The Counting of Iron-Absorbed Small Intestinal CellsJoe HalfenECE 53312/19/05Introduction:The project is to research a reliable and invariant way to count the number of cellscontaining iron. The cells with iron are stained with a blue dye, and this dye can vary from slide to slide in its effectiveness. Currently, the number of cells is counted manually. This manual counting is subjective, very laborious and time consuming. The subjective nature to the counting comes from the varying shades of dye. The slides can also have thousands of cells which lead to the massive amount of time required to count manually. With the above troubles in mind an automated computer solution is sought. This process would fully automate and batch the image files and count the number of cells in each image. The initial pitch is for a UW-Madison project headed by Dr. Pamela King in the Department of Pediatrics. The project calls for a Plug-In to be built into their current software. This project will use Matlab to determine some of the steps to be used if such a plug-in is to be written.Approach:The approach to this problem is not unique. The image is taken and processed. The stages of the processing involve preprocessing to set up the image for a final threshold to convert to a binary image. This binary image can then be processed again. After all processing is done pixel relationships are used to count the number of regions in the image.This particular project’s primary focus is on the initial preprocessing and results. The reason for this is time constraints on the project and lack of resources. Many different preprocessing methods are available for this type of problem, and the most efficient method(s) is subject to many different constraints specific to the problem itself. The goal is to determine the most effective and efficient or, at the very least, one good method to preprocess the images.Work Performed:The slides are stained with a dye, so exploiting this fact was a primary focus of this project. The previous project used the RGB colors and thresholding on those colors to statistically determine the number of cells on an image. This project will focus on isolating the cells themselves and attempting to count them. RGB color will still be used to try to isolate the cells. First the exact characteristics of how the RGB colors contributeto the entire picture were explored. It was found that the cells that were stained appeared to be darker in the green values and lighter in the red values. The following are the four test cases used of one of the images submitted by the Department of Pediatrics for this project. Each image is show as a 256 x 256 cut of the original images with varying numbers of cells as well different separations. The reason for 256 x 256 images is to reduce processing time.Test Case 1Test Case 2Test Case 3Test Case 4The green part of the image has the cells that are to be counted which are darker than the rest of the image. The red part of the image’s cells is very light. This difference was used to try to eliminate the red component of the images making the cells to be counted very dark in a composite image of the red and green images. After a trial and error process, the following equation was used to eliminate the cells color and recombine this image. imagecombine dredgreen_2255Once the image was recombined a threshold was picked again through trial and error to help pick out a good representation of the target cells. The reason why this initial processcould not be done automatically is that the cells are unique to this problem, and a trial anderror method to see what worked had to be used.The threshold decided upon was approximately 115 intensity on a 0 to 255 scale. In the actual Matlab code an intensity of .46275 is shown, because the image is represented as a double precision image in Matlab which is on a scale from 0 to 1. This intensity value very accurately removed everything except the cells to be counted over the test cases.Results:The following are processed images for each test case and each whole image given by theDepartment of Pediatrics. Test Case 1 Test Case 2 Test Case 3Test Case 4 Image 1 Image2 Image3A built-in Matlab algorithm was used to do the counting of these binary images for initial results.Discussion:Preliminary results are very promising in this avenue of approach. The counts on some of the simpler images are very accurate and show that this is a legitimate way to try to solve the counting problem. There are however problems with the binary images that will need to be addressed in the next phase of preprocessing. Foremost is the reliance of a threshold level that was chosen to fix the pictures available. This trail and error processwill not be a sound tactic for arbitrarily scanned images with varying levels of dye and quality. A way to choose a “good” value will have to be determined if the value in the project does not work over most cases.Another problem is the combination of some cells into one region if they are next to each other. For the counts in the results a 4-nearest neighbor approach was used to try to stem the miscounts from more than one cell being combined in a region, but this approach is not reliable over all cases. To combat this problem a combination of dilation and watershed techniques along with some smart processing would be a good solution. The dilation could smooth over some of the holes in the larger regions of multiple cells. Next the processing could use the size in pixels of the average cell in the image. This size could be stated by the user or predetermined by the program itself. The watershed method could then be applied to separate the cells. Another possibly faster method wouldbe to just calculate the size of the region and divide it by the average size of the cells.Lastly, the inaccurate result in Test Case 3 is another possible issue that could appear in the program. There are no cells that are to be counted in this image, but the algorithm still states there is one cell. This is because the whole region of the image is counted as one cell. A size algorithm presented in the above paragraph to fix cells being combined into one large region could also be detrimental to this problem. If the whole image is first counted as one cell, then the algorithm could subsequently break up the region into the amount of cells it would
View Full Document
Unlocking...