Reducing the speckle noise in ultrasound images plays an important role in image post-processing, e.g. image segmentation, edge detection and image registration. In this paper, a novel filtering method based on extreme learning machine (ELM) is proposed for ultrasound image denoising. We aim at generating a denoising model learned from the input samples. In the training step, pixel features from the noisy image and the target values from the original image are fed to ELM to learn a speckle reduction model. In the testing step, we obtain the denoising image via the learned model. A pilot study is performed to compare the proposed algorithm with other classical denoising methods using brain tumor ultrasound images. The experimental results show that our proposed method outperforms other methods in reducing speckle noise, and is effective to preserve image details.