PDF version of this document
We have introduced a model-based approach to assessing the accuracy of non-rigid registration, without the need for ground-truth. We have also described validation experiments where we progressively perturbed the initially good registration of a set of images, and found a monotonic relationship between our model-based measures and the degree of perturbation. We found that this behaviour was qualitatively identical to that obtained using a 'gold standard' method of assessment, based on the overlap of ground-truth anatomical labels associated with the images. A quantitative comparison of the two approaches demonstrated that one of the model-based measures, specificity, provides a more sensitive measure of misregistration than the overlap-based approach. This is not as surprising as it might seem at first sight, since the model-based approach uses all the full intensity information in the registered images, whereas the overlap-based approach uses a more impoverished representation of image structure. We tested different variants of the two approaches, and found that the model-based approach worked best when shuffle distance was used to measure separation in image space, whilst the overlap-based approach worked best when a label complexity weighting was applied.
These results are important, because they suggest that the performance of NNR algorithms can be compared objectively, using just the registered images they produce, and that the quality of registration can be assessed in routine applications of NRR, without the need for any additional information. It is important to note that, our approach does not depend on the specifics of the registration method used, or on the particular form of generative model constructed from the registered data - it can be applied to the results of registration, whatever the NRR algorithm used, and different forms of generative model could easily be substituted.
Acknowledgements: The authors would like to thank David Kennedy of the Centre for Morphometric Analysis at MGH for segmented brain data. The work was supported by the EPSRC/MRC-funded Medical Image and Signal IRC (GR- /N14248/01), Integrated Brain Image Modelling (EPSRC GR- /S82503/01) and Modelling, Understanding and Predicting Structural Brain Change (EPSRC GR/S48844/01).