Comparison of CT noise reduction performances with deep learning-based, conventional, and combined denoising...

Area of Science:

• Medical Physics
• Radiology
• Image Processing

Background:

• Deep learning algorithms show promise for reducing noise in computed tomography (CT) images.
• Conventional noise reduction methods have long been employed in image processing.

Purpose of the Study:

• To compare the effectiveness of deep learning-based, conventional, and combined denoising algorithms for CT images.
• To evaluate noise reduction, noise power spectrum, and task transfer function for different denoising approaches.

Main Methods:

• Comparison of a conventional adaptive 3D bilateral filter, a 2D deep learning algorithm, and their combination.
• Quantitative analysis using noise power spectrum and task transfer function on phantom and clinical CT data.
• Calculation of effective dose saving factors.

Main Results:

• The 2D deep learning algorithm demonstrated significant CT noise reduction.
• Combining the 2D deep learning algorithm with the 3D conventional filter further improved noise reduction.
• Analysis of noise power spectrum and task transfer function supported the observed denoising effects.

Conclusions:

• Deep learning-based noise reduction is effective for CT images.
• Conventional 3D spatial noise reduction can enhance the performance of 2D deep learning algorithms.
• Combined approaches offer superior CT image denoising capabilities.