Simultaneous Registration, Segmentation and Modelling of Structure in Groups of Medical Images

Abstract

We address the problem of extracting information from groups of medical images of the same anatomy. We describe an algorithm which simultaneously segments and registers a set of such images, incrementally constructing a model of their structure and the correspondences across the set. The framework explicitly models the fraction of each tissue type, rather than the expected intensity in each voxel, to decouple the model from details of the imaging sequence and modality. When estimating the optimal deformation field, the current image is compared to a reconstructed image, generated from the model tissue fractions and the current estimate of intensity distributions for each tissue type in the current image (i.e. an estimate of how the model would appear given the imaging conditions for that image). We describe the algorithm in detail and present results of applying it to a set of MR images of the brain.

Brain Image Segmentation with Gradient Information

“…Figure 2), with the calculation for image and tissue is as follows, Where and run over all pixels in the image. The fraction images from the 2D histogram method are then compared to those from the 1D histogram method [Petrovi¢ et al, 2007], by using Tanimoto overlap values with ground-truth [Crum et al, 2006] of each image on the 37-dataset. This table shows the 2D method has higher mean values than the 1D method for all tissue fractions, as well as smaller standard deviation for grey matter and CSF (see e.g.…”

“…Alternatively, as the value of at a particular intensity can be chosen to maximize the mixture-Gaussian probability over the whole range of [Petrovi¢ et al, 2007]:…”

“…However, if prior knowledge in terms of an atlas is not available, then a method that does not require training data for segmentation is necessary If we consider the classes of segmentation algorithm described in [Pham et al, 2000], then in terms of automated methods that do not require training, the only candidate is clustering-type methods. For dense data, clustering becomes in effect a Gaussian mixture model, hence we come to the Gaussian mixture model methods as used by Petrovic et al [Petrovi¢ et al, 2007]. These methods are more sophisticated than the simple Gaussian mixture models, since they also consider partial-voluming effects.…”