PDF version of this entire document

Summary and Discussion

This chapter showed how the problem of assessing NRR could be transformed into one of measuring model quality. It then introduced an approach which enables distances between images to be measured. It defined Specificity and Generalisation as measured of appearance model quality based on a measure of inter-image distance. It introduced a practical method of measuring inter-image distance and showed that, given distance measures and variety of images derived from a model, the quality of this model can be evaluated.

The measures used to evaluate models are Specificity and Generalisation. By considering images that are non-rigidly registered and building a model from them, it is then possible to assess NRR, using Specificity and Generalisation.

The next chapter utilises this new framework and explores its behaviour. Systematic experiments are used to show that the method works. Although the approach outlined above is principled, it still requires experimental validation and this can be achieved by measuring the change in model quality as the correspondence of an initially-registered set of images is progressively perturbed. The results are then compared with those obtained using a state-of-the-art assessment method, based on the overlap of ground-truth pixel/voxel labels. The results demonstrate that, not only is the proposed approach capable of assessing NRR reliably without ground truth, but that it also provides a more sensitive measure of misregistration than the overlap-based approach.

In Chapter 8, the new method is also applied to compare the performance of different registration algorithms on several sets of MR images of the brain, demonstrating that the method is able to discriminate between different methods of registration in a practical setting. The methods developed for measuring model quality are also valuable - in their own right - for other purposes. They can be used to compare different methods of model building [].