A self-supervised guided knowledge distillation framework for unpaired low-dose CT image denoising.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Low-dose computed tomography (LDCT) reduces X-ray radiation but introduces noise and artifacts, hindering medical diagnosis.
• Supervised deep learning methods for LDCT denoising require paired data, which is difficult to obtain in real-world scenarios due to positioning errors and patient movement.
• Unpaired learning offers a more feasible approach for LDCT image enhancement by utilizing non-aligned datasets.

Purpose of the Study:

• To develop a novel unpaired learning framework for improving low-dose computed tomography (LDCT) image quality.
• To address the challenge of acquiring paired data for supervised deep learning in LDCT denoising.
• To enhance noise suppression and detail preservation in LDCT images.

Main Methods:

• A two-stage network training framework named Self-Supervised Guided Knowledge Distillation (SGKD) was developed.
• Stage one involves a self-supervised cycle network for initial LDCT image quality improvement and dataset generation.
• Stage two employs a knowledge distillation strategy using complementary datasets to further refine denoising performance.

Main Results:

• The proposed SGKD framework demonstrated effective noise suppression and detail preservation in LDCT images.
• Experiments on simulated and real-world clinical datasets validated the method's performance.
• Qualitative and quantitative results showed superior outcomes compared to existing state-of-the-art network models.

Conclusions:

• The novel SGKD framework provides an effective solution for unpaired learning in LDCT image denoising.
• This approach overcomes the limitations of supervised methods by utilizing readily available unpaired data.
• SGKD significantly improves LDCT image quality, aiding in more accurate medical diagnoses.