Sohan Rashmi Ranjan MS, Presenter: Employee, General Electric Company

Rakesh Mullick PhD, Abstract Co-Author: Employee, General Electric Company

Manasi Datar, Abstract Co-Author: Employee, General Electric Company

Cathleen Cooper, Abstract Co-Author: Employee, General Electric Company

The goal is to define a strategy for validating image registration methods in the absence of ground truth, which are hard to get. Usually manual registration done by an expert is used for validation of an automated registration method. However there is an inherent inter- and intera- variability associated with manual registration process. Thus, it is important to ascertain the source of validation error when the margin of error is of the order of a voxel. Our approach is to use the notion of consistency in registration transforms as an evidence in support of a registration result.

In each of our study, three different images of a patient of same or different modalities were selected. Registration was performed, both manually by an expert and by our automated method, for the following image pairs: first image with the second image, second image with the third image, and first image with the third image. As registration is a transitive process, the result for registering the first image with the third image should be a combination of results for registering the first image with the second image, and the second image with the third image. Thus a check on the consistency of results may be used as an evidence in support of accepting accuracy of the method. For rigid registration, a registration result comprises of translation and rotation components. We demonstrate this approach for validating our automated rigid registration method by comparing corresponding results with manual registration results. Twenty sets of such triplets were identified in our validation studies of over 400 cases. Both manual and automated registration were performed, and results compared.  

On available results, the order of translation and rotation consistency error by automated registration and manual registration were 0.5 voxel and 1.5 voxels, and 0.05 and 0.12 degrees, respectively. The elaborate assessment of the registration is currently in-progress and will be part of the formal presentation.

Consistency in registration transform may be used as an evidence of registration accuracy. This assertion is useful in registration validation, as ground truth are usually not available.

Validation of registration method used in therapy and planning is crucial. An evidence to support registration result in the absence of ground truth may thus be of immense value to a practitioner.

Ranjan, S, Mullick, R, Datar, M, Cooper, C, Significance of Consistency in Registration Results in Validation Studies.  Radiological Society of North America 2008 Scientific Assembly and Annual Meeting, February 18 - February 20, 2008 ,Chicago IL. http://archive.rsna.org/2008/6020587.html