Скачать или смотреть The K-means clustering added to FluoRender for segmentation and tracking

We added another commonly used clustering method,
K-means to FluoRender for segmentation.
This video demonstrates the testing of the new method.
Currently, the cluster number is always 2.
It is able to generate acceptable segmentation results for
most test cases.
In the example, cells were segmented using thresholding only.
Some of them were under-segmented because a good
threshold value could not be easily determined.
K-means is fast to calculate and guaranteed to
generate the desired number of clusters.
However, the segmented structures may not follow
their natural boundaries.
Although estimating the cluster number for K-means
has been a challenge for the method itself.
We can set the cluster number based on tracking
information when this method is fully integrated.
For example, when a track splits.
It could be either a mitosis event, or the splitting
structures can be merged because of over-segmentation.
When two or more tracks merge, it is most likely to be
an under-segmentation.
We would actually hope to cluster the under-segmented
structure based on the track number from previous time point.
Furthermore, we can apply other more sophisticated
clustering method on the K-means results.
Therefore, a graph-based framework for multiconstraint
4D tracking can be built for more accurate and intuitive results.