Low Noise Electrocardiogram (ECG) has been widely used for heart disease diagnosis. The anisotropic median-diffusion is the filter obtained by intercalating a median filtering in each diffusion step. We propose to use anisotropic median-diffusion to filter noisy ECG signals. We describe how to estimate appropriate parameters of the proposed filter. We validate our method using ECG signals from the MIT-BIH databases (many of them with premature ventricular contraction) and compare our method with other filtering methods. Experiments show that the proposed technique can effectively remove the noise without changing the instants and amplitudes of events, as well as preserving the morphologies of ECG signals in sections of the QRS complex.