Motion-compensated 4DCT reconstruction from single-beat cardiac CT scans using convolutional networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Imaging

Background:

• 4D cardiac CT reconstruction is crucial for diagnosing heart conditions.
• Motion artifacts significantly degrade image quality in cardiac CT.
• Existing methods struggle to accurately reconstruct dynamic cardiac motion.

Purpose of the Study:

• To develop a novel deep learning-based method for single-heartbeat 4D cardiac CT reconstruction.
• To improve the accuracy and reduce artifacts in dynamic cardiac CT imaging.
• To enhance the diagnostic value of cardiac CT through motion compensation.

Main Methods:

• A deep learning approach was used to split a single cardiac cycle into multiple phases for reconstruction.
• A supervised registration network generated deformation fields between cardiac phases.
• The FBP-and-wrap method combined with a post-reconstruction network addressed motion compensation and artifact removal.

Main Results:

• The method demonstrated improved Root Mean Square Error (RMSE) and Structural Similarity Index Measure (SSIM) compared to traditional methods.
• Reduced blurring was observed in the reconstructed images.
• Validation with simulation data from 40 4D cardiac CT scans confirmed the method's efficacy.

Conclusions:

• The proposed deep learning method effectively reconstructs 4D cardiac CT from a single heartbeat.
• This approach significantly enhances image quality by mitigating motion artifacts.
• The technique offers a promising advancement for cardiovascular imaging and diagnosis.