Domain adaptive noise reduction with iterative knowledge transfer and style generalization learning.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Low-dose computed tomography (LDCT) denoising faces challenges due to lack of paired training data and domain gaps in real-world scenarios.
• Supervised methods struggle with unpaired data, while unsupervised methods often yield suboptimal results.
• Existing methods fail to effectively leverage the strengths of both supervised and unsupervised learning for LDCT denoising.

Purpose of the Study:

• To propose a novel domain adaptive noise reduction framework (DANRF) for effective LDCT denoising.
• To address the domain gap problem by integrating knowledge transfer and style generalization.
• To improve the performance of LDCT denoising in practical imaging scenarios.

Main Methods:

• Developed a domain adaptive noise reduction framework (DANRF) integrating knowledge transfer and style generalization.
• Employed iterative knowledge transfer with knowledge distillation using unlabeled target data and a pre-trained source model.
• Introduced a mean teacher mechanism for source model adaptation and iterative style generalization for dataset diversity.

Main Results:

• The proposed DANRF model demonstrated feasibility and effectiveness in multi-source LDCT image processing tasks.
• Experiments on multi-source datasets confirmed the model's ability to bridge domain gaps.
• The hybrid approach combining supervised and unsupervised learning significantly improved denoising performance.

Conclusions:

• The DANRF model offers a robust solution for practical low-dose CT imaging by effectively tackling domain gaps.
• The framework's hybrid nature combines the advantages of supervised and unsupervised learning for superior denoising.
• DANRF is well-suited for clinical applications, enhancing the quality of low-dose CT images.