Contents:
Current Research: Characterising Behavioural Phenotypes Using Automated
Image Analysis.
This work is funded by a joint BBSRC/EPSRC grant under the Bioinformatics
Initiative.
An adult male Sprague-Dawley rat in the elevated zero maze.
Background
Animal models are routinely used in studies of psychiatric and neurological
disorders, requiring systematic study of behaviour. There are various
simple automated systems available to record motion/movement patterns (e.g.
light
beam interruptions or vibration sensing devices), but the most versatile
approach is the analysis of video recordings of animals in a standardised
environment (water-maze, plus-maze,
zero-maze
etc.). Current automated systems for image analysis can be used to track
animals in video sequences, but they use only relatively simple techniques.
For instance, an obvious approach is to record the arena without the animal,
then calculate differences between this and subsequent images, identifying
large-enough regions of difference as being of interest. However, if you
look at the image above, you will notice that in addition to the animal
itself, there is also a reflection in the apparatus; both the animal and
it's reflection could be captured by the tracking system, leading to an
erroneous result for the position of the animal. Confusion is also possible
when there are other, necessary, varying elements in the scene, such as
run-number markers etc. Even without these problems, such systems only
provide position data; most current behavioural analyses
involve a skilled human experimenter observing and recording. This is an
extremely tedious task for the experimenter, with limitations on the number
of separate behaviours/events that can reliably be recorded. It is also
subjective and prone to operator bias. All these issues put considerable
constraints on the design of experimental scenarios.
Active Shape/Appearance Models
The great advantage of Active Shape/Appearance Model Techniques
is that the models contain the information as to the shape and appearance
(as the names suggest!) of the object in question, as well as information
on how these vary. So, a simple technique for detecting a white rat in
the above image might just look for all the large white blobs it could
find. It would find the rat, but it would also find the number markers,
and the highlights on the apparatus. Whereas with an Active Shape Model
(ASM), although it might try to fit to these things, it would enable us
to easily reject them, since the shape extracted from the image by the
model would lie well outside the normal variation seen in training data.
Overview
The Image Analysis task.
-
Initial Location of the animal: Relatively
simple, since for each candidate shape, the model provides us with a measure
of how rat-like it is, whether the background is as expected etc., which
allows elimination of other similar coloured objects and reflections.
-
Tracking: Once the position is established,
the animal can be reliably tracked between frames, since the change in
position is limited, and the change in shape will only be within certain
limits, which the model already knows about.
-
Extraction of Posture Information: In examples
such as the picture above, the rat occupies a relatively small portion
of the frame, and there is little information available other than the
shape of the outline. Extracting this data from the image is simplified
by the fact that the model does not try to describe arbitrary shapes, but
just the variations possible to actual rats. Hence, once the model has
fitted to the outline, its position, orientation, size and shape can be
described using very little data. However, experience has shown that this
is sufficient to distinguish postures.
-
Analysis of Behaviour: With
the data extracted from the images, we will then have a continuous track
for the animal in both position and shape space. As noted above, experience
has indicated that differentiable postures/behaviours (grooming, walking,
rearing etc.) lie in well separated areas of the abstract shape space.
Hence, it will be possible to take the combined tracks in position and
shape space, and interpret them in terms of definite behaviours.
Aims
A major aim of this research is to construct a
robust system which is able to track and identify key behaviours in a manner
which reproduces the combined efforts of current automated tracking systems
and human-keyed data. This would obviously eliminate the tedium for the
experimenter concerned, and also possible operator bias. But it is hoped
that the detailed shape information would allow investigation of effects
not detectable by human observers. For instance, it might be the case that
not only the frequency of a particular behaviour changes, but also the
exact way it is executed. This could possibly lead to greater sensitivity
of behavioural measurements, and greater efficiency in terms of useful
data extracted per animal.
References :
J. K. Shepherd, S. S. Grewal, A. Fletcher, D. J. Bill and C. T. Dourish
Behavioural and pharmacological characterization of the elevated
'zero_maze' as an animal model of anxiety
Psychopharmacology (1994) 116:56-64.
A paper which describes the zero-maze and the postures/behaviours
measured to detect anxiogenic/anxiolytic drug action.
These papers are only intended as a sample to illustrate the use
of various experimental scenarios and various methods of behavioural recording/analysis.
M. Ramanthan, A. K. Jaiswal and S. K. Bhattacharya
Differential effects of diazepam on anxiety in streptozotocin
induced diabetic and non-diabetic rats
Psychopharmacology (1997) 135:361-367.
Includes use of the elevated zero maze, also plus maze, open-field
exploratory behaviour and social interaction tests.
C. K. Kellogg and A. Lundin
Brain androgen-inducible aromatose is critical for adolescent
organization of environment-specific social interaction in male rats.
Hormones and Behavior (1999) 35:155-162.
Includes use of light-beam monitoring of locomotor activity, and
human scoring of social interaction.
L. M. Schrott and L. S. Crnic
Anxiety behavior, exploratory behavior, and activity in NZB x
NZW F1 hybrid mice: Role of genotype and autoimmune disease progression.
Brain, Behavior and Immunity (1996) 10: 260-274.
Includes use of the elevated plus-maze and scoring of line-crossing
and rearing in an open arena.
Links:
Measuring Behavior 2000
3rd International Conference on methods and techniques in behavioral
research, 15-18th August 2000, Nijmegen, The Netherlands
A recent report on open-area avoidance in anxious transgenic mice.
http://news.bbc.co.uk/hi/english/sci/tech/newsid_428000/428219.stm
'Stressed-out mice gain less weight and prefer safe
places'
A news release from the University of Michigan, which includes photographs
of mice head-dipping in the elevated plus maze.
http://www.umich.edu/~newsinfo/Releases/1999/Sep99/r092799e.html
The full text of the article.
Karolyi et al.
Altered anxiety and weight gain in corticotropin-releasing hormone-binding
protein-deficient mice.
PNAS
(1999) Vol 96, 20: 11595-11600.
'By genetically engineering a smarter than average
mouse, scientists have assembled some of the
central molecular components of learning and memory'
A Scientific American article on building a brainier
mouse, including movie of testing in the Morris water maze.
http://www.sciam.com/2000/0400issue/0400tsien.html
Publications/Conferences:
'Characterising Behavioural Phenotypes using Automated
Image Analysis',
C. Twining, C. Taylor, P. Courtney and C. Dourish
Talk given at Measuring Behavior
2000, Nijmegen.
'Robust Tracking and Posture Description for Laboratory
Rodents using Active Shape Models'
C. J. Twining, C.
J. Taylor and P. Courtney
To appear in 'Behavior
Research Methods, Instruments & Computers' , Measuring Behavior
Special Issue, August 2001.
'Kernel Principal Component Analysis and the construction
of non-linear Active Shape Models'
C. J. Twining and C. J. Taylor
Talk to be given at BMVC
2001, September 10th-13th, University of Manchester.
'An Information Theoretic Approach to Statistical Shape
Modelling'
Rhodri H. Davies, Tim F. Cootes, Carole J. Twining and Chris J.
Taylor
Talk to be given at BMVC
2001, September 10th-13th, University of Manchester.