Total-body low-dose CT image denoising using a prior knowledge transfer technique with a contrastive regularization...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Reducing radiation exposure in total-body computed tomography (CT) is crucial.
• Low-dose CT (LDCT) can lead to noise and artifacts, impacting diagnosis.
• Existing deep learning methods often over-smooth images by using only normal-dose CT (NDCT) as ground truth.

Purpose of the Study:

• To develop a method that better utilizes NDCT information for LDCT image reconstruction.
• To achieve high-quality LDCT images with preserved details and reduced artifacts.
• To explore advanced deep learning techniques for improved CT image reconstruction.

Main Methods:

• Proposed a novel intratask knowledge transfer (KT) method using a teacher-student consistency network (TSC-Net).
• Implemented a contrastive regularization mechanism (CRM) based on contrastive learning.
• Designed a dynamic enhancement module (DEM) incorporating attention and deformable convolutions.

Main Results:

• Ablation studies confirmed the effectiveness of KT loss, CRM, and DEM.
• TSC-Net demonstrated superior performance over state-of-the-art methods in quantitative and qualitative evaluations.
• Reconstructed images showed excellent results in clinical readings, proving their utility.

Conclusions:

• TSC-Net achieves superior performance in LDCT image reconstruction compared to existing approaches.
• The method successfully reconstructs high-quality CT images for clinical applications.
• Future work may involve fusing PET and CT modalities for further quality enhancement.