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

Abstract

Objective.low-dose computed tomography (LDCT) images suffer from severe noise due to reduced radiation exposure. Most existing deep learning-based denoising methods require supervised learning with paired training data that is difficult to obtain. To address this limitation, we aim to develop a denoising method that does not rely on paired normal-dose computed tomography data.Approach.we propose a self-supervised denoising method based on guided image filtering (GIF) that requires only LDCT images for training. The method first applies GIF to generate pseudo-labels from LDCT images, enabling the network to learn noise distributions between inputs and pseudo-labels for denoising, without paired data. Then, an attention gate (AG) mechanism is embedded in the decoder stage of a residual network to further enhance denoising performance.Main results.experimental results demonstrate that the proposed method achieves superior performance compared to state-of-the-art unsupervised denoising networks, transformer-based denoising model and post-processing methods, in terms of both visual quality and quantitative metrics. Furthermore, ablation studies are conducted to analyze the impact of different attention mechanisms and the number of AG mechanisms, showing that the proposed network architecture achieves optimal performance.Significance.this work leverages self-supervised learning with GIF to generate pseudo-labels, enabling LDCT denoising without paired data. The embedded AG mechanism, supported by detailed ablation analysis, further enhances denoising performance by improving feature focus and structural preservation.