In this paper, we propose to segment liver tumor within ROI (Regions Of Interest) of MR (Magnetic Resonance) images by combining the smoothed shock noise reduction filter, that we introduced in a previous work, and a fuzzy approach. Besides the whole pipeline dedicated to our application, the originality of our method is to combine fuzzy information (membership degree to clusters) and geometrical features (Zernike moments) in the final defuzzification step. Thanks to a dataset representative of the variability of tumors' geometries and pixel intensities, we demonstrate that our approach is accurate, in every cases of liver lesions of our benchmark. We also show that the combination of both our two tools (denoising filtering and fuzzy segmentation) is the reason of the efficiency of this methodology.