Automated measurement of stress scores in video recordings of laboratory mice
Veterinary Medical Science
Final Report Abstract
The goal of this project was to assess if current advances in the field of machine learning and computer vision can be applied to improve severity assessment in laboratory animal studies. To this end, we identified three areas which could benefit from automation and developed and validated algorithms in all three fields. In detail these three areas were: 1. Grimace scales, most importantly the mouse grimace scale (MGS). Grimace scales are a relatively novel method for visual assessment of severity, however their application is complex and requires extensive human work. To reduce the workload and minimize rater bias, we developed a set of methods that analyze video recordings, detect individual animals, extract images for scoring and ultimately perform the MGS scoring itself. The image processing workflow has been directly adapted from the human workflow, thereby allowing to directly map necessary human steps to individual algorithms. We were able to show that the method can perform a fully automated analysis of the MGS without the need for human interaction and that the results are clinically sound. Our results show that the neural networks do not globally score the entire image, but instead learn to associate individual scores with their corresponding facial parts, thereby fully applying the MGS procedure. Our method has been applied to novel, unseen data and it was able to differentiate between experimental and control animals. Additionally, we were able to show an increase in the MGS scores of animals after a carbon tetrachloride injection while the blindly assessed control group showed no such increase. Our methods have been also applied to other grimace scores where we were able to demonstrate that the developed algorithms can be adapted to other species as well. 2. Open field analysis, in which a single animal needs to be tracked in a closed environment. While open field analysis has already been addressed by commercially available and scientifically published algorithms, we extended current methods by adding a novel keypoint detection method and showing its robustness in an open field setting. We subsequently extended the method to allow the detection of multiple keypoints of an animal, thereby allowing additional tasks such as pose analysis. Additionally, we were able to perform open field tracking in a complex case in which established software failed to track the animal reliably. The results for both single and multi-keypoint detection have been published and are currently being applied by project partners in an animal laboratory to perform a large-scale analysis of pig movement in the open field. 3. Home cage monitoring with the goal of developing a system for the continuous camera-based monitoring of rodents in their home environments. Such a system would allow precise activity monitoring and improve comparability between different sites if multiple labs are equipped with identical devices. To assess the feasibility of this system, we developed a setup with multiple cameras that can be combined with a standard rodent home cage. Subsequently, we developed methods for the detection and identification of animals and also for the detection of multiple keypoints for each animal.
Publications
-
Semi-automated generation of pictures for the Mouse Grimace Scale: A multi-laboratory analysis (Part 2). Laboratory Animals, 54(1), 92-98.
Ernst, Lisa; Kopaczka, Marcin; Schulz, Mareike; Talbot, Steven R.; Struve, Birgitta; Häger, Christine; Bleich, André; Durst, Mattea; Jirkof, Paulin; Arras, Margarete; van Dijk, Roelof Maarten; Miljanovic, Nina; Potschka, Heidrun; Merhof, Dorit & Tolba, Rene H.
-
Robust Open Field Rodent Tracking Using a Fully Convolutional Network and a Softargmax Distance Loss. Informatik aktuell, 242-247. Springer Fachmedien Wiesbaden.
Kopaczka, Marcin; Jacob, Tobias; Ernst, Lisa; Schulz, Mareike; Tolba, René & Merhof, Dorit
-
Computation of Traveled Distance of Pigs in an Open Field with Fully Convolutional Neural Networks. Informatik aktuell, 229-234. Springer Fachmedien Wiesbaden.
Kopaczka, Marcin; Ernst, Lisa; Schulz, Mareike; Tolba, René & Merhof, Dorit
