A dual-domain deep learning-based reconstruction method for fully 3D sparse data helical CT.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Helical CT is crucial for clinical diagnosis.
• New helical CT prototypes with sparse detectors offer lower dose and cost but face data scarcity for reconstruction.
• Reconstructing images from limited data is an ill-posed inverse problem.

Purpose of the Study:

• To develop a deep learning-based method for reconstructing 3D images from sparse helical CT data.
• To address the challenge of insufficient projection data in novel helical CT systems.
• To improve image quality and enable clinical use of low-dose, cost-effective CT scanners.

Main Methods:

• A deep learning network incorporating a Radon inverse operator and slice disentanglement for 3D reconstruction.
• A three-subnetwork architecture: CNN in projection domain for data estimation/conversion, analytical operator for domain transfer, and CNN in image domain for refinement.
• End-to-end training on simulated data from the AAPM Low Dose CT Grand Challenge dataset.

Main Results:

• The deep learning method successfully reconstructed images from sparse helical CT data.
• Both visual and quantitative evaluations on simulated and clinical datasets showed encouraging results.
• The method effectively addressed the ill-posed inverse problem, demonstrating its potential for clinical application.

Conclusions:

• The proposed deep learning approach provides an effective solution for 3D image reconstruction in sparse helical CT.
• This method can overcome data insufficiency challenges in novel CT systems.
• The technique shows significant potential for clinical adoption, enabling lower radiation doses and reduced costs.