2009 Winter Project Week SegmentationAtlases
 Return to Project Week Main Page |
Key Investigators
- MIT: Michal Depa
- MIT: Koen Van Leemput
- MIT: Mert Rory Sabuncu
- MIT: Polina Golland
Objective
We are exploring different methods of constructing brain probabilistic atlases and their impacts on segmentation algorithms. These atlases contain for each voxel of the image the probabilities that it belongs to each of the labels being investigated. The goal is to compare the results of these algorithms when using probabilistic atlases created in different manners in order to establish guidelines for the appropriate choice in various situations.
Approach, Plan
Our approach is to use the model-based segmentation algorithm described in the MICCAI 2008 (Van Leemput et al) reference below with several different probabilistic atlases. These are traditionally created from a set of manually labeled training data by counting the relative occurrence of labels in corresponding locations in the training images. However, the use of more advanced methods could yield better results. In the model-based segmentation algorithm, atlases are formed using Bayesian inference (Van Leemput). A new method could involve creating a set of atlases for different modes of the population using clustering (M. R. Sabuncu et al).
Our plan for the project week is to discuss these different techniques and determine which data sets would yield the most interesting contrasts between them. If time permits, we would also like to start running experiments.
Progress
References
- K. Van Leemput et al. “Model-Based Segmentation of Hippocampal Subfields in Ulta-High Resolution In Vivo MRI.” MICCAI 2008.
- K. Van Leemput. “Encoding Probabilistic Brain Atlases Using Bayesian Inference.” IEEE Transactions on Medical Imaging 2008 (accepted).
- M. R. Sabuncu, S. K. Balci, and P. Golland. “Discovering Modes of an Image Population through Mixture Modeling.” MICCAI 2008.
