Cross Validation

Once a registration method has been validated to some degree of accuracy, it is possible to use the results of this method for cross-validating other methods. Although this sounds reasonable in principle, it is not an optimal validation strategy in practice. It is possible that the performance of the already validated method will not be as accurate as expected because of failure to reproduce the conditions of the original validation. In addition, except for blinded prospective studies, reporting biases and tuning of algorithms to the validation environment may cause reported accuracy of the already validated method to exceed the true accuracy. Finally, validation of the original method may not have been performed correctly, making the cross-validation inaccurate in a way that will likely reflect poorly on the newer method. These types of problems may accumulate across a chain of cross-validation studies to give rise to highly misleading results. Whenever possible, cross-validation should be limited to data sets registered with highly accurate methods that can reasonably be considered gold standards.

0 0

Post a comment