reconstructions that can reveal more details of patients. However, X-ray noise and the blurring effect caused by detector binning disrupt the acquisition of high-resolution images. Both image quality degradation factors can be addressed using image restoration techniques. In particular, in the field of computational image restoration, local implicit neural representation-based techniques that allow flexible adjustment output image resolution have been proposed. However, integrating local implicit neural representation techniques into CT imaging has significant challenges in terms of time and memory consumption. In this work, we propose a method to integrate the representation technique into CT imaging to perform continuous CT image reconstruction with high memory and time efficiency. We aim to apply local implicit neural representation-based techniques to CT imaging, which is computationally intensive due to excessive data size in both the encoder and decoder. To address this inefficiency, we remove unnecessary components of the sinogram data and employ the sinusoidal basis as a decoder to reduce time and memory consumption in both the encoding and decoding processes. As a result, our proposed technique offers lower memory and time resources demand compared to existing methods, and also shows higher reconstructed image quality. Leveraging these advantages, our method allows radiologists to quickly modify the resolution of the selected region of interest (ROI) image to help diagnoses detailed anatomical structures.