Figure 1.
Behaviour priorities employed to combine all pairwise classifiers hα⇒β in order to describe the behaviour of a mouse with a single class.
“1” is the highest priority. The blue mouse is the one actively performing the action. An arrow indicates movement, a circle the lack thereof.
Figure 2.
Algorithm diagram summarizing the behaviour classification phases.
The “Position Tracking” is composed of a pipeline of three modules. The blob detection module initialises the system, estimates the foreground shapes (i.e. it locates possible mice), and filters out unfeasible structures; the temporal watershed module identifies mouse positions and shapes, and their directionality; the mice matching module tracks the identities of each mouse. Then, a feature vector composed of 13 measurements describes relative position, movement and attitude of mice for all possible pairs. Finally, the continuous action description for the mice is generated thanks to our Temporal Random Forest approach, which evaluates ensembles of decision trees through time.
Figure 3.
Examples of challenging mice interactions.
The red, blue and yellow lines represent the contours detected by the tracking algorithm. (a) The whole mice arena; (b,d,f,h) details of the algorithm output; (c,e,g,i) unprocessed IR details.
Table 1.
Number of manual interventions every 30sec used to correct the identities exchange od the tracking algorithm on dataset A.
Table 2.
Behaviour agreement among the two graders (top section) and quality of the system compared to the two graders (middle and bottom sections) on Dataset A (higher values are better ).
Table 3.
Average time difference (in seconds) between the overall duration of each behaviour as classified by the graders or by the system (less is better).
Figure 4.
Overall mouse behaviour for one mouse in Dataset A-1 expressed in total time, as specified by the two human readers and automatically generated by our system.
The results obtained by the classifier are comparable to those of the human scorers.
Table 4.
Comparison of classification performance between the method presented in this paper and the classification approach of Burgos-Artizzu et al. [23].
Figure 5.
Fully automated analysis comparing the interactions of C57BL/6J (N = 10) and BTBR (N = 6) mice of all experiments with two animals per cage (Dataset B).
The top graphs show the overall occurrence of each social and non-social interaction. We generated a different graph for “StandTogether” behaviour since its classification as either social or non-social is arguable. The bottom graphs show the aggregate comparison of social and non-social interaction. The null hypothesis that the C57BL/6J mouse and BTBR mice show a similar social/non-social behaviour is rejected in both cases by a two-class, two tail t-test that assumes equal variance. This shows impaired social activity in the BTBR case. The significance values of the t-test are (*) p<0.05, (**) p<0.005, (***) p<0.0005.
Figure 6.
Fully automated analysis comparing the interactions of C57BL/6J mice in cages of two (N = 6) or four (N = 8) cage mates (Dataset B).
The top graphs show the occurrence of each social and non-social interaction. For the same reasons exposed in Fig. 5 the “StandTogether” behaviour is shown in a separate graph. The bottom graphs show the aggregate comparison of social and non-social interaction. The null hypothesis that mice show a similar social/non-social behaviour, regardless of the number of mice interacting, is rejected in both cases by a two-class two tail t-test that assumes equal variance. This shows an increase of social activity when C57BL/6J mice can interact with more littermates. The significance values of the t-test are (*) p<0.05, (**) p<0.005, (***) p<0.0005.