Processes

From TLMtracker

Jump to: navigation, search

In this section the predefined segmentation processes are described.

Contents

General

There are a number of more or less complex predefined processes for the segmentation of images, which come together with the program. Basically a process consists of a number of image processing pipelines. The image processing operations in these pipelines are executed sequentially on an input image resulting in a segmented image.

These predefined processes can be used as they are for the segmentation of images or they can be used as a starting point for the creation of new processes. Usually it will be necessary to modify them for an optimal segmentation of a new set of images. For all predefined processes it is assumed that in the brightfield images the cells are darker than the background, so the first operation is an inversion to make them brighter than the background. If this is already the case the processes can easily be modified by removing the first inversion operation.

User defined processes can also be saved and reused after loading. The following processes are preinstalled per default:

Threshold

This is a process using a simple manually selected global threshold value for the segmentation of images. It is separated in two process pipelines, but the second pipeline contains only the label operation. This can be useful if it is necessary to do manual corrections for certain images. In this case it is just necessary to select the paint-tool and add the manual corrections to the end of the first process pipeline, which is already selected by default.

Pipeline 1: The first step in pipeline 1 is a smoothing of the input brightfield image with a median filter with radius 1 to remove some of the noise in the image. After that the background is removed to get a more uniform appearance of multiple images. Following the threshold operation for the actual segmentation of the image is executed. Because sometimes there remain black dots inside the segmented cells, these are filled by a fill gaps operation. Finally objects smaller than 20 pixels are removed because they are probably artifacts.

Pipeline 2: The source image of pipeline 2 is the final image of pipeline 1, so that the image processing can continue at this state. The only operation executed here is the labeling of the previously achieved binary image, to get an indexed segmented image for the further determination of the cell properties.

RATS

This is a process using the RATS (Robust Automatic Threshold Selection) algorithm for the segmentation of images. It is separated in two process pipelines, but the second pipeline contains only the label operation. This can be useful if it is necessary to do manual corrections for certain images. In this case it is just necessary to select the paint-tool and add the manual corrections to the end of the first process pipeline, which is already selected by default.

Pipeline 1: The first step in pipeline 1 is a smoothing of the input brightfield image with a median filter with radius 1 to remove some of the noise in the image. After that the background is removed to get a more uniform appearance of multiple images. Following the RATS operation for the actual segmentation of the image is executed. Because sometimes there remain black dots inside the segmented cells, these are filled by a fill gaps operation. Sometimes the RATS operation produces artifacts, which can be removed by the following opening by reconstruction. Finally objects smaller than 20 pixels are removed because probably these are artifacts too.

Pipeline 2: The source image of pipeline 2 is the final image of pipeline 1, so that the image processing can continue at this state. The only operation executed here is the labeling of the previously achieved binary image, to get an indexed segmented image for the further determination of the cell properties.

Watershed

This is a process using markercontrolled watershed-transformation for the segmentation of images. It is separated in five process pipelines. The first four pipelines create the necessary markers for the cells and the background. In the fifth process pipeline the actual segmentation takes place.

Pipeline 1: In the first pipeline the markers for the cells are created. The first step is a smoothing of the input brightfield image with a median filter with radius 1 to remove some of the noise in the image. After that the background is removed to get a more uniform appearance of multiple images. Following the RATS operation for a segmentation of the image is executed. Because sometimes there remain black dots inside the segmented cells, these are filled by a fill gaps operation. Sometimes the RATS operation produces artifacts, which can be removed by the following opening by reconstruction. The erosion shrinks the cells to a smaller size to produce the actual marker. Finally objects smaller than 5 pixels are removed because probably these are artifacts.

Pipeline 2: In the second pipeline the background markers are created. The source image is the final image from pipeline 1, the cell markers. A distance transformation, a watershed transformation and a roots operation are executed to get a system of lines lying between the cell markers, consequently marking the background of the image.

Pipeline 3: The third and the fourth pipeline combine the cell and background markers. In the third pipeline the background markers from pipeline 2 are dilated which results in thicker lines.

Pipeline 4: In the fourth pipeline first, the cell markers and the dilated background markers are combined in such a way, that afterwards there is a narrow zone without cell markers next to the background markers. The reason for this is that when the cell marker image and the background marker image are overlapped in the last step of this pipeline, the cell markers must not touch the background markers. Otherwise these touching cells will be recognized as background too and not as individual cells.

Pipeline 5: In fifth and last pipeline a gradient operation is executed on the original brightfield image and the markers from pipeline 4 are combined with it via a impose minima operation. Finally a watershed transformation is executed to achieve the actual segmentation of the image.

Active contours

These are processes using different active-contours methods for the segmentation of images. They are separated in two process pipelines, but the second pipeline contains only the label operation. This can be useful if it is necessary to do manual corrections for certain images. In this case it is just necessary to select the paint-tool and add the manual corrections to the end of the first process pipeline, which is already selected by default.

Chan-Vese method

Pipeline 1: The first step in pipeline 1 is a smoothing of the input brightfield image with a median filter with radius 1 to remove some of the noise in the image. After that the Chan-Vese active-contours method is executed for the actual segmentation of the image. This algorithm yields only the contours of the cells, so these are filled by a fill gaps operation to get the cell areas. Finally objects smaller than 5 pixels are removed, because they are probably artifacts.

Pipeline 2: The source image of pipeline 2 is the final image of pipeline 1, so that the image processing can continue at this state. The only operation executed here is the labeling of the previously achieved binary image, so get an indexed segmented image for the further determination of the cell properties.

GAC method

Pipeline 1: The first step in pipeline 1 is a smoothing of the input brightfield image with a median filter with radius 1 to remove some of the noise in the image. After that the GAC active-contours method is executed for the actual segmentation of the image. This algorithm yields only the contours of the cells, so these are filled by a fill gaps operation to get the cell areas. Finally objects smaller than 5 pixels are removed, because they are probably artifacts.

Pipeline 2: The source image of pipeline 2 is the final image of pipeline 1, so that the image processing can continue at this state. The only operation executed here is the labeling of the previously achieved binary image, so get an indexed segmented image for the further determination of the cell properties.

Hybrid method

Pipeline 1: The first step in pipeline 1 is a smoothing of the input brightfield image with a median filter with radius 1 to remove some of the noise in the image. After that the Hybrid active-contours method is executed for the actual segmentation of the image. This algorithm yields only the contours of the cells, so these are filled by a fill gaps operation to get the cell areas. Finally objects smaller than 5 pixels are removed, because they are probably artifacts.

Pipeline 2: The source image of pipeline 2 is the final image of pipeline 1, so that the image processing can continue at this state. The only operation executed here is the labeling of the previously achieved binary image, so get an indexed segmented image for the further determination of the cell properties.

Personal tools