Ultrasound imaging can help in choosing the needle trajectory for epidural anesthesia but anatomical features are not always clear. Spatial compounding can emphasize structures; however, features in the beam-steered images are not aligned due to varying speeds of sound. A non-rigid registration method, called warping, shifts pixels of the beam-steered images to best match the reference image. Linear prediction is used to find the warping vectors and decrease computational cost. An adaptive median-based combination technique for compounding is also investigated. The algorithms are tested on a spine phantom and human subjects. The results show a significant improvement in quality when using warping with adaptive median-based compounding.