Motivation
How is the pool of adult Neural Stem cells maintained thoughrout life to keep producing neurons that will be integrated in the hippocampus? These newborn neurons are important for learning new memories, forgetting. Neurogenesis is activated by physical activity, espacially running in mice and inhibited in the case of depression or stress (antidepressor drugs promote neurogenesis).
Biologist have developped 2-photon microscopy methods to follow the fate of neural stem cells and their progeny in vivo. The bottleneck of the pipline is the image processing, We worked on developing image processing methods to automatically segment cells from the pictures.
Session 1
In the first session I told you about the experimental paradigm biologist are setting up to directly address the production of neurons in the adult hippocampus. They are using two photon microscopy to directly image neural stem cells and their progeny in living mice. From previous indirect experiments we have an idea about the rough genealogical relationships between Neural Stem cells, astrocytes, neuronal precursors, neuroblasts and neurons. The dataset of images allow us to directly address the question by observing the transitions (cf image of the first session).
The image dataset is composed of 63 time series (called spots) with dimensions x,y,z,t. Pictures are composed of 512x512 pixels (corrsponding to 0.5x0.5microns), 30 to 60 z (number variables betweens spots and also between time points within a spot, black frames have been added so that every time point of a serie has the same number of z), 4 to approximatively 18 time points, which makes a total number of cells of about 10 000 pictures all taken together.
Goals:
- segments cells from the background, on the 3D stack or on a smart projection of the z images
- classify cells based on their morphology to identify different types (cf Bonaguidi et al, review). First try some simple clustering methods (k-mean...)
- try to automatize the cell tracking in time
We agreed on first trying simple method of image processing. Someone pointed to the Point Cloud Library that might be useful to visualize the data in 3D and try out some algorithms already implemented.
We can also try some convolution neural networks.
Image sample
Example of a movie (Here all z planes have been collapsed). Images kindly provided by Gregor Pilz and Sara Bottes (HIFO, Zurich).
Review cell morphology (2.4 MB)
Clonal analysis initial paper (2.4 MB)
==============================================================================================================================
Methods for 3D segmentation:
- 1st method: simple method taking advantage of the image properties: high contrast. Uses exponential filter.
Drawback: uses K-mean clustering, so one needs to define beforehand how many cells we expect in the picture.
- 2nd method: gradient based (central difference gradient), find centroids in 3D
Improvements:
- improve the segmentation of the dendrites
- define a metric to compare the quality of the different methods compared to a human expert ground truth.
Next steps:
Train an SVN or a CNN to learn features specific for each cell type (Neural Stem Cells, precursors, neuroblasts, neurons)
Develop a GUI for the biologists to easily correct the automatically segmented cells
==================================================================================================
Original picture (maximum projection)
Results:
Method 1 : 3D segmentation only by thresholding
Method 2: gradient based segmentation
Segmented cells on the original picture
Segmented cells, edges are shown.
Individual segmented cells