hashemi_071306.ppt

Progress Report II

Mahboubeh Hashemi

Taiwan, Hsinchu (NCHC)July 14th 2006

Project Proposal

In order to create a three dimensional map of neural networks and their functional links the location of cells relative to each other needs to be determined. It is optimal to be able to use computational techniques, such as image processing software, as oppose to manually determining the location of these cells.

The current image processing software, designed and developed by a group in Taiwan, Dr. Lin and his team, uses flow solutions in detecting shapes. However, the application and accuracy of the software as a tool in biological experiments needs to be examined. Ultimate goal would be to understand the software, use images and movies of cells firing up in a network to run the code and analyze and interpret the data.

Goals Accomplished

➢ Searched for several algorithms for boundary detection and have made a list of the different possibilities to approach this problem (identifying the boundary and location of biological cells).

➢ Among those I am currently working on the Converging-Squares-Algorithm (this method is explained in the next couple of slides).

➢ Familiarized myself with C++ programming language, although there are many aspects of it that I still need to learn and work on.

➢ Have written a code which will check to see whether an image is square (i.e. k X k dimensions). If the image is not squared the code will add necessary number of black rows and columns to the image. This is because the Converging-Squares algorithm works only with squared images.

➢ Further I have developed another code which can change the respective intensity value of any pixel in the image to black (or zero). Note that the pixel number is obtained form the user and the only criteria is that it has to be within the dimensions of the image, the dimensions are given to the user.

Converging-Squares Algorithm

Converging-Squares Algorithm is an efficient means for determining the peak location in an image. For our purposes this method can be used to detect the centroid or regions of maximum intensity in each cell.

The algorithm also works in cases where there are multiple peak regions in the image (i.e. images which contain many many cells) and when the approximate size of the regions is known (currently, for this purpose, the diameter of the cells will be approximated). The image is examined in raster order (a scan pattern in which an area is scanned from side to side in lines from top to bottom) until a pixel I(x,y)T is found whose intensity exceeds a preselected threshold. The raster search is then suspended and a Converging-Squares sequence is initiated whose initial square is centered on (x,y)T, whose side length is set equal to the know maximum region diameter (for our purposes the side length of the square can be picked to be twice as long as the approximate diameter of the cell in order to ensure that it is completely enclosed by the square).

Converging-Squares Algorithm (continue)

After the peak within the region is found, pixels in the initial square area are set to zero to ensure that no further peak search will be undertaken for the same region. The raster search is then continued for further peaks. Thus, while scanning after reaching a pixel with an intensity value that exceeds the preselected threshold (for now it'll be selected based on trial and error) a square of size k X k would surround the cell, for the first iteration. This square is then divided into 4 overlapping sub-squares of length (k-1) X (k-1) each. The 4 sub-squares can be though of as pixels of an image of size 2 X 2, with each pixel representing an average over the actual pixels of each sub-square. So, by selecting the pixel with the highest intensity we are actually selecting the sub-square containing pixels of highest intensity relative to other ones. Then, in the second iteration that selected sub-square with the highest intensity itself will be again divided into 4 squares and the process discussed above gets repeated. As the square shrinks with each iteration the resolution increases. Eventually, the final comparison would be among 4 pixels of the original image which contains the peak. Then the raster search continues for detection further peak points in the image.

Current Goals

➢ Apply the Fortran code (designed and written by a previous NCHC employee) which is based on LSM method (described in the previous report) to the image of the cells and observe and analyse the results.

➢ Become more and more familiar with C++ language and work on

my programming skills .

➢ Work on both the algorithm and coding of Converging-Squares method.

Future Goals

➢ Each method and algorithm should be examined and applied to the data (images and movie clips, of biological cells and behaviour of neural networks in response to a stimulus, that are obtained from the UCSD lab).

➢ Result of each method should then be compared to the example of manual cell boundary detection (note that so far the best and most accurate tool for the purpose of cell boundary and location has been the human eye).

➢ Finish and implement on the Converging-Squares algorithm.➢ Obtain more knowledge on the molecular biology, neuropathology and

neurophysiology of the nervous system in order to be able to answer questions such as:

• What is the purpose of identifying cell boundaries, locations and ultimately neural networks?

• How will these finding help in the field of medicine, biology and specifically in clinical applications aimed at the regenerations and neuroprotection of the injured CNS (central nervous system)?

Pictures

The left picture is a group picture at top of an observation tower in YYL. The right is a picture taken from a water fall in Wuali (a city in Taipei county).

Acknowledgement

Many Thanks to:

Stephen ChenDr. Fang-Pang Lin