An efficient dual-domain deep learning network for sparse-view CT reconstruction.

Area of Science:

• Medical Imaging
• Deep Learning
• Computational Science

Background:

• Sparse-view CT reconstruction is crucial for reducing radiation dose.
• Traditional methods often struggle with image quality degradation due to limited projection data.
• Deep learning offers a promising avenue for enhancing sparse-view CT reconstruction.

Purpose of the Study:

• To develop an efficient deep-learning based dual-domain reconstruction method for sparse-view CT.
• To assess the clinical value and performance of the proposed method using objective and subjective evaluations.
• To investigate a model with minimal training parameters and comparable running time.

Main Methods:

• Designed two lightweight networks (Sino-Net and Img-Net) for projection and image domain restoration.
• Utilized end-to-end training on prospectively collected clinical thoraco-abdominal CT projection data.
• Performed quantitative analysis of Hounsfield Unit values, noise properties, SNR, and CNR.
• Conducted subjective evaluation by radiologists on image quality, structure conspicuity, and confidence.

Main Results:

• The dual-domain network achieved competitive results in noise and artifact elimination.
• Fine structure details, edges, and contours of anatomic structures were well-restored.
• The method demonstrated good computational performance on clinical data with a 1/6 sparse rate (384 views).

Conclusions:

• An efficient dual-domain deep learning network was developed for sparse-view CT reconstruction from commercial scanner data.
• The study provides a framework for organ-based image quality assessment in sparse-view CT.
• The findings may facilitate organ-specific dose reduction strategies through sparse-view imaging.