The validation of medical image processing algorithms is usually difficult due to the absence of known ground truth. As the real value of the parameters being estimated by the algorithms is not known, one as to refer to a reference value (or gold standard) to assess the computation when such a reference is available. A solution that has been recently proposed is to establish a bronze standard [1] by considering the "exact result" as an unknown variable that has to be estimated (along with its accuracy).

This method is applied on the assessment of medical image registration algorithms that are of high interest and widely developed. The method is based on the registration of all possible pairs of images by many registration methods (different from the one to evaluate) in order to better exploit the redundancy of information. We are using four different registration algorithms in our implementation of the bronze standard method: (1) Baladin and (2) Yasmina are intensity-based. The former uses a block matching strategy while the later optimizes a similarity measure on the complete images using the Powel algorithm. (3) CrestMatch is a prediction-verification method and (4) PFRegister is based on the ICP algorithm. Both CrestMatch and PFRegister register features (crest lines) extracted from the input images. These algorithms are further described in [1]. All produce a transformation and the bronze standard is computed from these results.

Medical image registration is rather compute intensive (in the order of minutes per pair of 3D images to register). The complexity of the bronze standard assessment method comes from the fact that the more pair of images and the more algorithms are available for registration, the more accurate are the results. A typical run involves registering a database of tens to thousands of pair of images through a couple of algorithm (each one may be used with different parameter sets). The amount of data to manipulate and the cost of computations is out of reach of standard computers. Yet the application is rather easy to distribute over a grid as the application workflow shows a lot of inherent parallelism, especially data parallelism.

The application workflow is described in Scufl and the computations are orchestrated through the Taverna workflow manager (MyGrid project) or MOTEUR, our home made workflow engine. Both engines have been interfaced [2] with the EGEE middleware.

[1] "Evaluation of a New 3D/2D Registration Criterion for Liver Radio-Frequencies Guided by Augmented Reality", S. Nicolau, X. Pennec, L. Soler, and N. Ayache, Intl. Symp. on Surgery Sim. and Soft Tissue Model., pp 270-283, Juan-les-Pins, France, 2003.

[2] "Grid-enabled workflows for data intensive applications". T. Glatard, J. Montagnat, X. Pennec, Computer Based Medical Systems (CBMS'05), special track on Grids for Biomedicine and Bioinformatics, Dublin, Ireland, June 23-24, 2005.