At first glance, the white plastic box with a bright orange floor looks like something for storing children's toys. However, the box isn't used to store Lego bricks; it contains real mice – with the aim of minimising their suffering. "This box allows laboratory animals to be observed in a humane and standardised way, whether by us here in Zurich or by researchers on the other side of the world," says Oliver Sturman, Head of the 3R Hub. The Hub is the point of contact at ETH Zurich for questions relating to the 3Rs – Replace, Reduce, Refine (see box).
For demonstration purposes, Sturman places a black plastic mouse in the box. Inside the box, whose front wall and lid are made of black acrylic sheets, it is pitch dark. "This is important, so that the animals feel comfortable and unobserved," says the neuroscientist. "When they are first placed in the box, they sniff about and explore the surroundings – which is natural behaviour. After a while, they get used to it and sometimes even fall asleep."
Two cameras – one from above and one from the front – film what is happening inside through the sheet. An infrared lamp allows the cameras to see in the dark.
Pain can be perceived in the face
The two cameras automatically record the mouse's body and face, providing indications on how the animal is feeling. This allows for the detection of subtle signs of pain and discomfort that are often reflected in the facial expressions of rodents – a narrowing of the eyes, a bulging of the nose and cheeks, or a change in ear position or whisker direction.
An algorithm then assesses the mouse's facial expression in real time. The new system, which the researchers have called the GrimACE, allows a rapid and precise assessment of whether animals are suffering and may need additional pain relief.
Current method time-consuming, subjective and imprecise
Facial expressions have long been used to detect and respond to potential pain and suffering in lab animals. The so-called Mouse Grimace Scale was developed for this purpose: each of the signs of pain and distress listed above is assessed on a scale from 0 (not present), 1 (moderately present) to 2 (obviously present).
To this end, scientists observe the animals from the cage side and compare their facial expression with detailed reference images on pictorial charts. This process is time-consuming and subjective.
Also, it is difficult for the human eye to gauge as the mouse's face may not be clearly visible. In addition, being observed by humans can cause additional distress in the animals.
Like a passport photo booth
The GrimACE system, on the other hand, allows an immediate, humane and objective assessment. As soon as the mouse is in the box, the video recordings begin. The system automatically selects the most significant frames and rates the features that could indicate an increased pain level.
Automated methods of facial recognition already existed, underscores Sturman. "What was missing was a complete, standardised, end-to-end system." The accuracy of algorithms diminishes if the surroundings are not identical or if the camera is sometimes placed nearer or further away.
We could liken the system to a photo booth for passport pictures, says Sturman. "As we all know, these machines are always built the same: a stool that is positioned a fixed distance from the camera, a white background and a dark curtain – all that ensures you get a successful photo, whoever and wherever the machine is used."
One kit for everything
The whole system including the software was developed by staff members from the 3R Hub – and is now being shared with the whole world as an open-source kit. "The idea is that as many users as possible can assemble and use it in a straightforward and standardised way – and that the data will then be comparable," stresses Sturman.
As with all computer vision and machine learning methods, the system continuously improves when it is trained on more image data. "The more people that use GrimACE, the less bias there will be."
Machine versus human
In a study, Sturman and other researchers from ETH Zurich tested the new system. They explored the question of whether the GrimACE can automatically and reliably detect pain in laboratory mice following brain surgery – and whether it provides comparable or even better results than trained human raters. They presented their findings in a paper that was recently published in the journal LabAnimal.
For the study, the researchers recorded images of the mice before and after brain surgery. After the surgery, the animals were given various painkillers in doses recommended by expert guidelines. The mice were operated for a different scientific goal, and the welfare assessment could run in parallel.
One expert viewed thousands of images of the mice before and after surgery with the naked eye as usual and assessed them manually. In parallel, the researchers also had the images assessed by the GrimACE. The result was that the automated assessments were a very close match with the expert's ratings.
Three people, three different ratings
The researchers also compared the ratings of three different people. Their assessments differed significantly.
This is not because the experts didn't do a good job, says Sturman, rather it is due to the subjective nature of rating. "We secretly gave all three raters the same images to assess, to check whether their own scores were consistent." And they were: individually, each person rated the images very consistently. One person gave both high and low scores. Another tended to give every image a lower score. And the third person gave all the images a higher score.
"This is where we see the strength of the computer as it delivers standardised results," says Sturman. Uniform assessment is important for animal welfare, emphasises the Head of the 3R Hub. This ensures an appropriate level of support for laboratory animals – in all laboratories. "If someone always assesses that an animal is not in pain, animals will suffer needlessly. And if someone always gives overly high scores, there is a risk that experiments are abandoned unnecessarily."
Besides facial features, the researchers also studied animal behaviour in their study on the suitability of the GrimACE. For this, a high-resolution camera from above recorded various points on the mouse's body. Features such as varying distances between individual points, changes in angle between two points, and acceleration of points provided indications of the mice's state. In such data, machine learning algorithms look for subtle differences that are barely visible to humans.
Worldwide interest
As soon as it was launched, the GrimACE met with widespread interest, says Sturman. "We've already received a number of email enquiries, for example from the US and UK."
To ensure that as many researchers as possible at ETH Zurich have access to the automated system, the 3R Hub recently installed a GrimACE System in the ETH Phenomics Center (EPIC).
Staff at the 3R Hub are already planning to further develop the GrimACE technology. It is not yet clear whether they will then patent the system and market it as a spin-off. "We're currently sharing our knowledge and technology in collaborations and are focusing on mutual data exchange to improve the system," says Sturman. "Our primary concern is to improve animal welfare."
3R principles
The 3R principles describe an ethical approach to animal experimentation. They stand for Replacement, Reduction and Refinement. The 3Rs aim to minimise the use of animals in scientific experiments, optimise animal welfare and promote alternative methods. ETH Zurich implements the 3R principles in animal experimentation, conducts its own research into the topic and set up the ETH 3R Hub in 2024 to consolidate 3R research efforts and to advise and support researchers.
References
Sturman, O., Schmutz, M., Lorimer, T. et al. GrimACE: automated, multimodal cage-side assessment of pain and well-being in mice. Lab Animal (2026). DOI: 10.1038/s41684-026-01695-9
