Low-Dose CT Image Synthesis for Domain Adaptation Imaging Using a Generative Adversarial Network With Noise Encoding...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning (DL) for low-dose X-ray imaging assumes consistent data distributions.
• Low-dose computed tomography (LDCT) images from different scanners exhibit varying noise, violating this assumption.
• Domain shift exists between simulated and clinical LDCT data, hindering model generalizability.

Purpose of the Study:

• To address the domain adaptation problem in low-dose computed tomography (LDCT) image processing.
• To develop a method for generating paired LDCT datasets with diverse noise characteristics.
• To improve the performance of DL-based LDCT processing methods through realistic image synthesis.

Main Methods:

• Proposed a novel generative adversarial network (GAN) incorporating noise encoding transfer learning (NETL), termed GAN-NETL.
• Developed a noise encoding operator to extract noise style and incorporated it into the GAN generator.
• Utilized transfer learning to adapt noise types from source to target domains for realistic synthesis.

Main Results:

• Demonstrated the feasibility and effectiveness of the GAN-NETL model in LDCT image synthesis using public and private datasets.
• Successfully generated paired LDCT datasets with different noise styles.
• Showcased improved performance of DL-based LDCT denoising using synthesized data.

Conclusions:

• The proposed GAN-NETL model effectively solves the domain adaptation problem in LDCT image processing.
• GAN-NETL enables realistic synthesis of LDCT images with target noise characteristics.
• Synthesized data using GAN-NETL can enhance the performance of downstream DL applications like image denoising.