This research work addresses a major denoising problem in Magnetic Resonance (MR) Images. Magnetic Resonance Imaging (MRI) is a powerful and e ective di- agnostic tool in basic research, clinical investigation, and disease diagnosis since it provides both chemical and physiological information about the tissue. MR Images are a ected by Rician noise during acquisition phase which decreases the image quality, image analysis and becomes di cult to diagnose it accurately. This thesis is an attempt to suppress low and high categories of Rician noise from MR data in such a manner to enhance the diagnostically relevant image content. Supervised and unsupervisedltering techniques are applied to suppress the Rician noise hence improving its quality for diagnostic process. A new supervisedltering model, based on genetic programming (GP), is proposed that evolves an optimal composite mor- phological supervisedlter (FOCMSF ) by combining the gray-scale mathematical morphological operators. (FOCMSF ) is evolved through evaluating thetness of sev- eral individuals over certain number of generations. The proposed method does not need any prior information about the noise variance. In the domain of unsupervisedltering, three techniques are proposed. These are collaborative techniques based on statistical and fuzzy logic. Fuzzy similarity based non local meanslter (FSNLM) is designed to non-locally search out similar and non-similar regions of a noisy pixel. Fuzzy weights are assigned to these regions on the base of similarity. Then the noisy pixel is replaced with the fuzzy weighted average of these regions. Another hybridlter is proposed that combines FSNLM and local order statisticallters to suppress Rician noise. This hybridlter uses the strengths of non-local and locallters and adaptively calculates the fuzzy weighted estimation of the noisy pixels. Another non local fuzzy weighted Enhanced LMMSE (Linear Minimum Mean Square Estimator) is designed. The aim of this approach is to handle adaptively the low and high levels of variation of Rician noise and to estimate a closed-form of Rician distributed signal. It estimates the noise free pixel value based on similarity of the non-local neighborhood pixels around a window of certain prede ned radius. Similarity is computed using fuzzy logic approach which is served as fuzzy weights in enhanced LMMSE module for accurate estimation of noise free pixel value. The proposed schemes handle the problem with better accuracy than several well knownltering schemes NLM, LMMSE, Wavelet based techniques etc. and therefore can be considered as original contribution of this research work. The pro- posed schemes handle the problem of Rician noise at low and high noise variances on smooth as well as detailed regions where existing methods fail due to multifar- ious nature of this noise. The improved performance of the developedlters are investigated using the standard MRI dataset and its performance is compared with previously proposed state-of-the art methods. Detailed experimentation has been performed using simulated and real datasets based on well known quantitative mea- sures. Comparative analysis demonstrates the superiority of the proposed schemes over the existing techniques.