A Dual CT Image Denoising Approach using Guided Filter and method-based Noise in the NSST Domain

Low-dose Computed Tomography (LDCT) images are often corrupted by Gaussian noise owing to electronic interference and environmental factors during image acquisition. This type of noise significantly degrades image quality, obscures fine structural details, and hinders accurate medical interpretation. To overcome this, the current study aims to develop an effective denoising method that suppresses Gaussian noise, preserves edges and sharp features, and improves the overall visual quality. The proposed method outperforms existing denoising techniques in terms of Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), Entropy Difference (ED), Feature Similarity Index Measure (FSIM), and Root Mean Squared Error (RMSE), thereby improving overall diagnostic precision.

The proposed denoising method integrates the Nonsubsampled Shearlet Transform (NSST), guided filtering, and BayesShrink thresholding method in a dual-stage process. Initially, the NSST process decomposes the noisy CT image into approximation and detail components. The approximation component is improved using guided filtering to preserve fine structural details, whereas detail components are denoised using BayesShrink thresholding for noise reduction. A method noise-based approach is then evaluated and further enhanced by applying NSST and BayesShrink methods to the residual part. Finally, the denoised outputs from the two stages are fused to reconstruct the final denoised CT image. The comparative evaluation of the proposed method against Discrete Wavelet Transform (DWT), NSST with bilateral filtering, Method noise-based Convolutional Neural Network (CNN), NSST with Bayes shrinkage, NSST with Wiener filtering, and Stein’s Unbiased Risk Estimate Linear Expansion of Thresholds (SURELET), and Tetrolet transform. Quantitative evaluation metrics (PSNR, SSIM, ED, FSIM, and RMSE) were used across varying noise levels (σ = 5,10,15,20), confirming its consistent superiority in noise suppression and edge preservation.

The proposed denoising method consistently outperformed all other standard methods at varying noise levels. It achieved higher PSNR and SSIM values, lower RMSE values, and enhanced ED and FSIM values. These experimental outcomes demonstrate superior denoising performance in terms of both noise reduction and edge detail preservation.

The integration of NSST with guided filtering and Bayesian thresholding significantly improves the denoising ability in LDCT images without degrading the fine image details. The iterative method noise-based refinement process improved the denoising performance. Although computationally intensive, the proposed method is clinically applicable, specifically in lower-dose imaging scenarios.

This study demonstrates a hybrid denoising approach that combines NSST, BayesShrink thresholding, guided filtering, and method noise-based refinement process. The method shows remarkable efficacy in suppressing Gaussian noise while preserving edge and structural details, thereby improving diagnostic quality in low-dose CT images.