Deep-learning-based spectral motion artifact correction on photon-counting cardiac CT images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Imaging

Background:

• Photon-counting CT (PCCT) offers improved image quality and reduced radiation dose.
• Cardiac and respiratory motion artifacts remain a challenge in PCCT.
• Deep learning (DL) shows promise for motion artifact correction.

Purpose of the Study:

• To evaluate an image-domain DL-based motion artifact correction method for PCCT.
• To assess the impact of spectral information (material basis images) on correction performance.
• To validate the method on simulated and clinical cardiac PCCT data.

Main Methods:

• Simulated PCCT imaging using XCAT phantoms for DL model training and validation.
• Two DL models were trained: one with and one without spectral information.
• Quantitative analysis of CT number accuracy in regions of interest and organs.
• Visual evaluation of image quality and assessment of cardiac function (wall motion, mechanical delay).

Main Results:

• CT number accuracy improved in 91% of regions and organs, with spectral information yielding the best accuracy in 61%.
• Both DL models enhanced visual image quality in simulated and clinical images.
• The DL model significantly improved left ventricular wall motion and mechanical delay estimation.

Conclusions:

• DL-based motion correction can substantially improve quantitative cardiac PCCT imaging.
• Spectral information significantly enhances the performance of DL motion artifact correction.
• The validated approach shows potential for more accurate cardiac CT analysis.