Continuous representation-based reconstruction for computed tomography

Computed tomography (CT) imaging has been developed to acquire a higher resolution image for detecting early-stage lesions. However, the lack of spatial resolution of CT images is still a limitation to fully utilize the capabilities of display devices for radiologists.

This limitation can be addressed by improving the quality of the reconstructed image using super-resolution (SR) techniques without changing data acquisition protocols. In particular, local implicit representation-based techniques proposed in the field of low-level computer vision have shown promising performance, but their integration into CT image reconstruction is limited by considerable memory and runtime requirements due to excessive input data size.

To address these limitations, we propose a continuous image representation-based CT image reconstruction (CRET) structure. Our CRET ensures fast and memory-efficient reconstruction for the specific region of interest (ROI) image by adapting our proposed sinogram squeezing and decoding via a set of sinusoidal basis functions. Furthermore, post-restoration step can be employed to mitigate residual artifacts and blurring effects, leading to improve image quality.

Our proposed method shows superior image quality than other local implicit representation methods and can be further improved with additional post-processing. In addition proposed structure achieves superior performance in terms of anthropomorphic observer model evaluation compared to conventional techniques. This results highlights that CRET can be used to improve diagnostic capabilities by setting the reconstruction resolution higher than the ground truth images in training.

Our proposed CRET method offers a promising solution for improving CT image resolution while addressing excessive memory and runtime consumption. The source code of our proposed CRET is available at https://github.com/minwoo-yu/CRET.