Post on 14-Jun-2015
description
Fuzzy C-Means Based Liver CT Image Segmentation with Optimum Number of
Clusters
The 5th International Conference on Innovations in Bio-Inspired Computing and Applications. June 23-25, 2014
Abder-Rahman Ali
The 5th International Conference on Innovations in Bio-Inspired Computing and Applications, June 23-25, 2014
Overview
Motivation Proposed Approach Optimal number of clusters Average Generalized Silhoeutte Results
Conclusion
The 5th International Conference on Innovations in Bio-Inspired Computing and Applications. June 23-25, 2014
Motivation
We investigate the effect of using an optimum number of clusters with Fuzzy C-
Means clustering, for Liver CT image segmentation
Is the optimum always the better?The 5th International Conference on Innovations in Bio-Inspired Computing and Applications. June 23-25, 2014
Proposed Approach
The 5th International Conference on Innovations in Bio-Inspired Computing and Applications. June 23-25, 2014
Optimal Number of Clusters
Generalized Intra-Inter Silhouettes
The 5th International Conference on Innovations in Bio-Inspired Computing and Applications. June 23-25, 2014
Optimal Number of Clusters (cont…)
Generalized Intra-Inter Silhouettes
The 5th International Conference on Innovations in Bio-Inspired Computing and Applications. June 23-25, 2014
Optimal Number of Clusters (cont...)
Generalized Intra-Inter Silhouettes
The 5th International Conference on Innovations in Bio-Inspired Computing and Applications. June 23-25, 2014
Compactness distance
Optimal Number of Clusters (cont...)
Generalized Intra-Inter Silhouettes
The 5th International Conference on Innovations in Bio-Inspired Computing and Applications. June 23-25, 2014
Separation distance
Optimal Number of Clusters (cont...)
Generalized Intra-Inter Silhouettes
The 5th International Conference on Innovations in Bio-Inspired Computing and Applications. June 23-25, 2014
silhouette
[-1, +1]
If bi-aj +: good clustering
If bi-aj - : poor clustering
Average Generalized Silhouette
The 5th International Conference on Innovations in Bio-Inspired Computing and Applications. June 23-25, 2014
• Average Generalized is considered in this work, since we are interested in the overall clustering quality of the entire dataset
• Average Generalized Silhouette returns a vector of silhouette values, one value for each data point (pixel)
• If one point has a silhouette value near 1, then its clustering is very good
Average Generalized Silhouette (cont…)
The 5th International Conference on Innovations in Bio-Inspired Computing and Applications. June 23-25, 2014
• If the silhouette is near -1, then the clustering of the point is very bad
• A silhouette value of 0 indicates an intermediate case
• Each silhouette is considered a measure of the clustering quality of the associated point
Results
The 5th International Conference on Innovations in Bio-Inspired Computing and Applications. June 23-25, 2014
The first and second images, the best number of clusters to be used is 3. For the third image, the best number of clusters to be used is 4. And, for the fourth and fifth images, the best number of clusters to be used is 2
Results (cont…)
The 5th International Conference on Innovations in Bio-Inspired Computing and Applications. June 23-25, 2014
where the optimal number of clusters to be used are larger than 2-clusters, based on Table (1) in the previous slide, they gave the best Jaccard index values. And, where the optimal number of clusters to be used are 2-clusters, choosing a random number of clusters in the range 3-5 gave better Jaccard index values
Results (cont…)
The 5th International Conference on Innovations in Bio-Inspired Computing and Applications. June 23-25, 2014
• The figure represents image (1) from the table
• Using 3-clusters, as recommended by the average silhouette value, shows more clearly the groundtruth than using 2-clusters
Conclusions
The 5th International Conference on Innovations in Bio-Inspired Computing and Applications. June 23-25, 2014
Choosing the correct number of clusters is very important in Fuzzy C-Means clustering
it was noticed that it is not always necessary that using the optimum number of clusters with FCM, as measured by the average silhouette value, always gives the best results in terms of Jaccard index
Thanks and Acknowledgement
The 5th International Conference on Innovations in Bio-Inspired Computing and Applications. June 23-25, 2014
http://www.egyptscience.net
Authors: Abder-Rahman Ali, Micael Couceiro, Aboul Ella Hassenian, Mohamed F. Tolba5, and Vaclav Snasel