Medical Image Denoising Based on Implementing Multi-Resolution Threshold Technique

Abstract

Medical image denoising is critical for accurate diagnosis and quantitative analysis, yet preserving fine anatomical detail while suppressing noise remains challenging. This paper proposes a multi-resolution thresholding (MRT) technique for medical image denoising that combines discrete wavelet transform (DWT) based decomposition with adaptive, scale-aware threshold selection and a post-processing spatial regularization step. The input image is decomposed into multiple resolution sub bands using a suitable orthogonal wavelet (e.g., Daubechies or Symlet), separating high-frequency noise components from low-frequency structural content. For each detail sub band, thresholds are determined adaptively using a hybrid criterion that blends robust noise estimation (median absolute deviation) with local signal activity measures, enabling preservation of edges and fine textures in anatomically important regions. Both soft and semi-soft thresholding rules are explored to reduce visual artifacts and Gibbs-like distortions. After inverse wavelet reconstruction, a multi-scale bilateral filter is applied selectively to remaining residuals to further suppress speckle and salt-and-pepper noise while conserving boundaries. The proposed MRT framework is evaluated on MRI and CT datasets corrupted with synthetic Gaussian and real-world noise. Quantitative results show consistent improvements in peak signal-to-noise ratio (PSNR) and mean square error (MSE) compared to classical denoising methods (Wiener, anisotropic diffusion) and several baseline wavelet-thresholding schemes, while visual assessment confirms superior edge preservation and lesion conspicuity. Computational complexity is modest and compatible with near real-time preprocessing on modern CPUs; the method also scales to 3D volumes by applying separable 3D wavelet transforms. MRT is therefore a practical, effective approach for clinical imaging pipelines where maintaining diagnostic features is essential.