Using SOM-based clustering to interpret optic nerve images.

摘要

An unsupervised Self-Organizing Map (SOM) based clustering technique has been applied to model the inherent classifications and sub-classifications of the optic nerve images obtained by Confocal Scanning Laser Tomography (CSLT) technology. One of the significant challenges in this study is to develop sophisticated approaches to facilitate the visualization, analysis, and tracking in time order of optic disc images. In our study we present a data mining framework that uses a combination of data clustering techniques (SOM and Expectation-Maximization) to group both normal control and glaucomatous optic disc data into meaningful clusters.;We present our results and conclude that our approach provides a good understanding of the inherent relationships among the morphological features of CSLT images as well as Zernike moments extracted from CSLT images of optic discs. Furthermore, our approach is not only capable of finding meaningful clusters but also identifying noisy images within a sequence of images, and a potential tool to track the time-series activity (progression) of glaucoma. We conclude that a study of the emergent clusters of patient may enhance our understanding of morphological features of optic nerve damage and may also lead to more informed clinical care of patients with glaucoma.