Comparison of surface thermal patterns of horses and donkeys in IRT images
Małgorzata Domino, Michał Romaszewski, Tomasz Jasiński, Małgorzata Maśko
CComparison of surface thermal patterns of horsesand donkeys in IRT images
Ma(cid:32)lgorzata Domino , Micha(cid:32)l Romaszewski , Tomasz Jasi´nski ,and Ma(cid:32)lgorzata Ma´sko Department of Large Animal Diseases and Clinic, Veterinary Research Centre andCenter for Biomedical Research, Institute of Veterinary Medicine, Warsaw Universityof Life Sciences (WULS–SGGW), 02-787 Warsaw, Poland;malgorzata [email protected] (M.D.); tomasz [email protected] (T.J.); Institute of Theoretical and Applied Informatics, Polish Academy of Sciences;44-100 Gliwice, Poland; [email protected] (M.R.); Department of Animal Breeding, Institute of Animal Science, Warsaw University ofLife Sciences (WULS–SGGW), 02-787 Warsaw, Poland;malgorzata [email protected] (M.M.) * Correspondence: malgorzata [email protected]; Tel.: +48-512-388-517
October, 2020
Abstract
Infrared thermography (IRT) is a valuable diagnostic tool in equineveterinary medicine however, little is known about its application in don-keys. The aim was to find patterns in thermal images of donkeys andhorses, and determine if these patterns share similarities. The study wascarried out on 18 donkeys and 16 horses. All equids underwent thermalimaging with an infrared camera and measuring the skin thickness andhair coat length. On the class maps of each thermal image, 15 regions ofinterest (ROIs) were annotated and then combined into 10 groups of ROIs(GORs). The existence of statistically significant differences between sur-face temperatures in GORs was tested both ‘globally’ for all animals of agiven species and ‘locally’ for each animal. Two special cases of animalsthat differ from the rest were also discussed. Our results indicated thatthe majority of thermal patterns are similar for both species however, av-erage surface temperatures in horses (22 . ± . ℃ )) are higher than indonkeys (18 . ± . ℃ ). It may be related to differences in the skin andhair coat. We concluded, the patterns of both species are associated withGORs, rather than an individual ROI, with higher uniformity of donkeyspatterns. a r X i v : . [ q - b i o . Q M ] O c t imple summary This study analyzes and compares the thermal patterns of horses and donkeysin infrared thermography (IRT) images. Thermal patterns were defined as sta-tistically significant differences between groups of ROIs corresponding to areaswith potential impact of large equine muscles. In our study, we used images ofhealthy and rested animals: sixteen horses and eighteen donkeys that formed adata set used in our experiments. We discussed our results, compared thermalpatterns between species and discussed special cases of animals identified as out-liers. Our results support the thesis about similarities in the thermal patternsof horses and donkeys.
Keywords
Infrared thermography; equids; thermal patterns; surface temperature; skinthickness; hair coat;
Infrared thermography (IRT) is a non-invasive imaging technique that allows todetect the radiant energy emitted by any object with a temperature above ab-solute zero. The radiated power detected by the thermal camera in the infraredspectrum is proportional to the fourth power of the object’s absolute temper-ature and is used to calculate the temperature of the target e.g. the surfaceof the animal’s body. Infrared radiation is often presented as a thermogramwhich is an image where the color gradient corresponds with the distribution ofsurface temperatures [27]. Furthermore, the relationship of temperature gradi-ents may create specific thermal patterns which may be used e.g. for assessingthe influence of load on saddle fit in horses [30] or the horses’ response to thetraining [15].IRT has been used as a diagnostic tool in equine veterinary medicine sincethe mid-1960s, particularly in the field of orthopedics, in the management oflameness [26, 8, 5, 6]. The surface temperature changes, reflecting heat emittedfrom overloaded or injured tissue, were considered a valuable indicator for iden-tifying areas of inflammation and blood flow alterations [13, 35]. This allowsto detect temperature changes before they can be detected by palpation [10, 2],and before the onset of other clinical signs of injury [2, 23]. Therefore, in recentstudies IRT was also applied to interpret changes in the surface temperatures ofthe thoracic region in the case of back pain diagnosis of equine athletes [9, 29],as well as results of the impact of load on saddle [30] or incorrect saddle fit [1].Moreover, the usefulness of equine IRT in the assessment of transient stressresponse during training [4, 20] and sport competitions [33, 3] has been demon-strated. Equine IRT seems to be highly related to thermoregulation and theincrease in blood flow due to exercise [20]. During physical exercise, metabolic2eat production increases as exercise intensity increases [11], and only a quar-ter of the energy used by a muscle is converted to mechanical energy. Theremaining three quarters are dissipated as heat [25]. Therefore, the radiant en-ergy emitted from the horse’s skin surface may be found as a product of basicmetabolic processes, exercise, and pathological conditions. However, it shouldbe kept in mind that the temperature measured from the body surface is relatednot only to the above internal conditions but also to the thermal properties ofthe skin and hair coat, and the thermal gradient between the skin surface andthe environment [32, 28].It is easy to see that IRT is widespread in the equestrian industry as avaluable tool to monitor the underlying circulation, tissue metabolism and localblood flow in response to different physiological, pathological, or environmentalconditions. However, little or no attention has been paid to application ofIRT to donkeys. The only work authors are aware of is the study of effectof season and age on daily rhythmicity of rectal temperature and body surfacetemperature during the cold-dry and hot-dry seasons in a tropical savannah [36].For the infrared measurement, the infrared thermometer and seven landmarksadapted from equine IRT were used. Although this study evaluated differencesin surface temperatures of donkeys of varying age groups under the changingenvironmental conditions, no studies to date have compared the thermal imagesof donkeys and horses obtained in the same circumstances. The scarcity of workson the imaging of donkeys motivates us to try to answer the question whetherthere are significant differences in thermal images of horses and donkeys. If theimages of these animals were similar, it could suggest that intensively researchedmethods for analyzing equine images are applicable to donkeys.We performed donkeys’ and horses’ imaging under the same environmentalconditions. Following previous equine researchers, we evaluated body surfacetemperatures in healthy animals. The normal thermal image has already beendescribed for e.g. the coronary band [21], distal forelimbs joints [14, 28], thethoracolumbar region [31], and the back and pelvic regions [17] in the horse. Itshowed a high degree of symmetry between the left and right sides of the body[27, 28] and reproducibility over hourly, daily, and weekly intervals up to 90%[31].The thermal images were manually segmented into fifteen regions of inter-est (ROIs) corresponding to underlying large muscles. Since the phenomenaobservable in thermal images often include more than one ROI, we combinedindividual ROIs into groups of ROIs (GORs) and examined the differences intheir mean temperatures. The differences, the occurrence of which has beenstatistically confirmed, constitute thermal patterns, which are the basis for thecomparison of both species and analysis of special cases (outliers). This com-parison is the main focus of our experiments. Our hypothesis is that thermalpatterns of horses and donkeys are similar.3
Materials and Methods
This section describes the methodology of acquiring, describing and visualis-ing our data, defining thermal patterns and assessing similarities between thepatterns observed in thermal images of horses and donkeys.
Eighteen donkeys (nine mares, seven geldings, and two stallions; mean age7 . ± .
04 years; mean height 119 . ± .
72 cm) and sixteen horses/poniescalled further horses (eight mares, six geldings, and two stallions; mean age7 . ± .
83 years, mean height 137 . ± .
33 cm) participated in the study. Thedonkeys and horses are privately owned and are housed in the same stable lo-cated in southern Poland in Luboch´ow. The owners of the animals consent tocarry out our research. The ethics approval was deemed unnecessary accordingto regulations of the II Local Ethical Committee on Animal Testing in Warsawand the National Ethical Committees on Animal Testing because all proceduresin the study were non-invasive and did not cause distress and pain equal to orgreater than a needlestick. The equids were fed three times a day with a doseof hay personalized to each animal to maintain an optimal, healthy conditionwithout obesity, and had daily access to a grassy paddock no shorter than 8h per day. Both during the study and the month preceding the study, equidswere not used in riding or harness. Before IRT imaging, physical examinationswere conducted to ensure that the equids were free from a preexisting inflam-matory condition. The imaging were carried out following the the internationalveterinary standards [18]. All donkeys and horses were clinically healthy, hadno apparent back or lameness problems, and demonstrated conformation typ-ical of their species and comparable growth. Two donkeys were excluded atthe stage of preliminary examination. The first of them due to plaque losses inthe hair coat caused by abrasions in transport in the week preceding the study,wheres the second due to the significantly longer hair length (7 . ± . . ± . To ensure the best possible conditions for comparison of collected thermal im-ages, the skin thickness, the hair coat length, and the constant thermal gradientbetween the skin surface and the environment were taken into account. Thestudy was performed in middle September, and all measurements were taken onthe same day under the same circumstances (ambient temperature 20.2 ℃ ; hu-midity 45%). A total of 68 images were taken in a closed space, protected fromwind and sun radiation, to minimize the influence of external environmentalconditions [22]. The imaged area was brushed and dirt and mud were removed4 a) SF-Skin of horse (b) SF-Skin of donkey Figure 1: Example of an ultrasonographic image taken over the 3rd lumbarvertebra: (a) the horse
H.1 ; (b) the donkey
D.3 . A subcutaneous fat plus skinthickness (SF-Skin) is highlighted.15 minute before imaging. The thermal images were acquired on the left andright sides at a 90 ° camera angle from a distance of approximately 2 m from thedonkey or the horse. During each imaging session two images of each individualwere taken. The images were positioned on the centre of the trunk. Images weretaken using an infrared radiation camera (FLIR Therma CAM E25, Brazil) withan emissivity (e) 0.99, by the same researcher (MM). The temperature rangewas standardized in the professional software (FLIR Tools Professional, Brazil)during the pre-processing of images at 10-30 ℃ level.Table 1: Measured features (mean ± SD) ofhorses (H.1-H.16) and donkeys (D.1-D.16):the length of hair coat and the thickness ofthe subcutaneous fat plus skin (SF-Skin).Animals Hair coat [cm] SF-Skin [mm]Donkeys 3 . ± . a . ± . c Horses 1 . ± . b . ± . d p-value < . < . Different superscript letters indicate significantdifferences between Horses and Donkeys for Haircoat (a, b) and SF-Skin (c, d) respectivelyaccording to the Mann-Whitney-Wilcoxon(MWW) test
After each IRT imaging, the ultrasonographic image was taken with an ul-trasound scanner (SonoScape S9, SonoScape, Shenzhen, China) using a linear5-12 MHz transducer (L752, onoScape, Shenzhen, China). Ultrasound scanswere performed with the transducer placed at the animal ´ s back, over the 3rdlumbar vertebra, perpendicular to the backbone; all the images were collectedon the left side of the animal [24]. The hair was trimmed at the measurementplace and ultrasound gel (Aquasonic 100, Parker, USA) was used as a coupling5 (a) Thermal map (b) ROIs (c) Extracted pixels Figure 2: Visualisation of a donkey
D.3 : (a) thermal data from the cameraas a thermal map; (b) annotated classes corresponding to selected ROIs (seeSec. 2.2.1); (c) extracted areas, used in our experiments.medium. The real time ultrasonographic examination was freezed, the imagewas saved, and the subcutaneous fat (SF) plus skin thickness (SF-Skin) mea-surement were obtained. An example of an ultrasonographic image is presentedin Fig.1. The hair coat samples were taken from the midneck about 5 cm belowthe base of the mane. The length of individual hairs were determined from arandom sample of five pulled strands, including the roots [16]. Average haircoat length and SF-Skin values are presented in Tab. 1.
Based on collected data, a data set was prepared that was later used in ourexperiments. The data set consists of images from a thermal camera and thecorresponding annotations in the form of class maps of main muscle areas. Everythermal image is a table of 320 x
240 pixels. The value in each pixel correspondsto the measured temperature value. A corresponding class map is a table, wherethe value in every pixel is the ROI number (or zero areas without annotations).An example class map is presented in Fig. 2. The class maps were producedby hand annotating the fifteen identified regions of interest (ROIs) in each im-age. The following ROIs corresponding to the underlying large muscles wereannotated:1. ROI 1 m. brachiocephalicus - a parallelogram-shaped area from the lat-eral surface of the Atlas, behind the angle of the mandible, to the regiosupraspinata of the scapula.2. ROI 2 mm. splenius capitis and cervicis - a triangle-shaped area from thelateral surface of the axis to the regio supraspinata of the scapula aboveROI 3.3. ROI 3 m. trapezius pars cervicalis - a triangle ranged from the middleof the neck to the regio cartilaginis of the scapula and along the regiosupraspinata of the scapula up to two-thirds of the length of the scapula.4. ROI 4 m. trapezius pars thoracica - a triangle ranged from the the regiocartilaginis of the scapula along the regio supraspinata of the scapula up6o one-thirds of the length of the scapula.5. ROI 5 m. latissimus dorsi - a triangle-shaped area from the regio in-fraspinata of the scapula, up to two-thirds of the length of the scapula,along the back to the tuber coxae.6. ROI 6 mm. glutei (superficialis and medius) - an irregular area in theregio tuberis coxae.7. ROI 7 m. biceps femoris - an oblong s-shaped area in the regio femoriscranially from the m. semitendinosus.8. ROI 8 m. semitendinosus - an oblong s-shaped area in the regio femoriscaudally from the m. biceps femoris.9. ROI 9 mm. in regio cruris - a rectangular-shaped area in the regio crurisbetween articulatio genus and articulatio tarsi.10. ROI 10 m. tensor fasciae latae - an irregular area between the in the regiotuberis coxae and the flank.11. ROI 11 m. obliquus externus abdominis - a trapezoid-shaped area fromthe lower two-thirds of the regio infraspinata of the scapula to the tubercoxae and the regio of processus xiphoideus sterni.12. ROI 12 m. pectoralis transversus - a triangle-shaped area behind the regioof olecranon to the regio of processus xiphoideus sterni.13. ROI 13 mm. in regio antebrachi - a rectangular-shaped area in the regioantebrachi between articulatio humeri and articulatio cubiti.14. ROI 14 m. pectoralis descendens - an irregular area in the projection ofregio infraspinata of the scapula.15. ROI 15 m. deltoideus - an irregular area in the projection of the regiosupraspinata of the scapula.
In order to facilitate replication of our results the data set and the experimentalsource code have been made available to the public under an open license. Our main goal was to find patterns in thermal images of both species anddetermine if these patterns share similarities. We define a thermal pattern as astatistically significant difference between two areas composed of ROI groups.Here we describe our methodology of finding and confirming the relevance ofthermal patterns. Data set location: Source code location: https://github.com/iitis/thermal_patterns.git Figure 3: Visualisation of a donkey
D.3 , divided into Groups of ROIs (GORs).
In this work our claims usually involve comparison of temperatures betweenareas (subsets of pixels) in a thermal image or images e.g. when we state ‘animalsA are warmer than animals B in area C’ this means that based on our sample, thesurface temperature of animals A is on average higher in this region. Therefore,when we claim that such statement is statistically significant, to confirm thiswe use the Mann–Whitney–Wilcoxon (MWW) test [7] (one-sided) which is anon-parametric statistical hypothesis test that allows to compare two relatedsequences of samples. We have chosen a non-parametric test due to the factthat temperature distributions in our ROIs are diverse and often non-Gaussian.The two sequences are obtained by randomly, uniformly sampling the comparedROI or groups of ROIs. The number of samples is the size of the smaller set– when comparing individual ROIs between animals, the difference in samplecount is usually less than 10%. Unless stated otherwise, we require that thep-value p < . The general idea is, that while we expect global differences between animalspecies e.g. on average, one species can have a higher surface temperature theother, we are interested in the existence of repeated dependencies between sur-face temperatures in different body areas of a given species. Such dependenciescan form a pattern that may be compared for both species. To identify thesepatterns, we use the following methodology:
Combining ROIs into groups of ROIs (GORs)
In the first step, basedon our observation that visible patterns in thermal images from our data setare often located in several ROIs, we have designated manually 10 groups ofROIs (GORs) for our analysis. The designed GORs, presented in Fig. 3 wereas follows:1. GOR 1
Neck , ROIs { , , } – represented an area with an impact ofmuscles located cranially from the cranial border of scapulae2. GOR 2 Frontquarter , ROIs { , , , , , } – represented an area with animpact of muscles located cranially from the spinous processes of scapulae8. GOR 3 Trunk , ROIs { , } – represented an area with an impact ofmuscles layed between the caudal border of scapulae and the vertical linedefined by tuber coxae, expecting area of m. pectoralis transversus
4. GOR 4
Hindquarter , ROIs { , , , , } – represented an area with animpact of the examined muscles of the pelvic limbs layed caudally fromthe vertical line defined by tuber coxae5. GOR 5 Rump , ROIs { , } – represented an area with an impact of twoROIs of the pelvic limbs layed the most caudally6. GOR 6 Dorsal aspect , ROIs { , , , } – collected the area with impact ofmuscles located above the horizontal line halfway up the trunk7. GOR 7 Ventral aspect , ROIs { , , , , } – collected the area withimpact of muscles located below the horizontal line halfway up the trunk8. GOR 8 Abdomen , ROI { } – represented an area with an impact ofmuscles layed between the caudal border of scapulae and the vertical linedefined by tuber coxae, expecting area of m. pectoralis transversus and m. latissimus dorsi
9. GOR 9
Groins (Girth and Flank), ROIs { , } – represented two areasmost covered by large muscles of thoracic and pelvic limbs with so girtharea and flank area10. GOR 10 Legs , ROIs { , } – represented two areas with an impact ofmuscles of the proximal parts of limbs, both thoracic and pelvic Comparing GORs temperatures
Our goal is to compare average temper-atures between designated groups of ROIs and test whether the difference isstatistically significant. We do it as follows:We have a set of animals of a given species A = { a , .., a } and a set ofgroups of ROIs defined in the previous paragraph G = { g , .., g } . A set T ag is a set of pixels (temperatures) of an animal a ∈ A from a group g ∈ G andthe mean value of pixels in a set is denoted by δ e.g. δ ( T ag ). For every pair ofgroups ( i, j ) ∈ G × G we compute a difference in average values of temperaturesin those groups for all animals i.e.∆ ( i,j ) = δ (cid:32) (cid:91) k ∈A T jk (cid:33) − δ (cid:32) (cid:91) k ∈A T jk (cid:33) . These differences are presented in our results as a matrix of differences M ∆ ∈ R |G|×|G| . Note that the matrix M ∆ is not symmetric as it shows temperaturedifferences and not their absolute values.Our next step is to apply the MWW test, described in Sec. 2.3.1 to verifythe statistical significance of the difference ∆ ( i,j ) for every pair of groups ( i, j ) ∈G × G . We do it in two ways: 9. Globally - for every pair ( i, j ) ∈ G × G we apply the MWW test to thewhole population i.e. we compare the union of sets (cid:83) k ∈A T ki with the unionof sets (cid:83) k ∈A T kj .2. Locally - for every pair ( i, j ) ∈ G × G we apply the MWW test separatelyfor every animal a ∈ A , by comparing the set T ai with the set T aj . Forour data set this results in 16 tests for every pair.In our results, outcomes of the MWW tests supplement the presented matrixof differences: results of the ‘global’ test are presented in the matrix itself (witha bold font) while results of ‘local’ tests are presented as a separate matrix M L ∈ R |G|×|G| + , where the value in every cell represents the number of animalsfor which the difference was significant. Thermal patterns
For a given animal species we treat a statistically signif-icant difference ∆ ( i,j ) between two groups of ROIs ( i, j ) ∈ G × G as a thermalpattern. A thermal pattern can thus be interpreted as a statement that basedon our data e.g.the Rump area is colder than the
Neck area. If the significanceis confirmed by the ‘global’ test but not for every animal in the ‘local’ test (i.e.the value in the matrix M L for this pair is not 16), it means that while thepattern emerges in a population, it is susceptible to individual differences ofanimals, and that there are animals that do not show it. If the pattern alsoappears individually in most (or all) of the animals tested, we consider it to bemore stable and reliable.Analysis of the structure of matrices M ∆ and M L for both species will bethe basis of our discussion about similarities in their IRT images. Here we describe visualisation techniques used to illustrate our observations.
In order to visualize the visible structures in IRT images from our dataset, thetemperature was presented in the form of a color map, modeled on the visiblepart of the electromagnetic spectrum, i.e. going from violet to red. To improvethe clarity of images, the zero values representing the areas outside the ROIsare shown in black. By manipulating the color map threshold values (assignedto its extreme colors), we visualise patterns common to all animals or highlightpatterns specific to a particular animal. An example of such a visualization ispresented in Fig. 11.Temperature distributions within a specific ROI are visualized using his-tograms where the y-axis is presented as a probability density, i.e. bin countsare divided by a total number of counts. Alternatively, we use boxplots wherethe box extends from the lower to the upper quantile values. The line in the10oxplot denotes the median, the whiskers denote the range of { q − . ∗ ( q − q ) , q − . ∗ ( q − q ) } where q q denote the first and the third quartiles andcircles denote outliers. An individual animal in our data set can be represented by a vector v i ∈ R d of d features corresponding e.g. to means or variances of temperatures in everyROI which leads to d ≥
15. Extraction and visualisation of data structures inhigh-dimensional space is often performed by using the Principal ComponentAnalysis [12] (PCA) and projecting data onto the first principal components.However, PCA uses the covariance matrix of data and since our data set containslimited number of examples, this makes computation of a reliable covariance ma-trix difficult. Therefore, to present our data we use the t-Distributed StochasticNeighbor Embedding (t-SNE) [34] algorithm that visualises data by giving eachexample a location in a two-dimensional map. An important feature of the t-SNE is that its output is non-deterministic which results from the fact that theoptimisation problem solved by the technique has a cost function that is notconvex. Since, in this work t-SNE is only used to highlight patterns emergingin data, we consider this acceptable and presented visualisations are selected asrepresentative examples after several executions of t-SNE.We will analyse the structure of t-SNE visualisation to investigate whetherROI features are characteristic for both species and allow to distinguish betweenthem. Features will be extracted with common statistics such as the mean,standard deviation, kurtosis and skewness. In addition, we will also check theimpact of mean normalisation by removing the global temperature of an animalfrom all pixels. This last experiment is aimed to test the distinctiveness ofdifferences between ROI temperatures of one animal.
This section describes our experiments and presents their results.
Unless stated otherwise, in all our thermal map visualisations, e.g. in the upperrow of the Fig. 11, presented color map temperature values t c were limitedeither to the common range of t c ∈ (cid:104) . , . (cid:105) ℃ , which are extreme values inannotated ROIs for all animals included in the study (not counting animals D.17 , D.18 , for reasons explained in Sec. 2.1).Alternatively, they were selected as extreme temperature values in ROIs fora given animal to highlight features of visible thermal patterns - these special Temperatures in ROIs – horses: t h ∈ (cid:104) . , . (cid:105) ℃ , E ( t h ) = 22 . ± . ℃ ,donkeys: t d ∈ (cid:104) . , . (cid:105) ℃ , E ( t h ) = 18 . ± . ℃ T e m p e r a t u r e (a) Horses
15 4 1 3 14 5 2 12 10 11 6 13 7 9 8ROI10.012.515.017.520.022.525.027.530.0 T e m p e r a t u r e (b) Donkeys Figure 4: Visualisation of temperatures in ROIs for of all animals: (a) horses;(b) donkeys. ROIs were ordered by their medians.cases are clearly indicated. When applying t-SNE for data visualisation, itsperplexity parameter was set to the value of 5.
Experiments were implemented in Python 3.6.9 using libraries: numpy 1.16.4,scipy 1.3.1, scikit-learn 0.22.1, matplotlib 3.2.2Experiments were conducted using a computer withIntel(R) Core i7-5820K CPU @ 330GHz with 64GB of RAM and with the Win-dows 10 Pro system. The running time of experiments could be measured inseconds.
A comparison of temperatures between ROIs is presented in Fig. 4. The imme-diate observation is that surface temperatures for horses are, on average, higherthan for donkeys, which is confirmed as statistically significant for every ROI(MWW test, see. Sec. 2.3.1). We can see that there are considerable variancesin ROIs temperatures, and many outliers. Fig. 5 presents example histogramsfor two ROIs where differences in mean temperatures of horses and donkeys aremost extreme. We can see how temperature distributions can be multi-modalwhich results from individual differences in animal surface temperatures. His-tograms of temperatures for all ROIs can be found in Fig. 14 in the Appendix.In order to visualise ROI features and assess their potential for distinguishingbetween species we used the t-SNR visualisation. Example results are presentedin Fig. 10. We can see that for the mean or standard deviation of temperaturesin ROIs, examples form two clusters corresponding to the species of animals.However, some examples are in the wrong cluster, which suggests that a subsetof animals could be missclassified. For features based on skewness and kurtosisno consistent structures are observed, which indicates that these features by12 D e n s i t y HorsesDonkeys (a) ROI 4
10 15 20 25Temperature0.000.050.100.150.200.25 D e n s i t y HorsesDonkeys (b) ROI 7
Figure 5: Histograms of temperatures for two ROIs where the difference ∆ t between mean values for the two animal species is: (a) the smallest (ROI 4,∆ t = 1 .
59) and (b) the largest (ROI 7, ∆ t = 5 . H.8,H.13 are globally higher, GOR 8
Abdomen is warm for horses
H.4, H.7, H.8,H.3 and GOR 4
Hindquarter for horses
H.4, H.7, H.8, H.10, H.13 .On the other hand, donkeys seem more uniform, we notice that temperaturevalues in GOR 5
Rump are usually lower than in other GORs, while in GOR 2
Frontquarter we observe warm areas. A comparison of histograms for fourselected GORs is presented in Fig. 12. We can see that the overlap betweenhistograms is more visible for the GOR 5 than for the GOR 2.To highlight the visible patterns, Fig. 11 presents individual thermal mapsfor two example animals. In plots (c) and (d) we can see that characteristicpatterns are usually associated with groups of ROIs rather than an individualROI.Thermal patterns for both species i.e. differences of temperatures betweendesignated GORs, prepared using the methodology described in Sec. 2.3.2 arepresented in Fig. 6. For both species, GORs
Rump and
Legs are consequentlycolder that others while GORs
Neck and
Frontquarter are warmer. The ma-jority of differences are globally significant, for horses there are five excep-tions:
Neck / Frontquarter , Trunk / Ventral aspect , Trunk / Abdomen , Ventral as-pect / Abdomen and
Rump / Legs . For donkeys there are only two exceptions:
Ventral aspect / Abdomen and
Trunk / Groins .13owever, as for the local significance of differences: for horses there are onlyfive patterns that consequently appear for all animals:
Dorsal aspect / Frontquarter , Groins / Trunk , Groins / Hindquarter , Legs / Neck , Legs / Frontquarter . On thecontrary, for donkeys, there are 21 of such patterns, which indicates that don-keys are individually more consistent with the global trend, which will be furtheraddressed in the discussion.A summary of pattern similarities between both species is presented in Fig. 7.Panel (a) presents patterns which are similar for both species e.g. the relationin temperatures between GORs
Rump / Neck is the same for both species (theGOR
Rump is colder than the GOR
Neck ) and is globally, statistically signif-icant, which is indicated with the green colour (the SPS class) in the image.Panel (b) presents the minimal number of animals in each species that sharethe corresponding pattern, e.g. for the pair
Rump / Neck , there are 15 horses and15 donkeys for which the pattern is also locally, statistically significant. We cansee that the majority of patterns fall under the SPS class which supports thethesis about similarities in patterns for both species. In addition, we notice thatdissimilar patterns are most common in GORs
Dorsal aspect and
Trunk . Thepossible explanation of this observation will be discussed in the next section.
Surface temperatures in horses are, on average, higher than in donkeys andtheir individual temperatures vary more within the species. This is largely dueto the differences in the thermal properties of the skin and hair coat. Both,a subcutaneous fat plus skin thickness and the length of hair coat were higherin donkeys than in horses (see values in Tab 1), which may indicate that theyprovide better thermal insulation. Recent results suggests that the hair coatproperties of donkeys and horses differed significantly [16]. Although the authorsindicate, contrary to us, a lower hair length in donkeys than in horses. Thismay be due to the considerable large seasonal variation in hair weight andlength typical for horses, but not for donkeys, or different breeds of horsesparticipating in our research (warmblood horses/ponies) and theirs (UK-nativecold blood horses/ponies)[16]. In other recent studies, the strong relationshipbetween BCS (body condition score) and SF-Skin, for both donkeys and horses,were demonstrated [19, 24]. In our research, the higher SF-Skin thickness indonkeys than in horses may indicate greater adiposity of donkeys and thus betterisolation. As a result, slight local changes in donkey surface body temperaturemay be difficult to observe. This makes the warm area visible in regio scapularisassociated with the GOR 2
Frontquarter particularly interesting. Additionally,it suggests the validity of animal temperatures analysis through comparing thecharacteristics of different regions of a given animal.14 e c k F r o n t . T r un k H i n d . R u m p D o r s . V e n t . A bd o m . G r o i n s L e g s NeckFront.TrunkHind.RumpDors.Vent.Abdom.GroinsLegs 0.05-0.05 -0.15 -0.36 -0.110.15 -0.210.36 0.210.11 (a) M ∆ , Horses N e c k F r o n t . T r un k H i n d . R u m p D o r s . V e n t . A bd o m . G r o i n s L e g s NeckFront.TrunkHind.RumpDors.Vent.Abdom.GroinsLegs -0.12-0.190.190.12 -0.03 1.42 3.14 3.88 0.65 2.06 1.87 1.31 3.170.03 1.45 3.17 3.91 0.68 2.09 1.90 1.34 3.20-1.42 -1.45 1.72 2.46 -0.77 0.64 0.45 1.75-3.14 -3.17 -1.72 0.74 -2.49 -1.09 -1.28 -1.84 0.03-3.88 -3.91 -2.46 -0.74 -3.23 -1.82 -2.01 -2.58 -0.71-0.65 -0.68 0.77 2.49 3.23 1.40 1.22 0.65 2.52-2.06 -2.09 -0.64 1.09 1.82 -1.40 -0.75 1.11-1.87 -1.90 -0.45 1.28 2.01 -1.22 -0.56 1.30-1.31 -1.34 1.84 2.58 -0.65 0.75 0.56 1.87-3.17 -3.20 -1.75 -0.03 0.71 -2.52 -1.11 -1.30 -1.87 (b) M ∆ , Donkeys N e c k F r o n t . T r un k H i n d . R u m p D o r s . V e n t . A bd o m . G r o i n s L e g s NeckFront.TrunkHind.RumpDors.Vent.Abdom.GroinsLegs 0 14 13 15 15 10 9 12 160 14 13 15 16 11 11 12 1614 14 8 9 13 10 11 16 513 13 8 6 6 14 13 16 515 15 9 6 11 15 13 15 815 16 13 6 11 14 12 10 1110 11 10 14 15 14 1 12 139 11 11 13 13 12 1 9 1312 12 16 16 15 10 12 9 1216 16 5 5 8 11 13 13 12 03691215 (c) M L , Horses N e c k F r o n t . T r un k H i n d . R u m p D o r s . V e n t . A bd o m . G r o i n s L e g s NeckFront.TrunkHind.RumpDors.Vent.Abdom.GroinsLegs 0 11 16 16 8 16 13 13 160 14 16 16 11 16 16 13 1611 14 16 16 13 16 14 6 1516 16 16 10 16 16 16 13 716 16 16 10 16 16 16 16 128 11 13 16 16 15 15 9 1616 16 16 16 16 15 5 15 1113 16 14 16 16 15 5 5 1613 13 6 13 16 9 15 5 1516 16 15 7 12 16 11 16 15 03691215 (d) M L , Donkeys Figure 6: Thermal patterns i.e. statistically significant differences betweenGORs (see Sec. 2.3.2). Upper panels present differences within one genre: (a)horses; (b) donkeys. E.g. the value M ∆[4 , = − .
12 in the [4 ,
0] cell in thepanel (a) is the difference between mean temperatures for the pair
Rump / Neck ,indicating that the
Rump
GOR is colder. Bold font indicates ‘global’ statisti-cal significance of this difference. Bottom panels present tables for (c) horses(d) donkeys, with the number of animals for which the corresponding tempera-ture difference in the table above is statistically significant considering individualthermal pattern of this animal. E.g. the value M L [4 , = 15 in panel (c), whichindicates that the pattern Rump / Neck is locally significant for 15 horses. Notethat the most stable patterns should be statistically significant simultaneouslyfor all data combined and for each of the 16 animals of a given species.15 e c k F r o n t . T r un k H i n d . R u m p D o r s . V e n t . A bd o m . G r o i n s L e g s NeckFront.TrunkHind.RumpDors.Vent.Abdom.GroinsLegs SPSSPHWSHWHCSHC (a) Global similarity of patterns N e c k F r o n t . T r un k H i n d . R u m p D o r s . V e n t . A bd o m . G r o i n s L e g s NeckFront.TrunkHind.RumpDors.Vent.Abdom.GroinsLegs 11 13 15 8 10 9 12 1614 13 15 11 11 11 12 1611 14 8 9 13 513 13 8 6 6 14 13 13 515 15 9 6 11 15 13 158 11 13 6 11 14 12 9 1110 11 14 15 14 12 119 11 13 13 12 5 1312 12 13 15 9 12 5 1216 16 5 5 11 11 13 12 03691215 (b) Min. no. animals with a given pattern
Figure 7: Comparison of thermal patterns for both species: (a) division of ther-mal patterns into six classes: SPS – denotes thermal patterns that are similarand globally statistically significant for both species; SP: similar patterns butnot significant; HWS: opposite patterns where horses are warmer (and donkeyscolder), which are statistically significant; HW: same as HWS but not signifi-cant; HCS: significant patterns where horses are colder (and donkeys warmer);HC: same as HCS but not significant. Note that the HWS class is the mostcommon which suggests global similarity of patterns; (b) the minimum numberof animals that confirm the global trend for classes { SP S, HW S, HCS } i.e. forboth species at least this number of animals share a given pattern. Note thatthe patterns where this value is high (e.g. 16, which is the maximum) are morereliable. H.1
H.2
H.3
H.4
H.5
H.6
H.7
H.8
H.9
H.10
H.11
H.12
H.13
H.14
H.15
H.16
Figure 8: Thermal maps of annotated ROIs for horses in our dataset.16 .1 D.2
D.3
D.4
D.5
D.7
D.7
D.8
D.9
D.10
D.11
D.12
D.13
D.14
D.15
D.16
Figure 9: Thermal maps of annotated ROIs for donkeys in our dataset.
We determined that patterns in IRT images from our data set are often visiblein groups of ROIs and proposed a methodology of assessing these patterns basedon the difference of temperatures in groups. Visualisation in Fig. 7 shows thatthese differences are similar for both species: looking at the plot (a), we can seethat 77 .
8% of patterns are similar and statistically significant, 8 .
9% of patternsare opposite and the rest of them cannot be confirmed statistically based on ourdata. We also see that for 88 .
8% of globally significant patterns, half or moreindividual animals from every species share this pattern. In our opinion, thissupports the thesis about similarities in IRT images of both species.Analysing the values in Fig. 6 we can see that the observation that donkeysare more ‘uniform’ in their GORs, results in larger maximum differences betweenGORs and the fact that more individual animals share the global trend then forhorses.As for opposite thermal patterns, we notice that they are usually associateswith GORs:
Dorsal aspect and
Trunk . Both groups cover a relatively large areaof the animal’s body, which raises the question of whether the more granularsegmentation of these areas will show further similarities.An important question is, whether the trends observed for our data set arecharacteristic of entire populations. As the number of cases is limited by practi-cal considerations, we believe that our results should be treated as a significantindication of the existence of the relationships we described. At the same time,we emphasize the need to further verify these conclusions for more data. Tofacilitate this, all research data related to this study is available to the public17 (a) Mean
200 150 100 50 0 50 1003002001000100200300 HorsesDonkeys (b) Std
150 100 50 0 50 100 150 2003002001000100200 HorsesDonkeys (c) Kurtosis
300 200 100 0 100 200 300 400 5004002000200400600 HorsesDonkeys (d) Mean (normalised)
Figure 10: T-SNR visualisation of the data set. Every dot represents an animaldescribed with features extracted from pixels of its 15 ROIs. Plots presentdifferent feature extraction statistics: (a) the mean; (b) the standard deviation;(c) the kurtosis; (d) the mean, after removing the global mean temperatureof an animal from all pixel values. Notice that features in plots (a) and (b)seem more distinctive for both species which results in more apparent clusters.However, the classes of examples in these clusters are mixed.18 .11 (a) Horse
H11 , t c ∈ (cid:104) . , . (cid:105) D.12 (b) Donkey,
D12 , t c ∈ (cid:104) . , . (cid:105) H.11 (c) Horse,
H11 , t c ∈ (cid:104) . , (cid:105) D.12 (d) Donkey,
D12 , t c ∈ (cid:104) . , . (cid:105) Figure 11: Selected examples of two animals from our dataset. The color mapvalues t c for images in the upper row are scaled to the common range, whichmakes them easy to compare: (a) horses; (b) donkeys. Images in the bottom roware scaled to the minimal and maximal temperatures in annotated ROIs of eachanimal, which highlights individual thermal patterns: (c) horses; (d) donkeys.E.g. warm horse’s GORs Abdomen and
Neck , cool donkey’s GOR
Rump , andwarm donkey’s GOR
Frontquarter .
10 20 30Temperature0.00.10.2 D e n s i t y (a) Frontquarter
10 20Temperature0.00.10.2 D e n s i t y (b) Hindquarter
10 20Temperature0.00.10.2 D e n s i t y (c) Rump
20 30Temperature0.00.10.2 D e n s i t y (d) Groins
Figure 12: Comparison of temperature histograms between animal species inidentified characteristic areas corresponding to selected groups of ROIs: (a)GOR 2
Frontquarter ; (b) GOR 4
Hindquarter ; (c) GOR 5
Rump ; (i) GOR 9
Groins . Horses are represented in red, donkeys are blue.19nder open licenses.
The two animals identified as outliers, i.e.
D.17-18 allow for an interestingstudy of how general or specific our proposed thermal patterns are. The visual-ization of the differences between the patterns of these animals and the rest ofthe donkeys is presented in Fig. 13. In panels (b, c) the green color indicates thecompliance of the animal thermal pattern with the global trend, i.e. the differ-ence for a given pair of GORs has the same sign and is statistically significantfor the animal. As we can see, thermal patterns for the
D.17 donkey with athick hair coat are usually in line with the global pattern, except for four pairsof GORs, where the statistical significance of the local animal pattern could notbe confirmed. The patterns of the donkey
D.18 are much less in line with theglobal pattern. This is an expected result, as the losses in the hair coat visiblyaffect its thermal characteristics in the image. This also suggests that the pro-posed thermal patterns may be the basis for creating temperature indexes ase.g. in [30] or features for detecting anomalies. Therefore, we speculate that alot of research methods applied successfully in equine veterinary medicine canalso be used for donkeys IRT imaging, provided that the visualization conditionsdescribed by us are met.The temperature difference matrices for these two animals are providedin Fig.15 in the Appendix.
We observed that characteristic thermal patterns of both horses and donkeysare usually associated with groups of ROIs (GORs) rather than an individualROI. Based on this observation we defined a thermal pattern as a statisticallysignificant difference between designated GORs for a given animal species. Wehave verified this significance both globally, for all data and locally, for individualanimals. We have shown how the majority of proposed thermal patterns aresimilar for both species. Noteworthy, the thermal patterns for donkeys are moreuniform then for horses, and donkeys are individually more consistent with theglobal trend. Note that, the proposed thermal patterns compare data fromone species or individual animals – in general, for animals form our data set,average surface temperatures for horses are higher than for donkeys which maybe related to differences in thermal properties of the skin and hair coat.
This work was conducted in the Veterinary Research Centre WULS (WCB)and the Center for Biomedical Research (CBB) supported by EFRR RPO WM2007–2013. Authors are grateful to the Mrs. and Mr. S(cid:32)lupski, owners of the20 .17 (a)
D.17
D.18 (b)
D.18 N e c k F r o n t . T r un k H i n d . R u m p D o r s . V e n t . A bd o m . G r o i n s L e g s NeckFront.TrunkHind.RumpDors.Vent.Abdom.GroinsLegs SNS (c)
D.17 and Donkeys N e c k F r o n t . T r un k H i n d . R u m p D o r s . V e n t . A bd o m . G r o i n s L e g s NeckFront.TrunkHind.RumpDors.Vent.Abdom.GroinsLegs SNS (d)
D.18 and Donkeys
Figure 13: Visualisation of differences between donkeys
D.17 and
D.18 , whichwere identified as outlier cases (see Sec. 2.1), and the rest of the donkeys i.e.animals
D.1-16 . The upper panels present thermal maps of the two cases: (a)Donkey
D.17 ; (b) Donkey
D.18 . Notice that
D.17 is colder than other animalsdue to its long hair length and that
D.18 has an unusual pattern of warm areasresulting from plaque loses in the hair coat. Bottom plots show differences intheir thermal patterns compared to the global pattern of other donkeys: (c)Donkey
D.17 compared to Donkeys; (d) Donkey
D.18 compared to Donkeys.The S class (green) indicates that the individual animal pattern is in line withthe global trend, class NS (red) indicates the opposite.21aria˙z agritourism farm in Lubochnia for letting them observe donkeys andhorses and for their help on site.
15 20 25[ROI 1] Temp.0.00.10.2 D e n s i t y
10 20 30[ROI 2] Temp.0.00.10.2 D e n s i t y
20 30[ROI 3] Temp.0.00.10.2 D e n s i t y
20 30[ROI 4] Temp.0.00.10.2 D e n s i t y
20 30[ROI 5] Temp.0.00.10.2 D e n s i t y
10 20[ROI 6] Temp.0.00.10.2 D e n s i t y
10 20[ROI 7] Temp.0.00.10.2 D e n s i t y
10 20[ROI 8] Temp.0.00.2 D e n s i t y
10 20[ROI 9] Temp.0.00.10.2 D e n s i t y
15 20 25[ROI 10] Temp.0.00.2 D e n s i t y
20 30[ROI 11] Temp.0.00.10.2 D e n s i t y
20 30[ROI 12] Temp.0.00.1 D e n s i t y
15 20 25[ROI 13] Temp.0.00.10.2 D e n s i t y
15 20 25[ROI 14] Temp.0.00.10.2 D e n s i t y
15 20 25[ROI 15] Temp.0.00.2 D e n s i t y
15 20 25[ROI 15] Temp.0.00.2 D e n s i t y Figure 14: Histograms of temperatures for all ROIs, the red colour denoteshorses, the blue colour denotes donkeys. The last plot (the LR corner) presentsthe combined histogram for all ROIs.
References [1] TZ Arruda, KE Brass, and D Flavio. Thermographic assessment of saddlesused on jumping horses.
Journal of Equine Veterinary Science , 31(11):625–629, 2011.[2] B B¨achi, T Wiestner, A Stoll, NM Waldern, I Imboden, and MA Weishaupt.Changes of ground reaction force and timing variables in the course of ha-bituation of horses to the treadmill.
Journal of Equine Veterinary Science ,63:13–23, 2018.[3] E Bartolom´e, MJ S´anchez, A Molina, AL Schaefer, I Cervantes, andM Valera. Using eye temperature and heart rate for stress assessment22 e c k F r o n t . T r un k H i n d . R u m p D o r s . V e n t . A bd o m . G r o i n s L e g s NeckFront.TrunkHind.RumpDors.Vent.Abdom.GroinsLegs -0.420.42 -0.01-2.38 -0.56-0.820.82 -0.750.01 2.38 0.750.56 (a) M ∆ , D.17 N e c k F r o n t . T r un k H i n d . R u m p D o r s . V e n t . A bd o m . G r o i n s L e g s NeckFront.TrunkHind.RumpDors.Vent.Abdom.GroinsLegs -0.42 0.530.42 -0.410.02 0.50-0.02 -0.44 -0.27-0.53 0.41 0.44 0.17 0.930.27 -0.17-0.50 -0.93 (b) M ∆ , D.18
Figure 15: Thermal patterns i.e. differences between GORs for the two donkeysidentified as outliers (see. Sec. 2.1): (a) Donkey
D.17 ; (b) Donkey
D.18 . Boldfont indicates the statistical significance of the difference for the given pattern.in young horses competing in jumping competitions and its possible influ-ence on sport performance.
Animal: an International Journal of AnimalBioscience , 7(12):2044, 2013.[4] M Becker-Birck, A Schmidt, M Wulf, J Aurich, A Von der Wense, E M¨ostl,R Berz, and C Aurich. Cortisol release, heart rate and heart rate variability,and superficial body temperature, in horses lunged either with hyperflexionof the neck or with an extended head and neck position.
Journal of AnimalPhysiology and Animal Nutrition , 97(2):322–330, 2013.[5] MA Cetinkaya and A Demirutku. Thermography in the assessment ofequine lameness.
Turkish Journal of Veterinary and Animal Sciences ,36(1):43–48, 2012.[6] O Ciutacu, A Tanase, and I Miclaus. Igital infrared thermography in assess-ing soft tissues injuries on sport equines.
Bulletin of University of Agricul-tural Sciences and Veterinary Medicine Cluj-Napoca. Veterinary Medicine ,63(1-2), 2006.[7] B Derrick and P White. Comparing two samples from an individual likertquestion.
International Journal of Mathematics and Statistics , 18(3), 2017.[8] AL Eddy, LM Van Hoogmoed, and JR Snyder. The role of thermography inthe management of equine lameness.
The veterinary journal , 162(3):172–181, 2001.[9] BPA Fonseca, ALG Alves, JLM Nicoletti, A Thomassian, CA Hussni, andS Mikail. Thermography and ultrasonography in back pain diagnosis ofequine athletes.
Journal of Equine Veterinary Science , 26(11):507–516,2006. 2310] CB G´omez ´Alvarez, M Rhodin, A Bystr¨om, W Back, and PR Van Weeren.Back kinematics of healthy trotting horses during treadmill versus overground locomotion.
Equine veterinary journal , 41(3):297–300, 2009.[11] KW Hinchcliff, AJ Kaneps, and RJ Geor.
Equine Sports Medicine andSurgery E-Book . Elsevier Health Sciences, 2013.[12] H Hotelling. Analysis of a complex of statistical variables into principalcomponents.
Journal of educational psychology , 24(6):417, 1933.[13] G Kastberger and R Stachl. Infrared imaging technology and biologi-cal applications.
Behavior Research Methods, Instruments, & Computers ,35(3):429–439, 2003.[14] LFS Machado, RL Dittrich, M Pavelski, AMC da F Afonso, I Deconto,PT Dornbusch, et al. Standardization of thermographic examination injoints of horses in training.
Archives of Veterinary Science , 18(4):40–45,2013.[15] M Masko, L Zdrojkowski, M Domino, T Jasinski, and Z Gajewski. Thepattern of superficial body temperatures in leisure horses lunged with com-monly used lunging aids.
Animals , 9(12):1095, 2019.[16] B Osthaus, L Proops, S Long, N Bell, K Hayday, and F Burden. Hair coatproperties of donkeys, mules and horses in a temperate climate.
Equineveterinary journal , 50(3):339–342, 2018.[17] M Pavelski, MS Basten, E Busato, and PT Dornbusch. Infrared ther-mography evaluation from the back region of healthy horses in controlledtemperature room.
Ciˆencia Rural , 45(7):1274–1279, 2015.[18] R Purohit. Standards for thermal imaging in veterinary medicine, 2009.[19] M Quaresma, R Payan-Carreira, and SR Silva. Relationship between ul-trasound measurements of body fat reserves and body condition score infemale donkeys.
The Veterinary Journal , 197(2):329–334, 2013.[20] V Redaelli, F Luzi, S Mazzola, GD Bariffi, M Zappaterra, L Nanni Costa,and B Padalino. The use of infrared thermography (irt) as stress indicatorin horses trained for endurance: A pilot study.
Animals , 9(3):84, 2019.[21] JG Rosenmeier, AB Strathe, and PH Andersen. Evaluation of coronaryband temperatures in healthy horses.
American journal of veterinary re-search , 73(5):719–723, 2012.[22] G Satchell, M McGrath, J Dixon, T Pfau, and R Weller. Effects of timeof day, ambient temperature and relative humidity on the repeatabilityof infrared thermographic imaging in horses.
Equine Veterinary Journal ,47:13–14, 2015. 2423] HJ Seeherman. The use of high-speed treadmills for lameness and hoofbalance evaluations in the horse.
Veterinary Clinics of North America:Equine Practice , 7(2):271–309, 1991.[24] SR Silva, R Payan-Carreira, M Quaresma, CM Guedes, and AS Santos.Relationships between body condition score and ultrasound skin-associatedsubcutaneous fat depth in equids.
Acta Veterinaria Scandinavica , 58(1):37–42, 2016.[25] EL Simon, EM Gaughan, T Epp, and M Spire. Influence of exercise onthermographically determined surface temperatures of thoracic and pelviclimbs in horses.
Journal of the American Veterinary Medical Association ,229(12):1940–1944, 2006.[26] WM Smith. Applications of thermography in veterinary medicine.
Annalsof the New York Academy of Sciences , 121:248, 1964.[27] M Soroko and K Howell. Infrared thermography: current applications inequine medicine.
Journal of Equine Veterinary Science , 60:90–96, 2018.[28] M Soroko, K Howell, and K Dudek. The effect of ambient temperature oninfrared thermographic images of joints in the distal forelimbs of healthyracehorses.
Journal of Thermal Biology , 66:63–67, 2017.[29] M Soroko, E Jodkowska, and M Zab(cid:32)locka. The use of thermography toevaluate back musculoskeletal responses of young racehorses to training.
Thermology International , 22(3):114, 2012.[30] M Soroko, D Zaborski, K Dudek, K Yarnell, W G´orniak, and R Vardasca.Evaluation of thermal pattern distributions in racehorse saddles using in-frared thermography.
PloS one , 14(8):e0221622, 2019.[31] BV Tunley and FMD Henson. Reliability and repeatability of thermo-graphic examination and the normal thermographic image of the thora-columbar region in the horse.
Equine veterinary journal , 36(4):306–312,2004.[32] TA Turner. Diagnostic thermography.
Veterinary Clinics of North Amer-ica: Equine Practice , 17(1):95–114, 2001.[33] Mercedes Valera, Ester Bartolom´e, Maria Jos´e S´anchez, Antonio Molina,Nigel Cook, and AL Schaefer. Changes in eye temperature and stressassessment in horses during show jumping competitions.
Journal of EquineVeterinary Science , 32(12):827–830, 2012.[34] L Van der Maaten and G Hinton. Visualizing data using t-sne.
Journal ofmachine learning research , 9(Nov):2579–2605, 2008.2535] S Westermann, HHF Buchner, JP Schramel, A Tichy, and C Stanek. Effectsof infrared camera angle and distance on measurement and reproducibilityof thermographically determined temperatures of the distolateral aspectsof the forelimbs in horses.
Journal of the American Veterinary MedicalAssociation , 242(3):388–395, 2013.[36] FO Zakari, JO Ayo, PI Rekwot, MU Kawu, and NS Minka. Daily rhythmsof rectal and body surface temperatures in donkeys during the cold-dry(harmattan) and hot-dry seasons in a tropical savannah.