Self-supervised learning for low-dose CT image denoising method based on guided image filtering.

Area of Science:

• Medical Imaging
• Deep Learning
• Image Processing

Background:

• Low-dose computed tomography (LDCT) imaging is crucial for reducing radiation exposure but results in significant image noise.
• Existing deep learning denoising methods often require paired normal-dose and low-dose data, which are difficult to acquire.
• There is a need for effective LDCT denoising techniques that do not rely on paired training datasets.

Purpose of the Study:

• To develop a novel self-supervised denoising method for LDCT images that eliminates the need for paired normal-dose data.
• To enhance denoising performance through the integration of an attention gate (AG) mechanism within a residual network architecture.

Main Methods:

• A self-supervised denoising framework utilizing guided image filtering (GIF) to generate pseudo-labels from LDCT images.
• Implementation of an attention gate (AG) mechanism in the decoder of a residual network to improve feature focus and noise reduction.
• Training the network using only LDCT images to learn noise distributions between the input and generated pseudo-labels.

Main Results:

• The proposed method demonstrated superior denoising performance compared to state-of-the-art unsupervised, transformer-based, and post-processing methods.
• Both visual quality and quantitative metrics showed significant improvements in denoised LDCT images.
• Ablation studies confirmed the optimal performance of the network architecture with the embedded AG mechanism.

Conclusions:

• Self-supervised learning combined with GIF effectively enables LDCT denoising without paired data.
• The integrated attention gate mechanism significantly enhances denoising capabilities, improving feature focus and structural preservation.
• This approach offers a viable solution for high-quality LDCT image reconstruction in clinical settings.