Noise-augmented deep denoising: A method to boost CT image denoising networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Low-dose computed tomography (CT) reduces patient radiation risk while maintaining image quality.
• Conventional deep denoising methods struggle with the complex noise structure in CT images.
• Noise in CT scans is influenced by global X-ray attenuation and patient-specific factors.

Purpose of the Study:

• To enhance existing deep denoising networks by incorporating additional image noise information.
• To improve the performance of CT image denoising algorithms.

Main Methods:

• Generated synthetic noise realizations for CT images.
• Integrated these noise images as additional input to deep denoising networks (noise augmentation).
• Demonstrated the approach, termed noise-augmented deep denoising (NADD), on modified CNN10, ResNet, and WGAN-VGG networks with 90% dose reduction.

Main Results:

• Noise augmentation significantly improved denoising network performance across all tested cases.
• NADD correctly removed noise artifacts that were previously mistaken for anatomical structures.
• Increased number of noise realizations led to better denoising performance.

Conclusions:

• Simulated noise realizations effectively enhance CT image denoising network performance.
• NADD is an architecture-agnostic method applicable to various deep denoising models.
• This approach offers a promising strategy for improving low-dose CT imaging.