Snekhalatha-Anbu, Venkatraman_ Journal of Engineering in Medicine-2015

Original Article

Proc IMechE Part H:J Engineering in Medicine2015, Vol. 229(4) 319331 IMechE 2015Reprints and permissions:sagepub.co.uk/journalsPermissions.navDOI: 10.1177/0954411915580809pih.sagepub.com

Automated hand thermal imagesegmentation and feature extraction inthe evaluation of rheumatoid arthritis

U Snekhalatha1, M Anburajan1, V Sowmiya2, B Venkatraman3 andM Menaka3

AbstractThe aim of the study was (1) to perform an automated segmentation of hot spot regions of the hand from thermographusing the k-means algorithm and (2) to test the potential of features extracted from the hand thermograph and its mea-sured skin temperature indices in the evaluation of rheumatoid arthritis. Thermal image analysis based on skin tempera-ture measurement, heat distribution index and thermographic index was analyzed in rheumatoid arthritis patients andcontrols. The k-means algorithm was used for image segmentation, and features were extracted from the segmentedoutput image using the gray-level co-occurrence matrix method. In metacarpo-phalangeal, proximal inter-phalangeal anddistal inter-phalangeal regions, the calculated percentage difference in the mean values of skin temperatures was foundto be higher in rheumatoid arthritis patients (5.3%, 4.9% and 4.8% in MCP3, PIP3 and DIP3 joints, respectively) as com-pared to the normal group. k-Means algorithm applied in the thermal imaging provided better segmentation results inevaluating the disease. In the total population studied, the measured mean average skin temperature of the MCP3 jointwas highly correlated with most of the extracted features of the hand. In the total population studied, the statistical fea-ture extracted parameters correlated significantly with skin surface temperature measurements and measured tempera-ture indices. Hence, the developed computer-aided diagnostic tool using MATLAB could be used as a reliable method indiagnosing and analyzing the arthritis in hand thermal images.

KeywordsHeat distribution index, thermographic index, rheumatoid arthritis, k-means algorithm

Date received: 1 September 2014; accepted: 13 March 2015

Introduction

Rheumatoid arthritis (RA) is a chronic autoimmunedisease that causes inflammation of blood vessels,development of bumps called rheumatoid nodules,weakening of bones and deformity of the joints leadingto long-term disability. It causes premature mortality,disability and compromised quality of life in the indus-trialized and developing world.1 The prevalence of RAranges from 0.5% to 1% worldwide in the general pop-ulation.1 The prevalence rate in India is 0.9%, almostequal to the world prevalence rate.2 RA is associatedwith severe disability and substantial morbidity.3,4

Symmons and Gabriel5 reported that mortality isgreater in patients with established RA in comparisonwith the general population. RA directly affects physi-cal function and mobility and results in substantialshort-term and long-term morbidity.

Thermal imaging is a non-contact, noninvasive,diagnostic imaging procedure based on skin

temperature measurement. It is a functional imagingmethod for analyzing physiological function related tobody temperature. Also, it has been used as a compli-mentary diagnostic tool in various clinical applicationsof rheumatology such as diagnosis of patellafemoralarthralgia6 and monitoring of skin temperature eleva-tions in scleroderma,7 RA,8 juvenile RA9 and hand andknee osteoarthritis.10,11

Jiang et al.12 hypothesized that thermography mayprovide a sensitive, noninvasive method to find disease

1Department of Biomedical Engineering, SRM University, Chennai, India2Department of Biomedical Engineering, Jerusalem College of

Engineering, Anna University, Chennai, India3Indra Gandhi Center for Atomic Research, Kalpakkam, India

Corresponding author:

M Anburajan, Department of Biomedical Engineering, SRM University,

Kattankulathur 603203, Chennai, Tamil Nadu, India.

Email: [email protected]

by guest on May 1, 2015pih.sagepub.comDownloaded from

severity both in animal models and in human studies.Glehr et al.13 used computer-assisted infrared thermo-graphy to diagnose anterior knee pain followingimplantation of artificial knee joints. Trafarski et al.14

demonstrated that thermography is useful in analyzingthe thermal reaction of the organism connected withlocal cryostimulation and monitoring and evaluatingthe progress of treatment in RA patients.

Thermography has been used as an effective tool inquantification of inflammation in hand joints of RA. Ithas been evident that thermographic changes in skintemperature over the areas of chronic inflammationaccurately reproduce the changes in other chemical andcellular process in RA.15 Various studies have been per-formed in last and recent decade for the assessment ofRA using thermography in hand regions.1618

Thermographic index (TI) and heat distribution index(HDI) have been used as quantitative measurement ofinflammation in rheumatoid joint.19,20

Several boundary edge detection techniques such asprobability-based edge detection techniques and cannyedge detection techniques have been applied in thermalimages for the extraction of region of interest (ROI).21

The edge detection approach is not suitable for seg-menting the thermal images because additional effort isrequired to connect the incomplete edges into a distinctand meaningful object boundary. Hence, it requires aneed of clustering-based segmentation techniques forextracting the ROI. Thermograms, either grayscale orpseudo color, are processed for detection of abnormalregions and quantification. However, temperature var-iations are not visible to naked eye. Hence, it is neces-sary to develop and analyze feature extractionalgorithms for abnormality detection and classificationof disease state and normal.22

The aim of our study was (1) to perform an auto-mated segmentation of hot spot region of the handfrom thermographs using the k-means algorithm and(2) to test the potential of features extracted from thehand thermograph and its measured skin temperatureindices in the evaluation of RA.

Patients and methods

Study design and population

This study was approved by our institutional ethicalcommittee (approval no. 35/iec/2010), and the informedconsent statement was signed by each participant. Afree medical screening camp to identify RA in the sub-urban South Indian population was conducted at SRMHospital and Research Centre, Kattankulathur,Chennai, Tamil Nadu, India, on 1519 February 2011.Patients with major systemic disease were included inthe study. However, patients who had undergone recentphysiotherapy, thermotherapy, phototherapy and feverwere excluded. According to the Indian Rheumatology

Association (IRA) consensus report 2008,23 individualswith persistent inflammatory arthritis with more thanfour joints, high Erythrocyte Sedimentation Rate(ESR) (RA . 20mm/h)/high C-reactive protein (CRP)(RA . 80mg/dL), positive immunoglobulin M (Igm)rheumatoid factor (RA . 20U/mL), radiographicchanges in juxta-articular osteopenia, erosions and jointspace narrowing were classified as RA. In all thepatients, both the hands were symmetrically affectedby RA.

A total number of 50 subjects, age ranging from 30to 75 years, were registered in the camp. According tothe diagnostic criteria of the study, 30 subjects (male/female (M/F) ratio=1:3) were found to have knownRA, and their mean age was found to be 45.36 11.4years. They had disease duration of 4.36 2.5 years, 15age- and sex-matched normal subjects (M/F ratio=1:3), their mean age was found to be 45.56 11.8 years,were included in the study. The remaining five subjectswere excluded according to the exclusion criteria of thestudy.

Thermal image acquisition procedure and analysis

The acquisition of thermal image was performedaccording to the guidelines recommended by theInternational Academy of Clinical Thermology.24 Allthe subjects were asked to remove all the ornamentsand were seated with their hands exposed for 15 min ina temperature-controlled room at 20 C with humidityof 45%50%. The subjects were then asked to stand inthe image capturing room with their hands kept in bothdorsal view and ventral view. The thermal camera waspositioned at a distance of 1.0m parallel to the hand,and a thermal image of both dorsal and ventral viewsof both right and left hand was taken in 5 s. All thethermal images were acquired in the same room for aperiod of 1 week in the forenoon session only, to avoidthe effect of cyclic variations in the atmospheric tem-perature in the study.

A hand-held thermal camera (ThermaCAM-T400;FLIR, Wilsonville, OR, USA) was used to image thehand region of both the groups of RA and normal.The thermal camera T400 utilizes the 3203 420 ther-mal element focal plane array (FPA) uncooled micro-bolometer detector system with minimum focusabledistance of 0.4m to infinity. The camera could measurethe temperature range of 220 C to 1200 C to anaccuracy of 2% with thermal sensitivity of 0.05 C. Theimages were stored and then analyzed using the soft-ware FLIR Quick report, version-1.2 and further pro-cessed with MATLAB version7.1 (MathWorks Inc.,Natick, MA, USA).

From the thermal image of the hand, the averageskin temperature (C) of a total of 28 index joints,which includes 10 distal inter-phalangeal (DIP) joints,8 proximal inter-phalangeal (PIP) joints and 10

320 Proc IMechE Part H: J Engineering in Medicine 229(4)


metacarpo-phalangeal (MCP) joints of both the handstaken in dorsal view, was evaluated using FLIR soft-ware. It was measured by positioning an availablesquare area tool of ROI of size 83 8mm in eachrespective joint in dorsal view. Furthermore, the aver-age skin surface temperature was measured in the ven-tral side (palm region).

Temperature indices

The HDI was defined as twice the standard deviation(SD) of all the temperatures within the ROI of theMCP joint. It represents the pattern and spread of tem-perature across joints.25 TI describes the peak tempera-ture of the joint and the area of each temperature band.It was used as a quantitative evaluation of skin thermalchanges in the ROI. The TI was calculated using theformula

TI=(PDT3a)

A1

whereP

represents the summation, DT represents thedifference between the mean skin temperature over themonitored area and the temperature of the isothermand a indicates the area (cm2) occupied by isothermand A represents the total area (cm2) of thermogram.The TI values were used to diagnose the individual, par-ticipated in the study as follows: (1)42: normal, (2) 34: osteoarthritis and (3) . 4: RA.26

Thermal image segmentation algorithm

Automated thermal image segmentation was performedby applying k-means algorithm to hand thermal imagesof both dorsal and ventral views. k-Means was anunsupervised clustering algorithm which classifies thedata based on k number of groups based on image fea-tures. The grouping was done by minimizing the sumof squares of distance between data points and the cor-responding cluster centroids.

A summary of the k-means algorithm27 is given hereas follows:

1. Input image was the hand thermal image taken indorsal/ventral view.

2. The RGB image was converted to HSV colorspace.The reasons for conversion of RGB to HSV weregiven as follows: (1) RGB defines color in terms ofa combination of primary colors, whereas HSVdescribes it using more familiar comparisons suchas color, vibrancy and brightness; (2) furthermore,HSV describes color similarly to how the humaneye tends to perceive it.

3. Three classes (k=3) were chosen and assigned asinitial centroidsClass 1: cluster 1; Class 2: cluster2; Class 3: cluster 3.

4. The distance between the centroid and each pixelof the input was computed.

5. According to the minimum distance criterion, thedata were clustered as follows:

(a) Cluster 1 to separate hot spot region;(b) Cluster 2 to separate background region;(c) Cluster 3 to separate other than hot spot

regions as well as background.6. The three centroids using the average of all HSV

points in each cluster were updated.7. Steps 16 were repeated until no points switch to a

new cluster, or until we hit a maximum number of25 iterations.

Statistical feature extraction

The feature extraction technique is implemented overthe segmented output image to extract the intensity fea-tures and statistical texture features. The features suchas mean, SD, smoothness, entropy, kurtosis, skewness,variance, contrast, correlation and energy are extractedfrom the segmented output image using gray-level co-occurrence matrix (GLCM) method28 and explained asfollows:

1. MeanThe mean is the average intensity value of the hotspot regions and was given as

Mean=X

iX

jiP(i, j) 2

2. SDThe SD is a measure of dispersion or variationfrom the mean value. A low SD represents thedata points to be very close to the mean valueand a high SD represents the data points arespread out over the large range of values

s=XN1i=0

XN1j=0

(i m)2p(i, j) !1=2

3

3. EntropyEntropy is the measure of randomness that canbe used to characterize the texture of the image.It is a quantitative measure of image information.If all the image pixels have the same gray level,minimum entropy is achieved; if the pixels have auniform distribution of gray levels or the image ishistogram equalized, maximum entropy isachieved

Entropy=X

iX

jP(i, j) log P(i, j) 4

4. SkewnessSkewness is the measure of the asymmetry of thepixel distribution around its mean. A symmetricdistribution has zero skewness and has equal val-ues for mean, median and mode. If the skewnessis positive, then the data are positive skewed orskewed right, meaning that the right tail of the

Snekhalatha et al. 321


distribution is longer than the left, the mean isgreater than the median and the mode is less thanthe median. If the skewness is negative, the dataare negatively skewed, meaning that the left tailof the distribution is longer than the right, themean is less than the median and the mode isgreater than the median

Skewness=

PP(i, j) m 2=N

h is3

5

5. KurtosisKurtosis refers to the measure of the peak of thearray of intensity distribution. Similar to skew-ness, kurtosis is a descriptor of the shape of theprobability distribution. There are three differentinterpretations of kurtosis: mesokurtic distribu-tion, leptokurtic distribution and platykurtic dis-tribution. A zero excess kurtosis indicates themesokurtic distribution which has a normalpeak around the mean. A positive excess kurtosisrepresents the leptokurtic distribution which has amore acute peak around the mean. A negativeexcess kurtosis shows the platykurtic distributionwhich has a lower, wider peak around the mean29

Kurtosis=X

P(i, j) m 4=Nh i

=s4h i

36

6. VarianceThe variance is the square of the SD. A smallvariance indicates that the data points are closeto the mean, whereas the large variancemeans that the data points are spread out fromthe mean

Var=X

iX

j(i m)2P(i, j) 7

7. ContrastThe contrast is the measure of the intensity con-trast between the pixel and its neighbors

Contrast=Xi=0

Xi=0

i, j2Pi, j 8

8. EnergyEnergy is given by the sum of the squared ele-ments in the GLCM. Energy measures the tex-tural uniformity that means pixel pairsrepetitions. It also measures the smoothness ofthe image. For more uniformly distributed pixelsin image regions, the smoothness level and energywere low. But in case of the nonuniform distribu-ted region, the smoothness and energy are high.Hence, in case of RA patients, due to uneven dis-tribution of temperature in hand regions, theenergy obtained was higher compared to thenormal30

Energy=Xi, j=0

(Pi, j)2 9

9. CorrelationCorrelation is a measure of linear dependencies ofgray level on those of neighboring pixels

Correlation=Xi=0

Xj=0

(i m)(j m)P(i, j)sisj

10

10. HomogeneityHomogeneity is used to determine whether fre-quency counts are distributed identically acrossdifferent populations

Homogenity=XN1i, j=0

pij

1+ i j 2 11

Statistical analysis

Data were analyzed using SPSS software package ver-sion 19.0 (SPSS Inc., Chicago, IL, USA). The measuredmean surface temperature, HDI, TI and feature extrac-tion parameters were compared between the RA groupand normal group using a Students t-test. TheKolmogorovSmirnov and ShapiroWilk tests wereperformed to test the normality of above-mentionedvariables which gave a significance value (p \ 0.05).The bi-variate (Pearson correlation) analyses were usedto obtain the correlations between HDI, TI, skin tem-perature measurements and feature extraction para-meters in the total population studied.

Results

HDI and TI analysis

The measured temperature indices correlated signifi-cantly with the feature extracted parameters in the totalpopulation studied (Table 1). The mean values of mea-sured HDI and TI in RA patients were highly signifi-cant (p \ 0.01) when compared to the normal group.The mean (6SD) HDI values of both RA and normalgroups were obtained as 0.79 (60.2) and 0.38 (60.1),respectively, and found to be statistically significant.Also, the measured mean (6SD) values of TI obtainedfor the RA group and normal groups were 4.18 (60.5)and 2.69 (60.5), respectively (Table 2). The percentageof patients whose diagnostic values of TI categorizedinto normal, osteoarthritis and RA was 24.4%, 8.8%and 66.6%, respectively.

Thermal image analysis and feature extraction

In the total population (n=45), the temperatureindices such as HDI and TI correlated significantly (p\ 0.01) with the feature extracted parameters(Table 1). The mean average skin temperature ofMCP1, MCP3, MCP4 and MCP5 correlated



Table 1. Correlation matrix between temperature parameter and statistical feature extracted parameter measured from handthermal image in total population studied.

Temperatureparameters

Feature extraction parameters

Mean Standarddeviation

Skewness Entropy Kurtosis Variance Contrast Correlation Energy Homogeneity

Thermal indicesHDI 0.4** 0.42** 0.39* 0.45** 0.39* 0.38* 0.56** 0.44** 0.44** 0.45**TI 0.46** 0.51** 0.44** 0.47** 0.44** 0.33* 0.61** 0.53** 0.54** 0.54**

Skin temperature at MCP joints, PIP joints and DIP jointsMCP1 0.45** 0.24# 0.32* 0.31* 0.25# 0.42** 0.25# 0.30* 0.33* 0.39*MCP2 0.46** 0.25# 0.34* 0.21# 0.20# 0.40** 0.27# 0.26# 0.28* 0.39*MCP3 0.48** 0.51** 0.52** 0.51** 0.43** 0.31* 0.55** 0.56** 0.56** 0.56**MCP4 0.46** 0.48** 0.48** 0.49** 0.42** 0.26 0.53** 0.30* 0.30* 0.41**MCP5 0.41** 0.26 0.42** 0.39* 0.26# 0.39* 0.50** 0.33* 0.28* 0.46**PIP2 0.3* 0.29* 0.37* 0.27# 0.48** 0.24# 0.29* 0.52** 0.3* 0.32*PIP3 0.27# 0.38* 0.41** 0.26# 0.46** 0.32* 0.38* 0.56** 0.32* 0.34*PIP4 0.27# 0.26# 0.38* 0.26# 0.48** 0.24# 0.26 0.51** 0.29* 0.31*PIP5 0.3* 0.3* 0.37* 0.31* 0.47** 0.29* 0.25# 0.45** 0.28* 0.29*DIP1 0.27# 0.09# 0.11# 0.2# 0.05# 0.019# 0.005# 0.03# 0.016# 0.05#

DIP2 0.24# 0.2# 0.24# 0.25\# 0.22# 0.05# 0.16# 0.21# 0.12# 0.16#

DIP3 0.23# 0.23# 0.27# 0.26# 0.21# 0.06# 0.15# 0.23# 0.14# 0.18#

DIP4 0.23# 0.16# 0.31* 0.25# 0.17# 0.0002# 0.07# 0.24# 0.16# 0.19#

DIP5 0.22# 0.15# 0.26# 0.26# 0.14# 0.22# 0.03# 0.22# 0.12# 0.16#

HDI: heat distribution index; TI: thermographic index; MCP: metacarpo-phalangeal; PIP: proximal inter-phalangeal; DIP: distal inter-phalangeal; NS:

not significant.

*p \ 0.05; **p \ 0.01; #NS.

Table 2. Measured mean average skin surface temperature of RA and normal in required region of interest.

Temperatureparameters

Region ofinterest

RA (N= 30),mean6 SD

Normal (N= 15),mean6 SD

% Difference Statisticalsignificance (p)

Thermal indices

HDI Whole hand 0.796 0.2 0.386 0.1 51.89 0.001TI Whole hand 4.186 0.5 2.696 0.5 35.64 0.001

Skin temperature measurements at jointsMCP MCP1 35.396 0.6 33.936 0.3 3.84 0.002

MCP2 35.406 0.6 33.866 0.2 4.35 0.003MCP3 35.526 0.7 33.646 0.2 5.29 0.001MCP4 35.336 0.6 33.766 0.2 4.44 0.003MCP5 35.406 0.6 33.866 0.2 4.35 0.004

PIP PIP2 35.426 0.6 33.806 0.2 4.79 0.002PIP3 35.446 0.4 33.716 0.2 4.88 0.001PIP4 35.396 0.4 33.726 0.2 4.71 0.004PIP5 35.226 0.7 33.676 0.1 4.60 0.001

DIP DIP1 35.216 0.5 33.876 0.4 3.80 0.005DIP2 35.406 0.6 33.866 0.2 4.35 0.001DIP3 35.246 0.6 33.556 0.4 4.79 0.001DIP4 35.226 0.5 33.576 0.3 4.68 0.004DIP5 35.156 0.5 33.756 0.1 3.98 0.005

Ventral side palm 35.396 0.7 34.436 0.2 2.71 0.004Measured statistical featuresMean Whole hand 5.03E2 056 5.868E2 05 2.89E2 066 1.08E2 06 94.25 0.0003Std dev Whole hand 8.54E2 056 7.63E2 05 2.08E2 056 9.58E2 06 74.59 0.0002Skewness Whole hand 0.0066 0.006 0.0016 0.0008 8.33 0.0009Entropy Whole hand 0.0026 0.002 0.00016 5.61E2 05 95 0.0002Kurtosis Whole hand 0.1286 0.15 0.0086 0.007 93.75 0.0004Smoothness Whole hand 5.1941E2 116 1.368E2 10 4.06E2 136 8.47E2 13 50.25 0.06Variance Whole hand 7.924E2 086 9.08E2 08 3.94E2 086 1.37E2 08 50.27 0.03Contrast Whole hand 6.5461E2 056 5.054E2 05 6.94E2 066 2.18E2 06 89.39 3.66E2 06Correlation Whole hand 0.00166 0.0014 0.00046 0.0003 75 0.0002Energy Whole hand 0.00146 0.001 0.000446 0.0003 71.42 0.0004Homogeneity Whole hand 0.00176 0.001 0.000466 0.0003 76.47 0.0002

RA: rheumatoid arthritis; SD: standard deviation; HDI: heat distribution index; TI: thermographic index; MCP: metacarpo-phalangeal; PIP: proximal

inter-phalangeal; DIP: distal inter-phalangeal.



significantly (p \ 0.05) with features such as mean,variance, skewness, energy, correlation, entropy andhomogeneity. As an overall comparison of all the MCPjoints, it was found that in MCP3 joint, the measuredmean average skin temperature was found to be highlycorrelated with the feature extracted parameters in thetotal population studied. Also, the mean skin surfacetemperature of PIP2, PIP3 and PIP5 correlated withmost of feature extracted parameters in the total popu-lation studied. But DIP joints had not shown any sig-nificant correlation between the mean average skinsurface temperature and the feature extractedparameters.

Table 2 shows the measured mean average skin tem-perature values at regions of DIP, PIP and MCP fromthe thermograph both in RA and normal groups. Itwas observed that the mean values of skin temperature,measured at all joints of the hand, were higher in theRA group, compared to its normal counterparts.

In MCP, PIP and DIP regions, the calculated percent-age difference in the mean values of skin temperatureswas found to be higher in MCP3, PIP3 and DIP3regions of interests in the RA group, as compared tothe normal group. These values were 5.3%[((35.522 33.64)/35.52)3 100], 4.9% [((35.442 33.71)/35.44)3 100] and 4.8% [((35.242 33.55)/35.24)3 100],respectively, and these values were statistically signifi-cant at p \ 0.001. Thus, in the RA group, the MCP3joint region showed highest percentage difference in themean values of skin temperature, as compared to nor-mal group. Figure 1(a) and (b) represents the dorsalview of the hand thermal image of normal subject andthe RA patient, respectively. In total, 14 ROIs (squareboxes) are shown with the minimum and maximumskin surface temperatures. These ROIs correspond toMCP (15), PIP (25) and DIP (15) joints of eachhand of a subject in the RA and normal groups.Similarly, Figure 1(c) and (d) represents the ventral

Figure 1. (a) Normal subjectdorsal view of the hand thermal image. The square ROIs indicate the boxes depicting the minimumtemperature ranges (30.6 C32.5 C) and maximum skin surface temperature ranges (31.5 C32.5 C) at the MCP, PIP and DIPjoints. (b) RA patientsdorsal view of the hand thermal image. The square ROIs indicate the boxes depicting the minimumtemperature ranges (31.9 C33.6 C) and maximum skin surface temperature ranges (33 C34.8 C) at the MCP, PIP and DIPjoints. (c) Normal subjectventral view of the measured thermal image. The square ROI indicates a box depicting the minimumtemperature (34.5 C) and maximum skin surface temperature (35.9 C) at palm region. (d) RA patientventral view of the handthermal image. The square ROI indicates a box depicting the minimum temperature (35.5 C) and maximum skin surfacetemperature (36.5 C) at palm region.



view of the hand thermal image of RA and normalgroups, respectively. The ROI indicated by the squarebox depicts the minimum and maximum skin surfacetemperature in the palm region (ventral side).

The thermal images of both hands in dorsal viewwere segmented for RA and normal subjects using thek-means algorithm (Figures 2 and 3). Figure 2 showsthe dorsal view of RA patient thermal image with thevarious stages of the segmented output image; specifi-cally Figure 2(a) shows the input image; Figure 2(b)

represents the cluster 1 image displaying the hot spotregions (absent for normal group); Figure 2(c) illus-trates the cluster 2 image depicting the background;Figure 2(d) displays the cluster 3 image showing theother regions; Figure 2(e) indicates the segmented hotspot regions (absent in case of normal) and Figure 2(f)represents the gray segmented output image. Figure 3represents the dorsal view of normal thermal image,where Figure 3(a) shows the input image; Figure 3(b)represents the cluster 1 image displaying the hot spot

Figure 2. RA patientdorsal view of various stages of the segmented output hand thermal image: (a) input hand thermal image, (b)cluster 1 image depicts presence of hot spot region, (c) cluster 2 image depicts the background, (d) cluster 3 image depicts the otherregions, (e) segmented output image depicts presence of hot spot region and (f) segmented gray output image with hot spot region.



regions; Figure 3(c) illustrates the cluster 2 imagesdepicting the background; Figure 3(d) displays the clus-ter 3 images showing the other regions; Figure 3(e) indi-cates the segmented hot spot regions and Figure 3(f)represents the gray segmented output image. Similarly,the ventral view of thermal images (Figures 4 and 5)demonstrates the segmented hot spot region if any pres-ent in RA and normal groups.

In the RA group (n=30), the feature extractedparameters were at significantly higher values

compared to normal group (p \ 0.05) as mentioned inTable 2. In comparison of RA and normal, the featureextraction parameters such as mean, skewness, entropy,kurtosis, contrast, correlation, energy and homogeneityshow high significance (p \ 0.01), but moderate signif-icance (p \ 0.05), achieved in features such as smooth-ness and variance.

A computer-aided diagnostic tool using MATLABcoding was developed for automated image segmenta-tion of hot spot regions, feature extraction and

Figure 3. Normal subjectdorsal view of the various stages of segmented output hand thermal image: (a) input hand thermalimage, (b) cluster 1 image depicts absence of hot spot region, (c) cluster 2 image represents the background, (d) the cluster 3 imagedepicts the other region, (e) segmented output image depicts the absence of hot spot region and (f) segmented gray output imagewithout any hot spot region.



classification of subjects into RA group and normalgroup based on the segmented hot spot region of thehand thermal image. Figure 6(a) and (b) showscomputer-aided diagnostic tool of thermal image analy-sis of hand-segmented image of RA patient and normalsubject, respectively.

Discussion

In this study, the mean average skin surface tempera-ture was measured at DIP, PIP and MCP of digits 15

which were greater for the RA group compared to thenormal group. Frize et al.31 in their thermal infraredimaging study predicted that second MCP, third MCPand the knee joints showed the greatest statistical dif-ference between the control subjects and the patients.In our study, among all the joints, the third MCP, thirdPIP and third DIP show statistically significant differ-ence in temperature between RA group and normalgroup. Hence, it was observed that in the total popula-tion studied, middle fingers have been affected withRA severely compared to other fingers in hand region.

Figure 4. RA patientventral view of various stages of the segmented output hand thermal image: (a) input hand thermal image, (b)cluster 1 image depicts presence of hot spot region, (c) cluster 2 image depicts the background, (d) cluster 3 image depicts the otherregions, (e) segmented output image depicts presence of hot spot region and (f) segmented gray output image with hot spot region.



The difference in temperature between RA and normalgroup in the palm region (ventral side) is less comparedto the dorsal side of the hand thermal image. Also, inthis study, from the dorsal view of the hand thermalimage, it was observed that hot spot region was foundat both wrist and joint regions. These findings had an

agreement with the result of thermal imaging study inRA conducted by Spading et al.25

Borojevic et al.32 found in their study that the RApatients had a mean surface temperature on the ventralside of the hand varying from 24.8 C to 36.5 C withthat of the healthy controls ranging from 24.8 C to

Figure 5. Normal subjectventral view of the various stages of segmented output hand thermal image: (a) input hand thermalimage, (b) cluster 1 image depicts absence of hot spot region, (c) cluster 2 image represents the background, (d) cluster 3 imagedepicts the other region, (e) segmented output image depicts the absence of hot spot region and (f) segmented gray output imagewithout any hot spot region.



35.5 C. In our study, RA patients had the mean aver-age surface temperature in the ventral side of the handas 35.39 C compared with that of the healthy controlsas 34.43 C. The percentage temperature differenceobtained between RA and normal groups was higher(2.71%) in our study compared to the study conductedby Borojevic et al. (1.63%). Hence, the reason forincreased temperature in the hands of RA was due tothe increased vascularity caused by synovitis and theproliferation of synovial cells. Also, the blood flow wasprevented from deep circulation to subcutaneous

circulation which causes an increase in temperature atthe surface of joints.

Collins et al. calculated the TI for quantitative anal-ysis of inflammation using thermography in arthritisusing multi-isothermal analysis. They had obtained themean TI of 3.77 in the rheumatoid knees and comparedthe isothermal pattern of inflamed knee with thederived TI before and after treatment.19 In our study,the mean TI attained was 4.37 in the rheumatoidhands, which was higher by 35.6% than the normalgroup. Also, there was a statistically significant

Figure 6. A computer-aided diagnostic tool for thermal image analysis of hand-segmented image: (a) sample RA patient and (b)sample normal subject.



(p \ 0.01) positive correlation observed between TIand feature extracted parameters in the total popula-tion studied. Salisbury et al.20 suggested the HDI as avalid temperature parameter for interpreting the RAinformation to detect abnormalities in the knee, wrist,elbow and ankle joints. Ilowite et al.33 validated andconfirmed that thermographic HDI is one of the bestmethods to assess inflammation of the synovial mem-brane for RA diagnosis. Spading et al. evaluated theRA patients with standardized thermographic methodsand calculated the HDI as twice the SD of all the tem-peratures in the pre-defined ROI. In addition to that,they suggested that thermal imaging could be used toimprove the assessment of disease activity in arthritisand could quantify clinically reasonable changes inarthritic joints in response to therapy.25 In our study,we calculated the HDI in the hand thermal image ofRA group and normal group as suggested by Spadinget al.; the results revealed that the percentage differenceof HDI was found to be higher by 51.8% in the RAgroup compared to normal group. To the best of ourknowledge, the correlation studies between temperatureindices and feature extracted parameters were limited.But in our study, the correlation between HDI and fea-ture extracted parameters was found to be positive andstatistically significant (p \ 0.05) in the total studiedpopulation.

Zhou et al. investigated the utilization of level setmethods to extract the boundary on thermal IRimages. An initial contour is grown adaptively using aspeed function based on edge/direction map. They alsoindicated that level set active contour method showedpromising results but are still unable to recover fullcontours for weak edges and poor contrast images.34

Frize et al. used image processing techniques such ashistogram and thresholding for analysis of hand ther-mal images in RA. But thresholding techniques do notperform well and often missed out important regions ofthe object of interest or segment larger regions than theexpected regions where fine details are lost. They alsoreported that features such as mode/max, variance andmaxmin calculations best discriminate between thetemperature measurement of control and patientgroups.35 In our study, k-means algorithm with threeclusters was used for effective segmentation of hot spotregions of hand thermal images in RA group. Also, 10statistical features were extracted from a segmentedoutput image to distinguish between RA group andnormal group. These features depict higher values forRA patients compared to normal group and had anagreement with findings of Frize et al.

The uniqueness of this study includes a computer-aided diagnostic tool using MATLAB coding devel-oped for automated image segmentation of hot spotregions, feature extraction and classification of subjectsinto RA group and normal group based on the segmen-ted hot spot region of the hand thermal image. Thepotential limitation of this study includes requirement

to validate the results for the largest number of samplesize.

Conclusion

In conclusion, in the population studied, the measuredmean average skin temperature of MCP3 joint highlycorrelated with most of the extracted features of thehand. Thermal imaging parameters such as skin tem-perature measurements, TI and HDI correlated signifi-cantly with feature extracted parameters in thepopulation studied. Hence, the developed computer-aided diagnostic tool using MATLAB could be used asa reliable method in diagnosing and analyzing thearthritis in hand thermal images.

Declaration of conflicting interests

The authors declare that there is no conflict of interest.

Disclosure

This article has been read and approved by all theauthors. This article is not under consideration by anyother publications and has not been publishedelsewhere.

Funding

This research received no specific grant from any fund-ing agency in the public, commercial or not-for-profitsectors.

References

1. Silman AJ and Hochberg MC. Epidemiology of the rheu-matic diseases. 2nd ed. New York: Oxford University

Press, 2001.2. Milind P and Sushila K. How to live with rheumatoid

arthritis? Int Res J Pharm 2012; 3(3): 115121.3. Pincus T. Long term outcomes in rheumatoid arthritis.

Br J Rheumatol 1995; 34(Suppl. 2): 5973.4. Sokko T. Work disability in early rheumatoid arthritis.

Clin Exp Rheumatol 2003; 5(Suppl. 31): 7174.5. Symmons DP and Gabriel SE. Epidemiology of CVD in

rheumatic disease, with a focus on RA and SLE. NatRev Rheumatol 2011; 7(7): 399408.

6. Devereaux MD, Parr GR, Lachmann SM, et al. Thermo-

graphic diagnosis in athletes with patellofemoral arthral-gia. J Bone Joint Surg Br 1986; 68(1): 4244.

7. Maeda M, Kachi H, Ichihashi N, et al. The effect of elec-

trical acupuncture stimulation therapy using thermogra-phy and plasma endothelin levels in patients withprogressive systemic sclerosis. J Dermatol Sci 1998;

17(2): 151155.8. Devereaux MD, Parr GR, Page Thomas DP, et al. Dis-

ease activity indexes in rheumatoid arthritis: a prospec-tive comparative study with thermography. Ann Rheu

Dis 1985; 44: 434437.9. Viitanen SM and Laaksonen AL. Thermography in juve-

nile rheumatoid arthritis. Scand J Rheumatol 1987; 16(1):

9198.



10. Varju G, Pieper CF, Renner JB, et al. Assessment of handosteoarthritis: correlation between thermographic andradiographic methods. Rheumatology 2004; 43(7): 915919.

11. Warashina H, Hasegawa Y, Tsuchiya H, et al. Clinicalradiographic and thermographic assessment of osteoar-thritis in the knee joints.Ann Rheum Dis 2002; 61(9): 852854.

12. Jiang LJ, Ng EY, Yeo AC, et al. A perspective on medi-cal infrared imaging. J Med Eng and Technol 2005; 29(6):257267.

13. Glehr M, Stibior A, Sadoghi P, et al. Thermal imaging asa non-invasive diagnostic tool for anterior knee pain fol-lowing implantation of artificial knee joints. Int J Therm2011; 14(2): 7178.

14. Trafarski A, Rozanski L, Straburzynska-lupa A, et al.The quality of diagnosis by IR thermography as a func-tion of thermal stimulation in choosing medical applica-

tions. In: Proceedings of the 9th international conferenceon quantitative infrared thermography, Krakow - Poland,25 July 2008. Available at: qirt.gel.ulaval.ca/archives/qirt2008/papers/03_13_17.pdf

15. Ring EF, Collins AJ, Bacon PA, et al. Quantitation ofthermography in arthritis using multi-isothermal analysis.II. Effect of nonsteroidal anti-inflammatory therapy on thethermographic index. Ann Rheum Dis 1974; 33(4): 353356.

16. Boas NF. Thermography in rheumatoid arthritis. AnnNew York Acad Sci 1964; 121(1): 223234.

17. Vardasca R. Hand thermogram standardisation withbarycentric warp model. In: Proceedings of the 4thresearch student workshop, University of Glamorgan,South Wales, March 2009, pp.7375.

18. Vardasca R, Ring EFJ, Plassmann P, et al. Thermal sym-metry of the upper and lower extremities in healthy sub-jects. Thermol Int 2012; 22(2): 5360.

19. Collins AJ, Ring EF, Cosh JA, et al. Quantitation of ther-mography in arthritis using multi-isothermal analysis. I.The thermographic index. Ann Rheum Dis 1974; 33(2):113115.

20. Salisbury RS, Parr G, De Silva M, et al. Heat distribu-tion over normal and abnormal joints: thermal patternand quantification. Ann Rheum Dis 1983; 42(5): 494499.

21. Bajwa U, Vardasca R, Ring EFJ, et al. Comparison ofboundary detection techniques to improve image analysisin medical thermography. Imag Sci J 2010; 58(1): 1219.

22. Selvarasu N, Nachiappan A and Nandhitha NM. Eucli-dean distance based color image segmentation ofabnormality detection from pseudo color thermographs.Int J Comp Theory Eng 2010; 2(4): 514516.

23. Mishra R, Sharma BL, Gupta R, et al. Indian Rheuma-

tology Association consensus statement on the manage-

ment of adults with rheumatoid arthritis. Indian J

Rheumatol 2008; 93(3): S1S16.24. http://www.iact-org.org (accessed on December 2014).25. Spading SJ, Kwoh CK, Boudreau R, et al. Three dimen-

sional and thermal surface imaging produces reliable

measures of joint shape and temperature: a potential tool

for quantifying arthritis. Arthritis Res Ther 2008; 10(1):

R10.26. Rovensky J, Clague R and Payer J. Thermographic

index. In: Dictionary of rheumatology. Vienna: Springer-

verlag, 2009, p. 212. DOI: 10.1007/978-3-211-79280-3_1104.27. Kanungo T and Mount DM, Netanyahu NS, et al. An

efficient k means clustering algorithm: analysis and imple-

mentation. IEEE Trans Pattern Anal Mach Intell 2002;

24(7): 881892.28. Haralick RM, Shanmugam K and Dinstein I. Texture

features for image classification. IEEE Trans Syst Man

Cybern Syst 1973; 3(6): 610621.29. Salvatore D and Reagle D. Schaums outline series of the-

ory and problem of statistics and econometrics. New York:

McGraw-Hill, 2002.30. Girisha AB, Chandrashekhar MC and Kurian MZ. Tex-

ture feature extraction of video frames using GLCM. Int

J Eng Trends Tech 2013; 4(6): 27182721.31. Frize M, Karsh J, Herry C, et al. Preliminary results of

severity of illness measures of rheumatoid arthritis using

infrared imaging. In: Proceedings of the international

workshop on medical measurements and applications,

Cetraro, 2930 May 2009, pp. 187192. New York:

IEEE.32. Borojevic N, Kolaric D, Grazio S, et al. Thermography

hand temperature distribution in rheumatoid arthritis

and osteoarthritis. Period Biol 2011; 113(4): 445448.33. Ilowite NT, Walco GA and Pochaczevsky R. Assessment

of pain in patients with juvenile rheumatoid arthritis:

relation between pain intensity and degree of joint inflam-

mation. Ann Rheum Dis 1992; 51: 343346.34. Zhou Q, Li Z and Aggarwal JK. Boundary extraction in

thermal images by edge map. In: Proceedings of the ACM

symposium on applied computing, Nicosia, 1417 March

2004, pp.254258. New York: ACM.35. Frize M, Adea C, Payeur P, et al. Detection of rheuma-

toid arthritis using infrared imaging. In: Proceedings of

the SPIE: medical imaging, vol. 79620M, 10 March 2011.

Bellingham, WA: SPIE.



Snekhalatha-Anbu, Venkatraman_ Journal of Engineering in Medicine-2015

Documents

Transcript of Snekhalatha-Anbu, Venkatraman_ Journal of Engineering in Medicine-2015