|
nif:isString
|
-
The experiment was carried out at the BSL-3 facilities of the VISAVET Centre of the Veterinary Faculty of the Universidad Complutense de Madrid (UCM). The in vivo experimental protocol was approved by the Committee of Animal Experimentation of the UCM (Regulation 2010/63/EU and Spanish Royal Decree 53/2013). The protocol included a detailed description of efforts to prevent unnecessary suffering of the animals, including humane endpoints and the use of anesthetics and euthanasic following Galindo-Cardiel et al. clinical evaluation guidelines [10]. Specifically, humane endpoints were reached when at least one parameter was scored as severe. Animals’ health and welfare was checked twice daily (morning and evening) by a clinical veterinarian. Pigs were also monitored 24h/day by videocameras that were checked regularly during the day.
The experiment lasted from 11 April 2015 (day 1) to 3 May 2015 (day 23), during which eight Landrace Large White healthy pigs from an authorized breeding centre (four months old and weighing approximately 40 kg) arrived together one day before and were enclosed together in one indoor pen maintained at constant temperature and humidity, with ad libitum access to food and water. They were continuously monitored using one camera in a fixed position. At least twice per day, animals were evaluated by veterinarians and technicians who monitored animal internal temperature, recorded clinical signs [11], kept the pen clean and controlled feed supplies. Once per day, blood and oral samples were analysed for ASF virus using a qPCR test [12], which allows the confirmatory diagnosis of ASF virus within hours of a sample. All animals were accommodated in the pen and allowed to acclimate for one day before the experiment. On day 11, all animals were simultaneously inoculated with an attenuated strain of ASF virus from a tick captured in Kenya in 2005 (ASFV Ken05/Tk1). The first positive qPCR detection of the virus occurred on day 15 in six of eight animals. Subsequently, some animals showed mild clinical signs but first evident clinical signs associated with ASF started from day 18 in five of eight animals. After this day, all animals spent increasing amounts of time lying down until ultimately dying from the disease. The last animal died on day 23. Based on this time course of infection, the experimental period was divided into 4 phases for analysis (Fig 1): (1) pre-infection (days 1-11), when animals were free of infection; (2) infection (days 12-15), after animals were inoculated with ASF virus but with no ASF virus detected; (3) qPCR detection (days 16-18), during which ASF virus was qPCR-detected in at least one animal; and (4) clinical detection (days 19-23), when clinical signs of ASF were evident in at least one animal, until the end of the experiment.
Figure data removed from full text. Figure identifier and caption: 10.1371/journal.pone.0183793.g001 Timeline showing the four analytical phases of the experiment.The experiment began on day 1 (kick off), and animals were inoculated with ASF virus on day 11. On day 15, six of eight animals tested positive for the virus by quantitative PCR. On day 18, clinical signs of ASF became evident in five of eight animals. The experiment ended on day 23.
A total of 541 hours of video footage were taken using a single camera equipped with night-vision capability at a fixed position in an upper corner of the pen (Fig 2A). Video footage was collected continuously throughout the 23-day experiment. The room was artificially illuminated from 7 a.m. to 9 p.m., while the camera recorded under night conditions from 9 p.m. to 7 a.m. The ‘U’-shape and the limited size of the pen as well as the tall metal enclosure did not allow individual tracking of the animals. Since the fixed camera filmed the same background throughout, and light intensity remained fairly constant within the day and night periods, it was easy to distinguish between animals in motion and static background, despite the standard definition (SD) resolution of the video which might make subsequent motion detection less sensitive.
Figure data removed from full text. Figure identifier and caption: 10.1371/journal.pone.0183793.g002 A) Original video frame (all RGB channels). B) The same frame in the red channel. C) Region of interest in the red-channel frame used for motion analysis.
Video was recorded at 6 frames per second (704 x 576 pixels) in RGB24 format, providing 24 bits in red, green and blue channels (RGB) for broad range of colour. We used the red channel for all analyses since it provided the best contrast for animal recognition (see Fig 2B). In addition, we analysed only a region of interest in which all animals were always observable (see Fig 2C).
Animal movements were recorded and digitally processed through the Optical flow algorithm based on the Horn-Schunck methodology [13] implemented in Matlab. This algorithm estimates the speed and direction of moving objects between consecutive images based on the movement of brightness patterns [14]. This estimation process is described below. Supposing that we know the value of a function E(x, y, t) determining the brightness of any point (x, y) in the (fixed) 2-dimensional domain D recorded in the video, at any time t ∈ [0, Tmax] during recording and considering a point following a trajectory (x(t), y(t)) in D, if this point maintains its brightness along the entire time interval, the value of E(x(t), y(t), t) remains constant for any t ∈ [0, Tmax]. Then, assuming that functions E, x and y are smooth, ddtE(x(t),y(t),t)=0,∀t∈[0,Tmax],(1) or, after applying the chain rule for differentiation, ∂E∂x(x(t),y(t),t)dxdt(t)+∂E∂y(x(t),y(t),t)dydt(t)+∂E∂t(x(t),y(t),t)=0,∀t∈[0,Tmax]. (2) In this equation, (dxdt(t),dydt(t)) represents the velocity of the point that at time t is in (x(t), y(t)). Since we cover the entire domain D with all possible trajectories, we deduce from Eq 2 that ∂E∂x(x,y,t)u(x,y,t)+∂E∂y(x,y,t)v(x,y,t)+∂E∂t(x,y,t)=0,∀(x,y,t)∈D×[0,Tmax],(3) where (u(x, y, t), v(x, y, t)) is the velocity of the point located at (x, y) ∈ D at time t. This can be expressed more compactly as Exu+Eyv+Et=0,inD×[0,Tmax]. (4) To compute the velocity function (u, v), the Horn-Schunck method [13] minimises the error ζ(u, v) given by ζ(u,v)=∫D(Exu+Eyv+Et)2dxdy+η2∫D(ux2+uy2+vx2+vy2)dxdy,(5) where η is a weighting factor that scales global smoothness. Using the theory of calculus of variation, the velocity (u, v) minimising Eq 5 satisfies the Euler–Lagrange equations of ζ, given by Ex2u+ExEyv=η2Δu-ExEt,ExEyu+Ey2v=η2Δv-EyEt,(6) where Δu and Δv are the Laplacians of u and v, defined as Δu=∂2u∂x2+∂2u∂y2andΔv=∂2v∂x2+∂2v∂y2. (7) In reality, the entire video scene consists of Nmax∈N frames (or images) instead of continuous time t ∈ [0, Tmax], E(x, y, t) takes values in {0, 1, ⋯, 255} and D is divided into a matrix of 704 × 576 pixels. Therefore, we considered the discrete function Ei,j,k ∈ {0, 1, ⋯, 255} as the measured average brightness of the pixel at the intersection of the ith row and jth column in the kth frame. Hence, the minimum (ui,j,k, vi,j,k) of Eq 5 at an arbitrary pixel (i, j) in the discrete domain D of 704 × 576 pixels per frame k ∈ {0, 1, ⋯, Nmax can be estimated iteratively as follows [13]: ui,j,kn+1=u¯i,j,kn-ExExu¯i,j,kn+Eyv¯i,j,kn+Etη2+Ex2+Ey2,vi,j,kn+1=v¯i,j,kn-EyExu¯i,j,kn+Eyv¯i,j,kn+Etη2+Ex2+Ey2,(8) where ui,j,k0=vi,j,k0=0, n = {1, ⋯, N}, N = 25 is the maximum number of iterations, η = 10 considering a large relative motion between frames [15], and (u¯i,j,kn,v¯i,j,kn) is the neighbourhood average of (ui,j,kn,vi,j,kn) computed using the convolution kernel [1 2 1; 2 0 2;1 2 1]/4(1+2) [16]. We computed (Ex, Ey) using the convolution kernel [−1 −2 −1;0 0 0;1 2 1] and its transposed form for each pixel in the first image and Et between consecutive images using the [−1 1] kernel [15, 17]. Finally, we computed the global motion in frame k as mk=∑i=1704∑j=1576ui,j,k2+vi,j,k2. (9) Eq 9 allows us to obtain a unique motion value for each frame k and to reduce motion analysis to time series analysis. We analysed the global motion, mk, in order to detect a significant reduction in animal motion following ASF infection. Nevertheless, two main facts had to be considered before the analysis. First, the video resolution caused a perturbation in the values of mk. Indeed, even when no motion was recorded, we observed that the background slightly changed between consecutive frames causing a problematic baseline noise in the time series. Second, as mentioned previously, the human factor altered the real values of global motion when the workers were in the region of interest of the screen. Thus, some intervals of time scored excessively high values of mk only when the workers were in the pen. To smooth the perturbation in mk, we considered a simple moving average filter by replacing each data point in the time series with the average of the previous K∈N data points (i.e. the previous K frames). In this work, we considered K = 90; that is, 15 seconds approximately (see Fig 3). Henceforth, all motion works were carried out throughout the new time series m^k=1K∑i=0K-1mk-i,∀k∈{1,⋯,Nmax}. (10) Figure data removed from full text. Figure identifier and caption: 10.1371/journal.pone.0183793.g003 Global motion mk and the corresponding moving average m^k for a randomly selected 2h of video footage.The moving average was performed over a window of K = 90 frames. The interval of time series boxed in red on the left correspond to the video image on the right, in which the animals were in motion.
Activity budgets relate to the use of an animal’s time including moving, standing or lying, feeding, drinking, social and aggressive behaviours [21]. In this work, three types of motion were proposed to classify animal activity during the experiment: (1) baseline, when all animals were lying or sleeping, such that motion was minimal or absent; (2) animal motion, when animals were walking, feeding or playing; and (3) human-related motion, when veterinary practices were carried out. Animal movements sometimes increased in the presence of humans, and also movements of the humans themselves contributed to overall motion when they entered the region of interest being recorded. Therefore we quantified the relative amounts of all three types of motion each day and sought to determine changes in animal motion unrelated to human interaction. To classify motion as baseline, animal motion or human-related motion, we used short intervals of S = 15 seconds (corresponding to 6S = 90 frames) and we joined consecutive values of m^k into these intervals as follows Mr=[m^1+6S(r-1),m^6Sr],∀r∈{1,2,…}. (11) We described each interval Mr using five statistical parameters [18]: minimum (Mrmin), first quartile (Mrq1), median (Mrmed), third quartile (Mrq3) and maximum (Mrmax). Firstly, we analysed animal behaviour before experimental infection in order to quantify healthy levels of motion. To do this, we grouped the short series Mr corresponding to the pre-infection phase by using the k-means algorithm (kmeans command in Matlab performed with squared Euclidean distance) with the five statistical values (Mrmin, Mrq1, Mrmed, Mrq3, Mrmax), which assigned the input observations into k clusters through step-wise cluster centroid estimation based on optimisation of within- and between-cluster dispersion [19]. To set up the optimal number of clusters k used in the k-means algorithm, we used the gap criterion algorithm (evalclusters command in Matlab), which scored the so-called gap statistic associated to a proposed number of clusters based on the minimisation of within-cluster dispersion [20]. Consequently, the maximum value of the gap statistic is associated with the optimal number of clusters (assuming more than 3 and less than 50). These k clusters should represent a different type of motion recorded during 15-seconds intervals, when the animals were not infected. For instance, intervals in which the animals showed a similar activity within 15 seconds should be grouped in the same cluster (i.e. where there is no high variation between Mrmin, Mrq1, Mrmed, Mrq3 and Mrmax). But depending on the number of animals in motion and the type of activity carried out, the mean value of motion may differ between intervals (i.e. Mrmed) and, consequently, such intervals may be grouped in different clusters. On the contrary, intervals in which the animals showed significant changes in the activity such as when they were walking and the human interaction accelerated the overall motion during an instant, should be grouped in clusters with a high variation between Mrmin, Mrq1, Mrmed, Mrq3 and Mrmax. After k-means analysis, all k clusters (and the Mr records therein) were assigned to the three types of motion relevant for this work: baseline (group G1), animal motion (group G2) and (3) human-related motion (group G3). Secondly, we classified the remaining Mr records to groups G1, G2 and G3 corresponding to post-infection periods (Fig 1). To do this, we built a support vector machine (SVM) pairwise classifier in Matlab. To build an accurate SVM classifier, half of the Mr records in the pre-infection phase were randomly assigned to the so-called training dataset and the other half of records to the so-called testing dataset. To train the SVM model, we used the svmtrain command with the training dataset; then, we evaluated the accuracy of the model using the svmclassify command with the testing dataset. Once the accuracy of the model was verified, it was then used to classify the remaining Mr records of the experiment.
Changes in animal motion as a result of infection: We expected that the daily activity budget (i.e. time spent lying/sleeping and feeding/walking) would change after infection with ASF virus. To measure this reliably, we used the Wald-Wolfowitz runs test (runstest command in Matlab) at significance levels of 90%, 95% and 99%, corresponding to respective p-values of less than 0.1, 0.05 or 0.01, which evaluated whether consecutive daily motion values were randomly distributed around the average value of the considered period. First, runs tests were performed during the pre-infection phase (days 1-11) to verify random distribution of motion values around the average value. Then, runs tests were repeated by expanding the window by one additional day until day 18, when clinical signs were detected. This allowed us to determine whether and when consecutive motion values significantly diverged from the average value, indicating ASF infection.
|
![[This page as RDF]](http://data.gesis.org/somesci/static/rdf-icon.gif){kind=icon}