Convolutional auto-encoder for image denoising of ultra-low-dose CT

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Ultra-low-dose computed tomography (CT) reduces radiation exposure but often suffers from image noise and artifacts.
• Effective denoising is crucial for maintaining diagnostic quality in ultra-low-dose CT (ULCT).

Purpose of the Study:

• To validate a patch-based image denoising method utilizing a convolutional auto-encoder neural network for ULCT.
• To compare the proposed method against large-scale nonlocal mean and block-matching and 3D filtering algorithms.

Main Methods:

• Acquisition of standard-dose (300 mA) and ultra-low-dose (10 mA) CT images of a chest phantom.
• Training a convolutional auto-encoder neural network with paired standard-dose and ULCT image patches.
• Visual assessment of denoised images by radiologists and technologists for artifacts, noise, vessel visualization, and overall quality.

Main Results:

• The proposed neural network method demonstrated statistically significant improvements in reducing streak artifacts, non-streak noise, and enhancing pulmonary vessel visualization compared to block-matching and 3D filtering (p < 0.05).
• Performance against large-scale nonlocal mean was also statistically superior across all assessed parameters (p < 0.05).
• Overall image quality showed no statistically significant difference between the proposed method and block-matching and 3D filtering (p = 0.07272).

Conclusions:

• A convolutional auto-encoder neural network can be effectively trained on standard-dose and ULCT image pairs for denoising.
• The proposed patch-based neural network denoising method shows superior performance compared to large-scale nonlocal mean and block-matching and 3D filtering in visual assessments.