An unsupervised two-step training framework for low-dose computed tomography denoising.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Image Denoising

Background:

• Low-dose computed tomography (CT) reduces patient radiation exposure but introduces image noise, hindering accurate diagnosis.
• Deep neural networks, particularly convolutional neural networks, show promise for CT image denoising.
• Supervised learning methods for CT denoising require large datasets of paired normal- and low-dose images, which are often unavailable.

Purpose of the Study:

• To propose a novel unsupervised, two-step training framework for CT image denoising.
• To utilize low-dose CT images from one dataset and unpaired high-dose CT images from another dataset.
• To improve the objective and perceptual quality of denoised low-dose CT images.

Main Methods:

• A two-step training process for a denoising network is employed.
• The first step involves pre-training the network using 3D CT volumes to predict center slices.
• The second step integrates the pre-trained network with a memory-efficient denoising generative adversarial network (DenoisingGAN) for enhanced quality.

Main Results:

• The proposed unsupervised framework demonstrated superior performance compared to traditional machine learning and self-supervised deep learning methods.
• Performance on phantom and clinical datasets was comparable to fully supervised learning methods.
• The denoising framework significantly improved both objective and perceptual quality of noisy CT images.

Conclusions:

• A novel unsupervised learning framework effectively denoises low-dose CT images, enhancing both objective and perceptual quality.
• The method eliminates the need for physics-based noise models or system-specific assumptions, ensuring reproducibility.
• The framework's general applicability extends to various CT scanners and dose levels.