Data Mining in Large Medical Image Databases

Abstract

The increasing amount of data stored in large databases leads the task of knowledge discovery, based on data mining techniques, to become an essential task to enable strategic decisions. However, the data mining techniques required by knowledge discovery tasks have high computational cost and impose complex management burdens over the analyst. The complexity derives from the diversity of tasks that may be applied in the data analysis and from the large amount of alternatives to execute them. The most common data mining tasks include data classification and grouping, outliers and exception identification, missing data prediction and dimensionality reduction. The high computational cost comes from the need of the algorithms to explore several alternative solutions, in different combinations, in order to obtain the desired information. Traditionally, the data are represented by numeric or categorical attributes in a table that represents one individual of the analyzed set in each tuple. Although the same tasks applied to traditional data are also necessary for images, the analysis complexity and the computational cost increase considerably, making applying the traditional techniques impractical. Therefore, special data mining techniques for large image sets need to be developed. This project aims at the development of new data mining techniques for image sets, an in particular at the data classification and grouping tasks. The work will be based on the Fractal Theory, and will explore the main features that differentiate images from traditional data with respect to their impact in data mining tasks. The Fractal Theory has been successfully employed in both individual image analysis and traditional data sets analysis, but not in large image sets, so this will be a novel research project that, based on the previous research activities of the involved group, has a high likelihood to earn interesting results.