A Review and Experimental Analysis of Denoising Techniques for Medical Images

Magnetic Resonance Imaging (MRI) and High-Resolution Computed Tomography (HRCT) are crucial for comprehensive diagnosis and treatment planning, as they provide detailed anatomical information. However, noise introduced during image acquisition often degrades the quality of these images, obscuring key anatomical features and complicating accurate diagnoses.

This study compared the performance of eight denoising algorithms: BM3D, EPLL, FoE, WNNM, Bilateral, Guided, NLM, and DnCNN. Both objective metrics, including Mean Squared Error (MSE), Structural Similarity Index (SSIM), and Peak Signal-to-Noise Ratio (PSNR), as well as perceptual quality metrics, such as NIQE, BRISQUE, and PIQE, were employed to assess their effectiveness.

BM3D consistently outperformed other algorithms at low and moderate noise levels, achieving the highest PSNR and SSIM values while preserving structural integrity and perceptual quality. For high noise levels, conventional algorithms, such as EPLL and WNNM, demonstrated competitive performance in homogeneous areas, preserving fine texture, but were limited by computational complexity.

One of the challenges in image denoising is preserving the finer detail structures of images while efficiently removing noise. Finding a balance between the reduction of noise and preservation of image integrity can be a lifesaving challenge, especially in cases where the images are in high detail, such as in the medical world.

This study highlights the trade-offs between denoising quality and computational efficiency among various algorithms for MRI and HRCT images. While BM3D remains a dependable choice for moderate noise levels, advanced deep learning-based methods, such as DnCNN, are better suited for handling significant noise variations without compromising critical diagnostic features.