Search

We investigate incorporating structural information from segmentation into a groupwise registration framework. Previous work by Petrovic et al., using MR brain images, showed that using tissue fractions to help construct an intensity reference image gives better results than just using intensity images alone. In their work, a Gaussian Mixture Model (GMM) was fitted to the 1D intensity histogram, then used to construct tissue fraction images for each example. The mean fraction images were then used to create an artificial intensity reference for the registration.By using only the mean, this discarded much of the structural information. We retain all this information, and augment each intensity image with its set of tissue fraction images (and also intensity gradient images) to form an image ensemble for each example. We then perform groupwise registration using these ensembles of images.This groupwise ensemble registration is applied to the same real-world dataset as used by Petrovic et al. Ground-truth labels enable quantitative evaluation to be performed. It is shown that ensemble registration gives quantitatively better results than the algorithm of Petrovic et al., and that the best results are achieved when more than one of the three types of images (intensity, tissue fraction and gradient) are included as an ensemble.

The 2009 MICCAI Head and Neck Auto-Segmentation Challenge was aimed at assessing algorithms for automated segmentation of the brain stem and mandible from CT images for use in radiotherapy planning. The organisers provided 10 datasets with associated manual segmentations by experts, and 8 test datasets on which the algorithms had to be run. In two of the test datasets (subjects 13 and 15) the field of view was truncated. The results were independently analysed by the organisers.To address the problem posed by the challenge we describe a method based on active appearance models (AAMs) which involves four stages: Initialisation with a parts-and-geometry model, search with a global AAM followed by search with local AAMs, then post-processing using linear regressors.Application of the method to the test images resulted in a mean (excluding subject 13) Dice overlap value of 76.1 +/- 5.1% for the mandible, and 72.9 +/- 24.6% for the brainstem. The method failed to segment both the mandible and brainstem in subject 13, and the whilst giving a successful segmentation of the mandible in subject 15, gave poor results on the brainstem. It takes about 10 minutes to run on a 64 bit Linux workstation with 2GB RAM, and a 2GHz pentium processor. We were encouraged by its performance on this dataset, and believe that its accuracy can be improved with some modifications to the segmentation pipeline.