We describe a method to reduce speckles of ultrasound images. We first segment an ultrasound image with speckle segmentation algorithm. We compute an adjacency graph representation of the speckles. For each speckle region, we compute its area, mean and variation. Based on the simularity (Fisher distance) of the adjacent regions, we merge neighboring regions into larger regions. The merge operation is conveniently mapped to graph operations and holes are avoided. The merge process iterates until the graph is stable. We then assign for each region a grayscale equals to the mean. This method have the advantage that it provides region segmentation of the ultrasound images as well as speckle reduction without loss of the edge contrast.