Automated identification and behaviour classification for modelling social dynamics in group-housed mice

Abstract

Mice are often used in biology as exploratory models of human conditions, due to their similar genetics and physiology. Unfortunately, research on behaviour has traditionally been limited to studying individuals in isolated environments and over short periods of time. This can miss critical time-effects, and, since mice are social creatures, bias results.

This work addresses this gap in research by developing tools to analyse the individual behaviour of group-housed mice in the home-cage over several days and with minimal disruption. Using data provided by the Mary Lyon Centre at MRC Harwell we designed an end-to-end system that (a) tracks and identifies mice in a cage, (b) infers their behaviour, and subsequently (c) models the group dynamics as functions of individual activities. In support of the above, we also curated and made available a large dataset of mouse localisation and behaviour classifications (IMADGE), as well as two smaller annotated datasets for training/evaluating the identification (TIDe) and behaviour inference (ABODe) systems. This research constitutes the first of its kind in terms of the scale and challenges addressed. The data source (side-view single-channel video with clutter and no identification markers for mice) presents challenging conditions for analysis, but has the potential to give richer information while using industry standard housing.

A Tracking and Identification module was developed to automatically detect, track and identify the (visually similar) mice in the cluttered home-cage using only single-channel IR video and coarse position from RFID readings. Existing detectors and trackers were combined with a novel Integer Linear Programming formulation to assign anonymous tracks to mouse identities. This utilised a probabilistic weight model of affinity between detections and RFID pickups.

The next task necessitated the implementation of the Activity Labelling module that classifies the behaviour of each mouse, handling occlusion to avoid giving unreliable classifications when the mice cannot be observed. Two key aspects of this were (a) careful feature-selection, and (b) judicious balancing of the errors of the system in line with the repercussions for our setup.

Given these sequences of individual behaviours, we analysed the interaction dynamics between mice in the same cage by collapsing the group behaviour into a sequence of interpretable latent regimes using both static and temporal (Markov) models. Using a permutation matrix, we were able to automatically assign mice to roles in the HMM, fit a global model to a group of cages and analyse abnormalities in data from a different demographic.