Technical note: Phantom-based training framework for convolutional neural network CT noise reduction.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Image Denoising

Background:

• Deep artificial neural networks (ANNs), including convolutional neural networks (CNNs), show promise for CT image denoising.
• A key challenge is acquiring paired high-noise and low-noise CT images for training, which is often impractical in clinical settings.
• This research addresses the need for CNN denoising model optimization using clinically feasible methods.

Purpose of the Study:

• To present a phantom-based training framework for CNN noise reduction.
• To demonstrate efficient implementation on any CT scanner.

Main Methods:

• A supervised learning approach using an image-based noise insertion technique to synthesize training data.
• Utilized patient image series and anthropomorphic phantom noise-only images, superimposed to create simulated low-dose CT images.
• Employed a modified U-Net architecture with mean-squared-error and feature reconstruction loss, tested on whole-body and abdomen-pelvis CT images.

Main Results:

• Achieved 75% noise reduction, 34% decrease in root-mean-square error, and 60% increase in structural similarity on quarter-dose CT images.
• Visual analysis and line profiles confirmed significant noise reduction.
• Demonstrated preservation of spatial resolution for anatomic features.

Conclusions:

• The phantom-based training framework effectively reduces noise and preserves spatial detail in CT images.
• The image-domain approach allows for straightforward implementation without requiring access to projection data.
• This method offers a practical solution for developing CNN denoising models in routine clinical settings.