Comparative Evaluation of Fast and Adaptive Non-local Means Algorithms for Rician Noise Reduction in Breast Magnetic Resonance Imaging

초록

Background: This study systematically compared the performances of fast non-local means (FNLM), conventional non-local means (NLM), and adaptive non-local means (ANLM) algorithms for Rician noise reduction in clinical breast magnetic resonance imaging (MRI). Materials and Methods: Rician noise with standard deviations of 0.05, 0.10, and 0.15 was synthetically introduced into pre-contrast T1-weighted breast MRI images obtained from 50 patients in a publicly available clinical dataset. For each noise level, the FNLM search window size was optimized using a root mean square error (RMSE)-based tuning procedure. The optimized FNLM was then quantitatively compared with NLM and ANLM. Image quality was assessed using RMSE, structural similarity index (SSIM), high-frequency error norm (HFEN), gradient magnitude similarity deviation, and edge preservation index (EPI). Computational efficiency was evaluated in a MATLAB (MathWorks) environment using central processing unit-based processing. Results and Discussion: The relative performance of FNLM varied according to noise level and evaluation metric. Compared with ANLM, FNLM achieved lower RMSE and HFEN and higher SSIM and EPI across most noise levels, while showing comparable or improved performance relative to NLM. Linear mixed-effects analysis confirmed significant algorithmic differences depending on noise severity. Regarding computational efficiency, FNLM was approximately 7.8–25.1 times faster than ANLM and 2.7–3.2 times faster than NLM across noise levels. Conclusion: The optimized FNLM algorithm provides competitive denoising performance while substantially improving computational efficiency in clinical breast MRI with Rician noise.

키워드