Figure 1. An example fingerprint (6). The dark lines are the ridges on the inside of the finger and the white lines are the valleys between the ridges.
Figure 2. Details showing two minutiae points in the fingerprint in Figure 1 (marked).
Main links:
Home | About us | News and publications | Centre-led programmes | Funding schemes | Information portal
Site map | Help | Contact us | Media Centre | Accessibility | Disclaimer | Acknowledgements | Glossary
Mr Jens Cameron, DiLab, Sweden; Dr Christina Jacobson, AstraZeneca R&D, Sweden; Dr Kenneth Nilsson, Halmstad University, Sweden; Professor Thorsteinn Rögnvaldsson, Halmstad University, Sweden
Laboratory rodents are today identified using methods such as tattoos, ear clips or implanted transponders. These methods are invasive and can cause pain and/or distress for the animal, with associated ethical issues. A new method is presented here which relies on each animal's innate physical characteristics for identification and is inspired by fingerprint identification of human individuals. The new method, which uses blood vessel patterns in the ear as unique identification traits, is non-invasive and painless and seems to be as accurate as fingerprint identification for humans.
It is sometimes necessary to identify individual animals used in pharmaceutical and other studies, for regulatory and breeding reasons, or due to experimental design (1,2). The increasing use of genetically modified animals increases the need for identification, in order to separate different genotypes within a breeding colony or study. If identification of laboratory animals was easy (i.e. required no special effort from laboratory personnel or no special equipment) and caused no adverse effects for the animals involved, then it would be even more widespread. Adopting easy to use, humane methods of identification would achieve refinement (e.g. identifying individual animals makes group housing possible in more cases) and perhaps also reduction (e.g. the ability to follow a single animal through an entire experiment increases the accuracy in comparisons and reduces the number of animals needed in an experiment).
Current practices for identifying laboratory rodents can be non-invasive or invasive. The main drawback with non-invasive methods is that they last for a short time only. For example, patterns can be shaved in fur but the fur can grow back within days and the procedure must then be repeated. Similarly, fur can be marked with dye but the marks can fade rapidly as the animals groom. Coat colours/patterns have been used to identify animals in a group but work only when clearly disparate; most of the time rodent experiments will involve animals of the same strain or genotype which are similar in appearance.
Invasive methods, which include ear punching/notching, ear tags, tattoos, subcutaneous electronic transponders (microchipping), tail clipping or even toe clipping (1,2), are more permanent but are not ideal from an animal welfare perspective and hence not used routinely. The ideal method would provide permanent identification with no adverse effects for the animals.
There are also practical drawbacks with current invasive methods: ear punches can become difficult to read after a few weeks; ear tags can become infected or fall off; tattoos can be unsuitable for pigmented animals or wear off; and inserted electronic transponders can be hazardous, even for large animals (1,3).
Several non-invasive methods exist for identifying humans based on physical traits (biometrics), such as fingerprints, face geometry, iris pattern and retina blood vessel pattern. Given this, it is surprising that hitherto methods of reliably identifying laboratory rodents using a biometric signature have not been developed. This is perhaps because of their small size, since biometric identification has been tested on larger animals; it is much more straightforward to transfer methods that have been developed for humans to animals of about the same size as humans. For instance, iris scan identification has been tested on horses (4), and retinal scan identification has been tested on cattle and sheep (5).
However, the application of biometrics for identification of horses/cattle/sheep is different from the intended laboratory application in one important aspect. In the former case one wants to know whether a particular animal is indeed the animal it 'claims' to be, which is called verification. The measured pattern is compared against only one reference pattern (the claimed identity). This means that one can afford to use quite complicated and slow methods for the comparison, since the comparison is done only once or on very few occasions. In the laboratory application, on the other hand, one wants to know who the measured individual is in a potentially large database of individuals. This is called identification and means that the measured pattern is compared against many patterns in the database. It is highly impractical to use slow and complicated algorithms in this case, since it is questionable if laboratory personnel would have the time to wait several seconds for each identification. To become a routine procedure, the identification probably needs to be as fast and uncomplicated as the bar code reader used in supermarkets.
In this article, we show that laboratory rodents can be identified by their ears using a method inspired by human identification using fingerprints, with techniques that allow matching of a query image against many reference images in a database. The intention is for this method to be developed so that it becomes as quick and uncomplicated as the supermarket bar code reader.
A fingerprint is a copy, made with, for example, ink or a solid state sensor, of the pattern of ridges and valleys on the inside of the finger (see Figure 1). Human fingerprints are unique for every individual and do not change during a person's lifetime unless they are physically treated or damaged. The ridges and valleys of the finger form discontinuous lines that bifurcate or end, which creates patterns called minutiae points (see Figure 2) whose positions and orientations can be compared between prints. If two fingerprints have many minutiae points in common (about 12) then they are assumed to come from the same individual.
|
|
|
|
Figure 1. An example fingerprint (6). The dark lines are the ridges on the inside of the finger and the white lines are the valleys between the ridges. |
Figure 2. Details showing two minutiae points in the fingerprint in Figure 1 (marked). |
There are many commercial fingerprint identification (verification) systems available today, where the fingerprint is detected using a small touchpad and minutiae points are automatically extracted using image processing filters that look at, for example, the symmetries of the local line pattern (7). Such commercial systems are highly accurate and are used by the police and immigration services.
We have developed a method whereby an image of a rodent's ear can be used for identification (8,9). Rodents' ears are thin and fairly free of fur; hence it is possible to get a detailed, high-contrast picture of the blood vessels in the ear by very briefly applying light to the back of the ear and photographing it from the front. The blood vessels form a tree structure (Figure 3) with branching points, which can be thought of as minutiae points similar to the minutiae in fingerprints (Figure 4).
|
|
|
|
Figure 3. An example blood vessel tree (from a C57BL/6J type mouse). |
Figure 4: The blood vessel tree in Figure 2, with minutiae points marked (minutiae points were automatically detected). |
These minutiae points can be extracted automatically by looking at the local orientation in the image; locations where the local orientation is very divergent (i.e. where there are lines with different directions) are labelled as minutiae points. The positions of these minutiae points can then be compared between two images and it is then possible to estimate the probability that two blood vessel trees are different (e.g. by testing the hypothesis that two blood vessel trees match by chance).
The matching of two sets of minutiae points is done by first
aligning the two sets, with respect to rotation and
translation, and then computing a score. The score is based on
the sum of the probabilities of observing each minutiae point
in the query given the reference (database) pattern. In this
way a score is computed for the match between the query animal
earprint and each animal earprint in the database. The
probability of the highest score is then computed, based on a
random match hypothesis, and if this probability is
sufficiently low (e.g. 5%) then that match is flagged as a
positive identification. In questionable cases, when the
highest score is close to the decision threshold, a
verification step is done in which the full query image is
matched against the best scoring database images. The risk of
erroneous identification is computed from the highest
verification score and then the same strategy as described
above is used to flag for a positive identification.
The method has been tested on a collection of mice (C57BL/6J)
at the AstraZeneca research and development (R&D)
facility in Lund, Sweden. Twenty mice were photographed in two
recording sessions. The mice were 7-8 weeks old at the first
session and 21-22 weeks old at the second session. In the first
session, five images were captured per mouse and ear; in the
second session, only one image was taken per ear. (For both
recording sessions another 30 mice of the same age were
photographed as a control group). One reference image for each
of the 20 mice was used as reference in the database; the
remaining images were used as queries.
The test demonstrated that each of the 20 animals in the database could be positively identified but that it was sometimes necessary to request a second image if the first image was of low quality. The test results are summarized in Table 1, where "P(missed identification)" means the risk of "missed identification" (i.e. that the system is unable to find a sufficiently certain identification for the query, which then requires a second image), and "false identification" means that the animal is erroneously identified (this only happened for animals that were not stored in the database). It is possible to vary the "false identification" and "missed identification" rates by tuning the decision threshold (certainty value column in Table 1). The entire matching process, excluding the time for the image capturing, takes about 1 second when running the algorithm on non-optimized software and hardware.
Table 1. Probabilities for false identification and missed identification when the decision threshold in the verification step is varied (certainty value = "high" means that the decision threshold is restrictive).
|
Certainty value |
P(false identification) |
P(missed identification) |
|
Low |
6.0% |
10% |
|
Medium |
1.5% |
13% |
|
High |
0.8% |
18% |
A handheld image capturing device is currently being developed by DiLab in Lund. The device allows simple capturing of the ear image while holding the awake animal with one hand. The scoring method is also being fine-tuned, so that the verification step is needed less frequently. The prototype system will be evaluated at AstraZeneca R&D in Lund. The goal is to capture an image of the awake animal's ear within 1 second.
A new biometric method for identifying laboratory rodents
from the blood vessel pattern in their ears has been
introduced. The method is not harmful to the animals and it
seems possible to design the equipment so that animals can be
identified while being handled for other purposes (i.e.
"on the fly"). This would mean that identifying
laboratory rodents can become a routine task, supporting
refinement in laboratory animal studies.
Work is underway on reducing the risk for "false
identification" and "missed identification" to
zero. Probably the best approach, to keep the method fast but
still increase precision, is to double-check uncertain
identifications with a more precise verification method that
takes a slightly longer time but does not require a second
image. One approach is to extract a graphical representation
for the blood vessel tree and match this graph to the database.
If this is done only for uncertain identifications, which arise
rarely, then the identification method will still be rapid.
All views and opinions expressed in this article are those of the author and do not necessarily reflect the views and opinions of the NC3Rs.
In this section:
Banners:
NC3Rs | 20 Park Crescent, London W1B 1AL | Tel: 020 7670 5331 | Fax: 020 7670 5178 | Email: enquiries@nc3rs.org.uk