The proposed non-linear image distortion method achieves considerable classification improvements on the used corpus of medical radiographs. The improvements are due to the introduction of local dependencies into the distortion model and the use of context information by using image gradients and local image regions. It is interesting to see that the use of the context information alone already improves the classification considerably, especially for the simple zero-order model. The results are improved by taking into account local dependencies without restricting the possible pixel displacement fields too strictly.