Deep learning enabled ultra-fast-pitch acquisition in clinical X-ray computed tomography

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Computed Tomography

Background:

• Fast acquisition speed in X-ray computed tomography (CT) is crucial for clinical applications.
• Helical scans improve acquisition speed, but high helical pitch (p) on single-source CT (SSCT) causes reconstruction artifacts due to data insufficiency (p > 1.5).
• Current limitations restrict helical pitch, hindering faster CT scan times.

Purpose of the Study:

• To develop a deep convolutional neural network (CNN) capable of correcting artifacts caused by ultra-fast helical pitch in CT scans.
• To enable faster CT acquisition speeds beyond current clinical limitations.

Main Methods:

• A customized CNN, UFP-net, was developed using residual learning and custom spatial-domain local and frequency-domain non-local operators for multi-scale feature representation.
• The network was trained and tested on 83 contrast-enhanced patient CT exams (chest, abdomen, pelvis) with consistent base scan parameters.
• Synthesized ultra-fast pitch (p=2, 3) helical CT data were generated from regular pitch (p=1) images; UFP-net input ultra-fast pitch images and output regular pitch images.

Main Results:

• UFP-net significantly improved image quality compared to standard filtered-back-projection (FBP) at ultra-fast pitch settings.
• At p=2, UFP-net achieved mean SSIM > 0.98 and mean rRMSE < 2.9%, outperforming FBP (SSIM < 0.93, rRMSE > 9.1%).
• At p=3, UFP-net demonstrated mean SSIM [0.86, 0.94] and mean rRMSE [5.0%, 8.2%], while FBP showed significantly lower SSIM [0.36, 0.61] and higher rRMSE [36.0%, 58.6%].

Conclusions:

• The UFP-net shows potential for enabling ultra-fast CT data acquisition without compromising image quality.
• The method demonstrated generalizability across different body parts when base scan parameters were consistent.
• This deep learning approach could facilitate faster patient throughput and improved diagnostic capabilities in clinical CT.