Contents
Understanding patterns and discovering associations, regularities and anomalies between anatomical structures and normal or abnormal function of the human brain is a fundamental goal in the neuroscience community. Current advances in brain image acquisition techniques have made available enormous amounts of remarkable high-resolution three-dimensional (3-D) image data. The availability of this data has already facilitated many advances in human brain mapping during the last decade. In addition to the continuous development of improved brain imaging techniques, greater computer capabilities and improvements in normalization techniques are leading to the creation of large databases of structure/function information. The analysis and exploitation of such large collections of medical image data still remains a problem though. Major issues in the current attempts for managing this data are the efficiency, effectiveness and robustness of the database and data mining tools used to extract knowledge (in the form of patterns, associations, etc). New tools for content-based (similarity) retrieval, association mining, classifications, etc, can have significant impact in this endeavor. Although considerable progress has been made in content-based image retrieval and classification for general types of images, progress in medical images and in particular in brain images has been very slow since global signatures that are usually employed in the former do not work in the brain imaging domain where the regions of interest occupy a small portion of the image.
The goal of this project is to overcome these problems and address the great need for developing efficient brain data mining tools for the analysis and management of large collections of brain images (from various imaging modalities) and associated clinical data. These automated tools enable interoperable brain image data representation that is easy to search. The main focus is on the management of the spatial regions of interest (ROIs). We are developing a general unified framework for managing ROIs regardless of whether these are lesions, tumors, areas of brain activation, or regions of (normal/abnormal) morphological variability of a variety of brain structures.
The first specific aim focuses on developing efficient methods for feature extraction and classification of ROIs in brain images. The second specific aim focuses on developing fast and effective database techniques supporting efficient retrieval of similar regions of interest in large brain image databases as well as spatial data mining tools for discovering associations between anatomic and other variables such as function, pathology, or response to drugs. Our third specific aim focuses on the integration of the above techniques with morphological analysis tools to correlate morphological changes to changes of other measurements such as functional, physiological, etc. We plan to evaluate and validate the classification, similarity searching and data mining techniques using real and simulated data and to demonstrate their utility in the analysis of large data sets from a number of epidemiological studies of brain morphology and function.
We are developing novel brain data mining techniques that are efficient, effective and robust. We expect that this work will advance our ability to analyze 3-D brain image data and to discover associations between spatial patterns, anomalies or normal variations and other non-spatial data. Focusing on the ROIs (e.g., lesions, tumors, areas of brain activation, areas of morphological variability) we work on analyzing their spatial characteristics (such as spatial distribution, shape, etc) and on developing effective and efficient methods for characterization, classification, and content-based retrieval of 3-D brain image data, and spatial mining tools for determining associations between structural and functional data obtained using brain imaging techniques and other variables obtained through clinical assessment. Efficiency here is very important; the methods should be scalable so that they can be applied to the analysis of very large data sets coming from multiple studies.
At this very early stage of development, as this work is newly funded, we have completed a preliminary implementation of database tools that allow efficient ROI similarity searches and classification. The similarity queries are handled by extracting features from ROIs, mapping them to a k-dimensional (feature) space and calculating the nearest neighbors of a certain query ROI in feature space. We have also developed preliminary tools for reduction of the dimensionality of this space. Based on these tools we can easily select the most discriminative features making efficient the retrieval of nearest neighbors.
2. Q. Wang, D. Kontos, G. Li and V. Megalooikonomou, ``Application of Time Series Techniques to Data Mining and Analysis of Spatial Patterns in 3D images'', in Proceedings of the International Conference on Acoustics, Speech and Signal Processing, (ICASSP) 2004, (to appear).
3. D. Kontos, V. Megalooikonomou, F. Makedon, ``Computationally Intelligent Methods for Mining 3D Medical Images'', in Lecture Notes in Artificial Intelligence, 3rd Hellenic Conference on Artificial Intelligence, Samos Island, Greece, May 2004, (to appear).