Magnetic Resonance Imaging (MRI) is a technique used to assess various regions of the body and is useful in medical diagnosis. Because of this, it is important to maintain the clarity of the MRI images that are often degraded by different noises, particularly Gaussian and Rician. In this paper a comprehensive evaluation of different denoising methods for brain MRI images, including classical, hybrid, and deep learning methods is conducted. The methods evaluated include BM3D (Block-Matching and 3D Filtering), PCA (Principal Component Analysis)-based denoising, PCA combined with median filtering, a proposed hybrid approach (combining PCA, median filtering, and bilateral filtering with Rician bias correction), and DnCNN (Denoising Convolutional Neural Network) trained on custom MRI image patches. The metrics used to assess these denoising methods include PSNR (Peak Signal to Noise Ratio), SSIM (Structural Similarity Index), and histogram-based metrics like Bhattacharya distance, intersection, and correlation. In summary, the DnCNN method shows the best results across all noise types and levels, demonstrating high structural preservation. However, the proposed hybrid method also demonstrates competitive results compared to DnCNN and better results in comparison to classical methods like BM3D and PCA-based denoising methods. From this comparison, it can be concluded that while the deep learning-based method is overall the best for denoising, the hybrid method can also be seen as an alternative in cases of limited training resources.
S. et al. (Mon,) studied this question.