IR video processing and thermal energy quantification
Based on BAT thermogenesis causing in an increase in skin temperature, a number of previous studies used a threshold segmentation technique to detect regions of increased IR signal, with the threshold calculated as the median value of the highest 25 % of pixels [22, 24]. However, this approach neglects spatial information in the image and cannot generate accurate data in terms of the detection results for analysis of BAT activity. In our study we modified the seeded region growing (SRG) technique [25] and applied it to IR video imaging to detect local regions of interest (ROIs), a term we use to refer to the ‘hot’ regions overlaying the C-SCV BAT depots that are delineated by our algorithm.
The standard SRG algorithm starts with the selection of a suitable seed and further expands it into spatially connected regions of similar characteristics. In our application, the seed is automatically determined by searching for the pixel of the highest temperature in the pre-defined image areas (e.g., the left or right C-SCV region). To increase the robustness of the SRG algorithm, we define a small local region \(A_{I}\) containing the seed pixel to ensure that a stable mean temperature can be estimated for the region. Let H be the set of all unallocated pixels that are adjacent to region A
I
:
$$H = \left\{ {\varvec{x} \notin A_{i} |N\left( \varvec{x} \right) \cap A_{i} \ne \emptyset } \right\},$$
(1)
where \(N\left( \varvec{x} \right)\) is the 8-neighbors of the pixel \(\varvec{x}\). For a pixel \(\varvec{x} \in H\), the SRG algorithm computes the difference \(\delta \left( \varvec{x} \right)\) between its temperature value and the mean temperature of the region \(A_{i }\) as a measure of how different pixel \(\varvec{x}\) is from the region it adjoins:
$$\delta \left( \varvec{x} \right) = \left| {T\left( \varvec{x} \right) - \mathop {\text{mean}}\limits_{{\varvec{y} \in A_{i} }} \left[ {T\left( \varvec{y} \right)} \right]} \right|,$$
(2)
where \(T\left( \varvec{x} \right)\) is the temperature reading of pixel \(\varvec{x}\). We then take \(\varvec{z} \in H\) such that
$$\delta \left( \varvec{z} \right) = \mathop {\hbox{min} }\limits_{{\varvec{x} \in H}} \left[ {\delta \left( \varvec{x} \right)} \right]\,\quad{\text{ and}}\quad\delta \left( \varvec{z} \right) \le T_{t} ,$$
(3)
and add pixel \(\varvec{z }\) into the region A
i
. This iterative process is repeated until \(\delta \left( \varvec{z} \right)\) becomes larger than a pre-defined threshold \(T_{t}\). Different from the standard application of SRG which involves assigning all unallocated pixels to the most homogenous regions, our application requires the detection of local ROIs which contain spatially connected pixels of similar temperature values. Therefore, we set the parameter \(T_{t }\) to control how tolerant the SRG algorithm will accept a neighboring pixel to grow the region. In our study, \(T_{t }\) was optimized for individual subjects to achieve reliable segmentation. In Fig. 1, we show the results for ROI detection and compare these with results based on the threshold segmentation technique. We demonstrate that our method can accurately detect local pixel clusters with closed boundaries overlaying the BAT depot in the C-SCV region. In comparison, the threshold segmentation technique can only identify a number of disconnected pixel clusters, a limitation which reduces its reproducibility and accuracy.
We applied this algorithm to 5 s of IR video sequence per data point, with a total of 150 frames per video sequence. Only 5 s of IR video sequence per data point was used in this pilot study since processing IR video sequences was a time-consuming process, with longer video sequences requiring longer processing times. Our algorithm calculated the number of pixels and the mean temperature value of pixels in the ROIs detected in each frame and then averaged these values across the entire 5 s (150 frames) of video sequence. The pixel count of the ROI was used to estimate the actual area of the ‘hot region’ overlaying the C-SCV BAT on the human body based on a simple calibration. Two 5-mm aluminum foil disks placed on the skin—one under each clavicle symmetrically spaced 18 cm apart and two other similar disks each positioned 8 cm above the respective subclavicular disks on the superior border of the trapezius muscle bilaterally—served as fiducial markers delineating a rectangle on a three-dimensional plane. A 3×3 matrix could then be computed to represent transformation between two spatial frames of reference for the registration of images against known linear dimensions and enable subsequent computation of the ROI area via pixel counts. As the aluminum foil disks had a distinct temperature of ~28–31 °C, compared to skin temperatures image registration with these fiducial markers also permitted precise detection of changes in the actual temperature to be distinguished from those caused by changes in body poses.
We next quantified BAT heat energy output in watts (W) by applying the Stefan–Boltzmann law:
$${\text{BAT heat energy output }}\left( {\text{W}} \right) \, = \varepsilon \times \sigma \times \, A \, \times \, T^{4} ,$$
(4)
ε: emissivity (0.98 for human skin);
σ: Stefan–Boltzmann’s constant (5.676 × 10−8 W/m2K4);
A: mean ROI area in m2 (based on number of pixels in detected ROI);
T: mean ROI temperature in Kelvin (averaged temperature values of all pixels in detected ROI).
The heat energy output of the left and right C-SCV BAT were summed to give the total heat energy emitted by the C-SCV BAT depots.
IRT of C-SCV regions
Subjects were seated in a relaxed and upright posture with arms adducted, away from all heat-emitting objects and at a 1.0-m distance from an IR thermal imaging camera fastened onto a tripod (model FLIR T440; FLIR Systems AB, Täby, Sweden) with thermal resolution at 76,800 (320 × 240) pixels. This positioning was to assure optimum visualization of the C-SCV regions. Thermal imaging was performed in the same room as the indirect calorimetric measurements, at a constant ambient temperature of 24 °C.
All IR thermal video recordings were performed over a standard recording period of 5 min. During the 5-min recording period, subjects were asked on three separate occasions to turn their heads to the right for 10 s while keeping shoulders still in the antero-posterior orientation, then instructed to turn their heads to their left for another 10 s in a similar fashion before turning their heads back to face the camera. This sequence of changes in head position was to allow for imaging of the front and both antero-lateral views of the C-SCV regions.
On the first visit, thermal imaging of the C-SCV regions was carried out while the subjects were exposed to a cold stimulus for 5 min. The cold stimulation consisted of the immersion of both hands and feet in water kept at 18 °C, a local cooling protocol adapted from Symonds et al. [22].
Capsinoid ingestion trials were conducted on each of the two subsequent test visits. Thermal data from the C-SCV regions were recorded continuously for 5 min before each subject ingested the test meal to obtain the baseline temperature. IRT (5 min per recording) was repeated at 30-min intervals over a 2.5-h period following ingestion of capsinoids or placebo. All thermal data were recorded in a radiometric IR video format and analyzed using the FLIR research IR software [26]. The IR videos were recorded at 30 frames per second.
Subjects
Between October 2013 and June 2014, 24 healthy, lean [body mass index (BMI) of <25 kg/m2] male volunteers were recruited (Table 1). Potential subjects underwent a screening session consisting of a questionnaire and measurement of BMI and fasting blood glucose. Exclusion criteria included pre-diabetes or diabetes, metabolic disorders, ingestion of medications, smoking, and recent weight change. This trial was registered at http://www.clinicaltrials.gov (ID: NCT01961674; registration date: 10/10/2013).
Table 1 Baseline characteristics of subjects
Test substances
Capsinoids and placebo capsules were provided by Ajinomoto Inc. (Tokyo, Japan). The capsinoids were extracted from the fruit of sweet chili pepper Capsicum anuum L. var. CH-19 plants, purified, and encapsulated as previously described [27]. Each gel capsule contained either placebo or 1.5 mg of capsinoids; the latter capsules contained capsiate, dihydrocapsiate, and nordihydrocapsiate in a 7:2:1 ratio, and 199 mg of a mixture of rapeseed oil and medium-chain triglycerides.
Study protocol
This was a double-blinded, placebo-controlled, randomized crossover study in which each subject completed two study visits and consumed (in random order) either 0 or 9 mg of capsinoids in capsule form. All capsules looked identical and contained either 0 (placebo) or 1.5 mg of capsinoids. Six capsules (1.5 mg each) were consumed during each test visit, together with a standardized portion of white rice containing 50 g carbohydrates. The ingestion of rice enabled a comparison of the thermic effect of food with diet-induced thermogenesis triggered by capsinoid-stimulated BAT as well as an assessment of whether BAT activation had any effect on glycemic response. Randomization was performed using a computer-generated random number list prepared by one of the authors with no clinical involvement in the study, such that both subjects and investigators involved in the study were blinded to the capsule assignment for each study visit.
On each test visit, subjects reported to the laboratory between 0800 and 0900 hours after an overnight fast of at least 10 h. Fasting blood glucose was measured at each visit using a glucose dehydrogenase method (HemoCue AB, Angelholm, Sweden). The mean fasting blood glucose level was determined by averaging the fasting blood glucose values obtained on both test visits. Subjects were asked to avoid strenuous physical activity, alcohol consumption, and consumption of chili or spicy foods on the day prior to the test visits.
Upon arriving at the laboratory, subjects were provided standard cotton singlets (to ensure adequate exposure of the neck and upper thorax for thermal imaging) and Bermuda shorts with an estimated Clo value of 0.2 [28]. The Clo unit provides a measure of thermal insulation provided by clothing [29]. Room temperature was maintained at 24 °C so that subjects were in thermal comfort conditions to avoid shivering that may confound the results. With the subjects resting in a supine position, oxygen consumption and carbon dioxide production were continuously recorded via indirect calorimetry for 45 min. The stable value of the last 10-min period was used to calculate the respiratory quotient (RQ) and resting energy expenditure (EE), the latter considered to be a good approximation of the basal metabolic rate (BMR). After 45 min, IR imaging of the C-SCV regions was performed by video recording for 5 min. The subjects then ingested capsules containing the placebo or 9 mg of capsinoids, with water ad libitum and a portion of white rice. During the 2.5 h following the consumption of the capsules and food, respiratory gas parameters were recorded continuously for 20 min of each of five 30-min time intervals, and the EE and RQ were recorded during the last 10-min period of each 30-min time interval. At 30-min intervals, IR imaging of the C-SCV regions was performed over a standard recording period of 5 min (see Fig. 2).
Indirect calorimetry
Indirect calorimetry (Quark RMR; COSMED) was used to provide a sensitive imaging-independent proxy of BAT activation. Changes in EE and RQ were measured since UCP-1-dependent heat production should elevate the metabolic rate and thus increase EE. Oxygen consumption and carbon dioxide production were measured using a ventilated hood system, with the subjects resting in a supine position. EE and RQ were estimated from measures of oxygen consumption and carbon dioxide production based on the Weir equation [30].
After an overnight fast, resting EE was measured for 45 min prior to consumption of the test meal (6 capsules + standardized portion of white rice). Following the test meal, postprandial EE was measured in 20-min blocks over the next 2.5 h.
Statistical analysis
Paired t tests were used to ascertain if there were differences in SCV heat production under cold stimulation conditions between the low-BAT (placebo) and high-BAT (capsinoids) groups. Data were expressed as mean ± standard error of the mean (SEM). A multi-level model constructed with the generalized structural equation model (gSEM) framework [31] was used to determine if there were significant differences in EE, fat oxidation, and SCV heat production between the capsinoids and placebo groups over time (0.5, 1, 1.5, 2, 2.5 h) in low-BAT, high-BAT, and all subjects, while adjusting for the respective baseline values. The coefficients were interpreted as the average difference in these outcomes between the capsinoids and placebo (reference) groups. The standard errors of the coefficients were adjusted with a robust procedure. As an advanced regression model, gSEM is ideal for handling longitudinal data where the outcomes are monitored over time as it captures all salient features (treatment effects, time effects, baseline effects), thus better reflecting the nature of our data [32]. Notably, gSEM has a higher statistical power and is superior to random-effects, repeated-measures analysis of variance since it captures more features and requires fewer assumptions, which therefore optimizes its capability of detecting a statistically significant difference even with a small sample size. All analyses were performed with Stata/MP version 14 (Stata Corp, College Station, TX), and the significance level for all statistical tests was set at 5 %.