Atlases are widely used to aid in Brain segmentation or tissue classification, they are typically constructed by summarizing information concerning image intensities and anatomical shapes from a set of training images that includes manual segmentations. Conventional atlas building methods rely on the averaging process of all training images, which can lead to significant loss of information. Recent research shows that selecting a subset of atlases that are similar to the test subject provides more accurate segmentation than those given by randomly selected atlases.
We proposed a fast,shape-based hierarchical matching approach based on Spatial Pyramid Matching (SPM) to perform a very efficient k-NN search for testing brains. We comparied the proposed method against known methods for k-NN lookup and evaluated the atlases built on these methods on a training set contains 259 brain MRIs and a testing set of 20 brain MRIs.
P. Zhu, S.P. Awate, S. Gerber, R. Whitaker. "Fast Shape-Based Nearest-Neighbor Search for Brain MRIs Using Hierarchical Feature Matching " .In Medical Image Computing and Computer-Assisted Intervention (MICCAI 2011), vol. 6892, pp. 484--491. 2011.
The goal for this project is to find geological structures or objects that indicate the presence of reservior of valuable natural resources like oil and gas. The geological structures we are interested in include faults, channels, domes,etc. Because of the complexity and variablity of these structures, it usually takes huge amount of time for the geologists to interpret the seimic data to find them.
In this project, we develop a framework to automatically detect the geological structures, using PCA based/mahalanobis distance based outlier/novelty detection methods. These interesting structures are detected as abnormalises which can be separated from the normal horizons or strata that consist of the backgound. On top of that, we propose a hierarchical clustering-based pipeline for geological structure recognition, which gives a stastical models for the target structures in terms of configuration of codes/labels, and then using the models to detect new objects of the same category in seismic data.