Medical Image Data Base Shape Indexing and Pathology Classification

The purpose of this research is to develop accurate medical pathology classification procedures and efficient indexing methods for medical data-bases. More specifically, we dealt with dental X-ray data, and we have extracted pathology shape descriptors using modal analysis, a physics-based simulation method for extracting the modes of deformation of a shape.  

• Following are the four pathology classes of interest, in their progressive evolution:

Classification results (features selected based on divergence criteria)

                                         Minimum Mahalanobis distance classifier               

                                              K-nearest neighbour classifier

Indexing in high-dimensional space (randomized divide and conquer)  

• We have tested a randomized divide and conquer geometric algorithm for k-nearest neighbour search. The results are good for spaces of moderate dimension (10-14). In the following, we plot the search performance for different dimensions and different number of data-base entries. The comparion is performed against linear search (dashed line).