V Congreso Latinoamericano de Ingeniería Biomédica (CLAIB2011)

Artificial neural networks and image features for automatic detection of behavioral events in laboratory animals
Carlos Fernando Crispim-Junior, José Marino-Neto

Última modificación: 2011-04-13 05:10


Animal behavior is a biological signal infrequently explored in signal processing research. Neuroscience studies assess changes in the expression of behavioral categories to examine the effects and mechanisms of action of drugs (e.g., on feeding or defensive behaviors), and to understand the underlying neurochemical and neurophysiological processes. Direct scoring of behavior is usually performed manually by a human observer, with the expected judgment errors (due to, e.g., varying experience, fatigue, and uncertainty about category definition). This paper examines the use of morphological and kinematic image features, processed by artificial neural networks (ANN), to automatically detect and score behavioral events on digital videos taken from rats placed into an open-field arena.  Video samples of rat (n=6, Wistar females, 250-300 g bw) were acquired using a camcorder, fixed perpendicularly to the arena center, and had their activities (locomotion, immobility, rearing and grooming) analyzed during 10 minutes after their placement on the experimental arena. From these samples, we extracted a set of image features (e.g., distance travelled, animal length, animal area) and their inter-frame changes, from the video frames included in each behavioral category, using locally developed software for animal tracking (EthoWatcher®). Statistical indices (mean, median, deviation measures) of each image feature were calculated for the samples of each behavior. These statistical indices were used as input parameters to train Multilayer Perceptron (MLP) ANN models. Training procedure used back-propagation algorithm and evaluation measures (10-fold stratified cross-validation, area under ROC curve – AUC). We correctly identify the 4 behavioral categories with AUCs (locomotion: 95.6%; immo-bility: 97.3%; rearing: 94.6%; and grooming 83.6%), suggesting that the combined use of morphological and kinematic features, and its variation representation over samples is able to successfully identify some of the behavioral  categories most used in neuroscience studies using open-field arenas.