Reducing windmill artifacts in clinical spiral CT using a deep learning-based projection raw data upsampling: Method...

Area of Science:

• Medical Imaging
• Computed Tomography
• Artificial Intelligence

Background:

• Multislice spiral computed tomography (MSCT) imaging can produce aliasing artifacts, or windmill artifacts, due to insufficient sampling along the z-axis.
• These artifacts manifest as bright streaks diverging from high-contrast structures, degrading image quality.

Purpose of the Study:

• To propose a deep learning-based approach as an alternative to the z-flying focal spot (zFFS) hardware solution for reducing aliasing artifacts.
• To enhance longitudinal sampling in MSCT by developing a supervised learning model for raw data interpolation.

Main Methods:

• A supervised learning model was developed to map input projections to the required rows for doubled z-directional sampling.
• The approach was evaluated using both clinical (40 patient scans) and synthetic (100 simulated scans) datasets for training and validation.
• Performance was assessed qualitatively and quantitatively on test sets of real patient scans and phantom measurements, including a simulation study on scan configurations.

Main Results:

• Deep learning models improved root mean square error by approximately 20% compared to neglecting doubled longitudinal sampling.
• Both clinical and synthetic training data effectively reduced windmill artifacts.
• Training with synthetic data yielded superior performance in artifact reduction compared to clinical data.

Conclusions:

• Deep learning-based raw data interpolation can enhance z-axis sampling and minimize aliasing artifacts in MSCT, offering an alternative to zFFS.
• Training with synthetic data showed particularly promising results for artifact reduction.
• This method provides a beneficial solution for CT scanners without zFFS hardware capabilities.