Data Mining Schemes for Medical Imaging

Abstract

In this research, we proposed a methodfor segmenting medical images using newlineSOM (Self Organizing Map) neural network. We than associate semantics to these newlineregions usingfuzzy reasoning. We have experimentedfor MRI (Magnetic Resonance newlineImaging) of brain images and digital mammogram images for breast cancer. The newlineexperimental data is drawn from the databases are available on the web named; newlineThe Whole Brain Atlas (database) given by Keith A. Johnson and J. Alex Becker, newlineand The Digital Database for Screening Mammography (DDSM) by Universityy of newlineSouth Florida. newlineA self organizing map is a well established unsupervised clustering property newlineconsisting of components called nodes or neurons. Pixels are clustered on the basis newlineof their grayscale and spatial features with a SOM network. Clustering separates newlinedifferent regions. These regions could be regarded as segmentation results newlinereserving some semantic meaning. Each node contains a corresponding weight newlinevector of same dimension. A random vector is chosen on every step of the learning newlineprocessfrom the initial data set and then the best-matching (the most similar to it) newlineneuron coefficient vector is identified. Select the winner which is most similar to the newlineinput vector. The distance between the vectors is measured in the Euclidean metric. newlineTrack the node which shows the smallest distance (this node is called as best newlinematching unit). Then update the nodes in the neighborhood of Best Matching Unit newline(BMU) by pulling them closer to the input vector. The result of neighborhood newlinefunction is an initial cluster center (centroids) for fuzzy c-means algorithms. newlineFuzzy c-means is a clustering method which allows to find the cluster centers. In order to accommodate the fuzzy partitioning technique, the membership matrir (U) is randomly initialized. This iteration will stop when the difference of update membership matrix and membership matrix is less than the termination criterion which lies between 0 and 1. newline