Maximum Likelihood Techniques for for Joint Segmentation-Classification of Multi-spectral Chromosome Images

Wade Schwartzkopf, Brian L. Evans, Alan C. Bovik
Embedded Signal Processing Laboratory
Laboratory for Image and Video Engineering
Department of Electrical and Computer Engineering
The University of Texas at Austin, Austin, TX 78712-1084



This work develops new methods for automatic chromosome identification by taking advantage of the multispectral information in M-FISH chromosome images and by jointly performing chromosome segmentation and classification. Chromosome imaging is a valuable tool for doctors and cytogenetic technicians. Extra chromosomes, missing chromosomes, broken chromosomes, and translocations (parts of chromosomes breaking off and attaching to other chromosomes) are indicators of radiation damage, cancer, and a wide variety of inherited diseases. There are currently over 325 clinical cytogenetics laboratories in the United States performing over 250,000 diagnostic studies each year involving chromosome analysis.

Traditional chromosome imaging has been limited to grayscale images, but recently a 5-fluorophore combinatorial labeling technique (M-FISH) was developed in which each class of chromosomes binds with a different combination of fluorophores. This results in a multi-spectral image, in which each class of chromosomes has distinct spectral components. Although M-FISH presents significantly more information than was available in traditional grayscale images, little research has been previously reported in the open literature.

The purpose of the research described in this work is to develop new methods for automatic chromosome identification. In particular, we (1) develop a maximum likelihood hypothesis test that uses this multi-spectral information, together with conventional criteria, to select the best segmentation possibility, (2) use this likelihood function to combine chromosome segmentation and classification into a robust chromosome identification system, and (3) show that the proposed likelihood function can also be used as a reliable indicator of errors in segmentation, errors in classification, and the chromosomes anomalies that can be diagnosed with M-FISH imaging.

We show that the proposed multi-spectral joint segmentation-classification method outperforms past grayscale segmentation methods in decomposing touching chromosomes, and we show that it outperforms past M-FISH classification techniques that do not use segmentation information. In addition, we show that the proposed likelihood function is a reliable indicator of abnormal chromosomes, as well as segmentation and classification errors.


This program accepts one MFISH image as input. It outputs an image of labelled connected components. Images can be in PGM, PNG, or raw formats. If input image is in PGM or PNG format, no height or width information is needed. The output image will be the same size and format as the input image. An MFISH directory structure is assumed such as used in the ADIR MFISH chromosome image database.

The syntax for calling this program:

mfSegment inputImage [MFISHdirectory outputImageFilename fileExt height width]

The default values for the optional arguments are:

This code requires the libpng library (which, in turn, requires the zlib library).

Last updated: August 26, 2002