PDF version of this entire document

next up previous
Next: Background Up: Data-Driven, Entropy-Based Measures for Previous: Data-Driven, Entropy-Based Measures for

Introduction

Over the past few years, non-rigid registration (NRR) has been used increasingly as a basis for medical image analysis. Applications include structural analysis, atlas matching and change analysis [5]. Many different approaches to NRR have been proposed, for registering both pairs and groups of images [3,19]. These differ in terms of the objective function used to assess the degree of mis-registration, the representation of spatial deformation fields, and the approach to minimizing the mis-registration with respect to the deformations. The problem is highly under-constrained and, given a set of images to be registered, each approach will, in general, give a different result. This leads to a requirement for methods of assessing the quality of registration.

Various methods have been proposed for assessing the results of NRR [8,10,16,15]. 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.

In this paper we present a method for evaluating the results of NRR that relies on the image data alone, and can thus be applied routinely without the need for ground truth. The method is based on the observation that, given a set of registered images, it is possible to construct a statistical model of appearance. If the registration is correct, this provides a concise description of the set of images. If it is incorrect, the performance of the model degrades. We base our assessment of the quality of registration on the quality of the resulting model, which can be evaluated using an entropy-based approach.

In the remainder of the paper we explore the background, explain the method in detail, and present validation results using data for which the correct registration is known.


next up previous
Next: Background Up: Data-Driven, Entropy-Based Measures for Previous: Data-Driven, Entropy-Based Measures for
Roy Schestowitz 2007-03-11