Registration is an important poblem in medical image processing and has been quite well studied. However, in most applications, we need to know more than just the value of the result: we need to evaluate its accuracy. In the statistical framework, this means going from first order to second order statistics.
We have developped a theoretical framework for handling uncertainty on geometric features such as points, oriented points or frames. In this demo, we show how it applies to a wide range of problems in medical imaging.
To apply our statistical framework, we extract from the 3D images features that are supposed to contain most of the geometric information.
To assess a statistical significance to our results, we need to register a sufficient number of images in roughtly the same conditions.
With 24 images of the same patient, we are able to study the stability of extremal points across observations. The ultimate goal would be to correlate some of these points with some known anatomical landmarks on the brain. We present here a preliminary study.
We have two MRI images of the same pelvis in two different configurations. The goal is to determine if there is a relative motion of the bones in the pelvis.
We have shown that the determination of second order statistics in medical images processing not only possible but necessary to ensure a correct interpretation of measurements. We have developped a generic framework to tackle these types of problems but there is still some work to be done for the matching and multiple registration parts, as for the uncertainty of curve, surface or volume based registration.
A more general framework allowing a non-rigid group action would also be very interesting for the development of statistics for computerized atlases.