Leukocyte Detection

leukocytedetection pic.jpg

Project Info

Title Advanced Biomedical Image Processing for Detecting Leukocytes In Vivo from Video Microscopy
Goal Design and build a fully automated leukocyte detection program.
Funding National Institutes of Health (NIH)


To detect leukocytes in digital video for the purposes of tracking.


Leukocytes (white blood cells) aid the body in fighting diseases and repairing tissue. In some people leukocytes accumulate in unwanted areas such as healthy tissue and joints which can result in heart disease and arthritis. Studying the behavior of leukocytes assists doctors in designing drugs to treat diseases and conditions such as Crohn's disease, stroke, multiple sclerosis, and atherosclerosis.

An algorithm that automatically detects the positions of leukocytes in a video could be used to initialize a tracking program or to count the number of cells that pass through a given line per unit time. Leukocyte velocity is a reliable predictor of inflammation intensity

Leukocytes appear both as bright and dark circles as shown in figure 1. The brightness or lack of brightness of a cell is due to the depth of the cell within a vessel. In a sequence, a bright cell may become dark.

leukocytedetection fig1.jpg
Fig 1. Sample Frame of Bright and Dark Leukocytes.

A level set method has been used to successfully detect the bright leukocytes. For each gray level, the algorithm looks for circular connected components (CC). If a CC falls within a certain size and the Area/Perimeter^2 is approximately 4*Pi, the CC is labeled as a potential cell. A mapping between gray levels determines whether a CC is a new cell, or a previously discovered one. To remove false positives, N consecutive frames are processed, and if the cell is not present in some percentage of N frames, then the potential cell is discarded.

Finding dark leukocytes has proven more difficult. Unlike bright cells, dark leukocytes don't not have a strong contrast with the background and tend to have broken edges. The following two algorithms have produced promising results:

Diffusion Hough: First, anisotropic diffusion is applied to the image. An edge detector (typically canny) is then applied. Finally, the hough transform designed to find circles is run. Points that correspond to large values in the hough space are considered to be center points of leukocytes.

Gradient Inverse Coefficient of Variation (GICOV): GICOV is a measure of the distribution of the edge strength along a contour. GICOV first collects the edge strength along a contour. The mean, m, and variance, v, is then computed. A high mean corresponds to a strong edge, and a low variance corresponds to a constant edge strength along the contour . GICOV computes the ratio m/v for different permutations of an ellipse centered at a provided point and returns the largest ratio. Leukocytes have larger ratios than non-leukocytes.

Using these two metrics, both leukocytes and non-leukocytes were ground truthed to determine the statistics under the two cases, leukocyte and non-leukocyte. A likelihood ratio was then constructed to classify potentital points.


Topic revision: r3 - 01 Jun 2014, AndreaVaccari
©2017 University of Virginia. Privacy Statement
Virginia Image and Video Analysis - School of Engineering and Applied Science - University of Virginia
P.O. Box 400743 - Charlottesville, VA - 22904 - E-Mail viva.uva@gmailREMOVEME.com