PDF version of this document

Introduction

Non-rigid registration (NRR) of both pairs and groups of images has in recent years increasingly been used as a basis for medical image analysis. Applications include structural analysis, atlas matching and change analysis [5]. The problem is highly under-constrained and a host of algorithms [4,18] that have become available will, given a set of images to be registered, in general produce different results.

Various methods have been proposed for assessing the results of NRR [8,10,15,14]. Most of these require access to some form of ground truth. One approach involves the construction of artificial test data, which limits application to 'off-line' evaluation. Other methods can be applied directly to real data, but require that anatomical ground truth be provided, typically involving annotation by an expert. This makes validation expensive and prone to subjective error.

We present two methods for assessing the performance of non-rigid registration algorithms. One of the methods requires ground truth to be provided a priori, whereas the other does not. We compare both methods on a registration of a set of 38 MR brain images and show them to provide a robust evaluation of registration success. Moreover, we demonstrate that both methods are in fact closely correlated if not interchangeable.