Machine Learning to Identify Cells in Zebrafish-Skin Patterns
Harita Duggirala (Computer Science, 2024)
haritaduggirala2024@u.northwestern.edu
Broad Motivation
deccanherald.com
Zebrafish
deccanherald.com
Zebrafish
deccanherald.com
Zebrafish (Danio rerio) share a similar genetic
structure to humans. Their black pigment cells
(melanophores) are related to the originating cells
for melanoma in humans.The skin patterns of
zebrafish are comprised of horizontal stripes of
different colored pigment cells. Accurately
quantifying the cell size of various pigment cells
in relation to their location on the skin is a crucial
step towards better understanding cell behavior.
Motivation for My Project
Determine the size and distribution of black
pigment cells (melanophores) in the stripes.
Magnified image of zebrafish stripes
Approach: Machine learning to extract cells from
fish images in an automated process
Challenges
● Machine learning methods are often a black box
● Fish pictures are taken under various lighting
conditions & microscopic lenses (see below)
Objectives
● Automatically extract black cells from a large set
of fish images
● Quantitatively describe how different machine
learning methods influence resulting
measurements
Method
● Identify black cells manually in small set of
images using Ilastik
● Utilize batch processing to generate predictions
of black cells for larger set of images
● Quantitatively describe effects of using different
initial training sets on accuracy of cell prediction
Ilastik Object Classification Ilastik Object Prediction
Role of Training Set on Results
Our results show that using different images for
training during machine learning has:
● Minimal effect on our predictions of cell size
● Significant effect on the number of cells we extract
Future Research
● Compare accuracy of different image processing
software to determine optimal method for
extracting cells from fish images
● Isolate and determine the individual effects of
different lighting and microscopic-lens conditions
on the accuracy of our cell predictions
Acknowledgments
● Close-up images of cells courtesy of Hans Georg Frohnhöfer and
Christiane Nüsslein-Volhard (Max Planck Institute for
Developmental Biology)
● The study resulting in this presentation was assisted by a grant
administered by Northwestern University's Office of
Undergraduate Research. However, the conclusions, opinions,
and other statements in this presentation are the author's and
not necessarily those of the sponsoring institution.
● Thank you to Dr. Alexandria Volkening for your mentorship and
support for this project.