Image Analysis

The extraction of numerical data from images has a long history. The use of a calibrated scale to extract length data from objects seen in a microscope is an example of image analysis. Much of the foundation of image analysis was developed using manual analysis techniques on photographic prints. Many of the standard analysis methods or algorithms were developed for use with images taken by aircraft or satellites. A large portion of the modern foundation of image analysis was developed at institutions such as the Jet Propulsion Laboratory at Caltech, and other Federal laboratories. Image analysis techniques are designed to extract data from the image, and to do so, many times the image must be manipulated or altered in order for certain features of the image to be recognized. For the purposes of this discussion, we will restrict our discussion of image analysis to computer software analysis of images, and will not discuss specialized mechanical or purpose-built electronic systems for special image analysis tasks. A key requirement for image analysis is that the portion of the image or the element within the image that is to be analyzed can be clearly differentiated from the rest of the image. This is in some ways similar to staining a specimen.If we use a stain that colors everything the same color, it might be quite difficult to see the nuclei within the cells. The same is true for image analysis. If we want to analyze nuclei, then they should be clearly different in terms of color or density if a monochrome image from the cell cytoplasm. Image analysis has many tools to assist in defining one part of an image from another.

However, these tools are not perfect, and a good rule of thumb is that ninety percent of successful image analysis takes place prior to placing the specimen on the microscope. The reason for such a statement is to emphasize that with proper specimen preparation and staining, the task of image analysis, and therefore the reliability and precision of the analysis, is greatly enhanced. As a simple example, to analyze an H & E type stain with a monochrome camera where the object is to identify cell nuclei is sometimes difficult. The reason is that both the red eosin stain and the blue hematoxylin stain may have very similar grey levels when viewed with a monochrome camera this also applies to black and white film. By choosing a counter stain that has a very different grey level appearance, it is easy to separate the nuclei from the cytoplasm. For this example, a good strategy would be to substitute Orange G or Naphthol Yellow for the eosin.The difference in the monochrome images will be astounding, and all of the detail we see in our full color view through the microscope will appear in the image. Another example of this type of specimen preparation manipulation for imaging is again an example where we wish to count nuclei, but the nuclei are in a very basophilic tissue, such as the nuclei of pancreatic acinar glands. In this example, the basophilic cytoplasm may stain so densely that the nuclei cannot be seen. In this case, an azure eosin stain can be used, and when buffered appropriately, will be highly selective for nuclei, and eliminate most of the cytoplasmic basophilia.

A common method used to separate structures within images is called thresholding. To threshold an image, we examine the grey level values of the objects we are interested in separating. Remember that all of the pixel values in the image will fall between 0 and 255, with 255 being white or the clear background in bright field microscope images. In a specimen stained for nuclei with hematoxylin, we expect the nuclei to be the darkest objects in the image have the lowest numbers, since black is 0. For this example, that we look at the pixel values of a few nuclei, and we see that the lightest parts of the nuclei have values of around 100. If we set a threshold to a value of 120, then we should get all of the nuclei, since every pixel that makes up nuclei should have a value of 100 or lower. When we threshold, we actually convert the image to a binary image.A binary image is one where we have only two values, either black or white. In the case of this nuclei example, the nuclei would be black, and the remainder of the specimen would be white. In most image analysis software packages, the selection of the threshold value is made relatively easy by having an interactive control that works like a slider. You just move the slider back and forth until only the objects of interest are seen, and everything else disappears.