Domain-adaptive denoising network for low-dose CT via noise estimation and transfer learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Low-dose computed tomography (LDCT) is crucial for reducing radiation exposure in medical diagnostics.
• Deep learning (DL) methods enhance LDCT imaging but struggle with domain shift due to varying noise and imaging conditions.
• Inconsistent feature distributions between training and testing data degrade DL model performance in LDCT denoising.

Purpose of the Study:

• To propose a novel domain-adaptive denoising network (DADN) for low-dose computed tomography (LDCT) imaging.
• To address the out-of-distribution problem in LDCT denoising caused by varying noise and feature distributions.
• To enhance the robustness and performance of DL models in real-world LDCT applications.

Main Methods:

• A novel network model integrating a reconstruction network and a noise estimation network was designed.
• A U-Net-based reconstruction network with spatially adaptive normalization modules was employed.
• A two-stage training strategy involved initial training on a simulated dataset followed by fine-tuning on a torso phantom dataset.

Main Results:

• The DADN model demonstrated strong performance in noise and artifact suppression on both public and clinical LDCT datasets.
• Visual inspection and quantitative analysis confirmed the model's effectiveness.
• The network successfully preserved image contrast and fine details during the denoising process.

Conclusions:

• A novel deep learning network (DADN) was developed to overcome domain adaptation challenges in LDCT image denoising.
• The proposed DADN model proves to be a feasible and effective method for DL-based LDCT image denoising.
• The study highlights the potential of DADN for improving the quality and reliability of LDCT scans.