Murphy Lab - Classification of Fluorescence Images
Michael V. Boland, former graduate student in
Mia K. Markey, former undergraduate in
The goal of this work is to develop methods that allow the numerical
description and subsequent classification of the patterns
characteristic of subcellular structures in fluorescence microscope
images of eukaryotic cells. Such data are generated on a regular
basis by labeling one or more cellular molecules with fluorescent dyes
(most often by using antibodies against specific proteins.) As
currently practiced, investigators identify patterns based on
experience or via comparison with patterns of known proteins. The
question we address is whether these patterns can be described in a
way that is amenable to further processing by a computer, thereby
enabling automation of their analysis.
There are three steps involved in the automated classification of fluorescence microscope images:
As such, we first generated a set of images that could be used with
the subsequently designed classification methods. This set consisted
of more than 300 images generated using five different fluorescent
markers (four antibodies to cellular proteins and a DNA stain). These
images are available for download.
- Image acquisition
- Feature extraction
We then proceeded to generate
numeric features to describe these images. Because of the nature of
fluorescence microscope images, we pursued features that were
invariant to translations and rotations of the fluorescence patterns
within the images. We also made every effort to choose features
without considering the five classes of patterns we had generated.
That is, we desired those features that were generic enough to be
useful with additional patterns we would generate in the future rather
than those tailored to our existing data. The first two sets of
features generated, then, were based on Zernike moments and Haralick's
To classify these features, we implemented both
a classification tree and a backpropagation neural network.
The combination of Haralick's texture features and the neural
network classifier allowed us to achieve 88% recognition on previously
unseen single images from the five classes described above.
Given the significant heterogeneity in the images in each of the
five classes we generated, we consider our classification results to
be quite good. Furthermore, we believe this result can be improved by
looking at many (i.e. 10-20) images of cells subject to identical
preparation and then generating a single classification for the entire
set (see References for details.)
Boland, M.V., Markey, M.K., Murphy, R.F. 1998. Automated Recognition
of Patterns Characteristic of Subcellular Structures in Fluorescence
Microscopy Images. Cytometry 33:366-375.