Abstract
We propose a new similarity measure, Combined Connectivity and Spatial Adjacency (CCSA), to be used in hierarchical agglomerative clustering (HAC) for automated segmentation of Self-Organizing Maps (SOMs, Kohonen [1]). The CCSA measure is specifically designed to assist segmentation of large, complex, functional image data by exploiting general spatial characteristics of such data. The proposed CCSA measure is constructed from two strong indicators of cluster structure: the degree of localization of data points in physical space and the degree of connectivity of SOM prototypes (as defined by Tasdemir and Merényi [2]). The new measure is expected to enhance cluster capture in large functional image data cubes such as hyperspectral imagery or fMRI brain images, where many relevant clusters exist with widely varying statistical properties and in complex relationships both in feature space and in physical (image) space. We demonstrate the effectiveness of our approach using the CCSA measure on progressively complex synthetic spatial data and on real fMRI brain data.
Original language | English (US) |
---|---|
Title of host publication | 12th International Workshop on Self-Organizing Maps and Learning Vector Quantization, Clustering and Data Visualization, WSOM 2017 - Proceedings |
Publisher | Institute of Electrical and Electronics Engineers Inc. |
ISBN (Electronic) | 9781509066384 |
DOIs | |
State | Published - Aug 29 2017 |
Event | 12th International Workshop on Self-Organizing Maps and Learning Vector Quantization, Clustering and Data Visualization, WSOM 2017 - Nancy, France Duration: Jun 28 2017 → Jun 30 2017 |
Other
Other | 12th International Workshop on Self-Organizing Maps and Learning Vector Quantization, Clustering and Data Visualization, WSOM 2017 |
---|---|
Country | France |
City | Nancy |
Period | 6/28/17 → 6/30/17 |
ASJC Scopus subject areas
- Artificial Intelligence
- Computational Theory and Mathematics