Open Access
Add to BibSonomy
Abstract
Analysis of morphological changes of the peritoneal membrane is an essential part of animal studies when investigating molecular mechanisms involved in the development of peritoneal fibrosis or testing the effects of potential therapeutic agents. Current methods, such as histology and immunohistochemistry, require time consuming sample processing and analysis and result in limited spatial information. In this paper we present a new method to evaluate structural and chemical changes in an animal model of peritoneal fibrosis that is based on hyperspectral imaging and a model of light transport. The method is able to distinguish between healthy and diseased subjects based on morphological as well as physiological parameters such as blood and scattering parameters. Furthermore, it enables evaluation of changes, such as degree of inflammation and fibrosis, that are closely related to histological findings.
© 2020 Optical Society of America under the terms of the OSA Open Access Publishing Agreement
1. Introduction
Peritoneal fibrosis is a serious clinical complication found in patients treated with peritoneal dialysis (PD). Patients receiving PD for extended periods of time can develop functional and morphological changes of the peritoneal membrane. These changes are characterized by injury to mesothelial cells and prolonged inflammation, which leads to development of fibrosis [1,2] or even mortality in up to 10% of patients receiving PD [3]. Since PD as a treatment option has many advantages in clinical practice, strategies to prevent peritoneal fibrosis are of great clinical interest.
To investigate mechanisms involved in the initiation and development of peritoneal fibrosis, or to test the effects of potential treatment strategies for humans, various rodent models have been developed, including chlorhexidine gluconate (CHX) induced mouse model [2–5]. The CHX induced mouse model is a frequently used model of peritoneal fibrosis with morphological similarities to those seen in humans such as inflammation, neoangiogenesis, fibroblast proliferation, production of collagen fibers which finally leads to fibrosis, thickening of the peritoneum, and peritoneal dysfunction [3,4]. In the first stages of the disease model, inflammation is the dominant process characterized by infiltration of various inflammatory cells, interstitial fluid, erytrocytes (blood congestion), and injured tissue cells [3].
To evaluate the effectiveness of treatment on different stages of peritoneal fibrosis in animals, various techniques can be used. Histological evaluation of abdominal biopsy is the gold standard. It offers unparalleled insight into changes on the tissue level but is spatially limited to the small sample and thus can miss localized changes outside of the sample. Additionally, the techniques rely on time-consuming sample processing and analysis by experienced personnel.
Hyperspectral imaging (HSI) is an emerging technique that combines spectroscopy and imaging with a broad spectrum of applications from food science [6], art analysis [7], remote sensing [8], medical sciences [9] and many others. The result of HSI is a 3D dataset comprised of two spatial and one spectral dimensions. Tissue native chromophore distribution and scattering properties can be calculated by fitting HSI spectroscopic data to light transport models, offering an insight into both physiology and morphology. A distinct advantage of HSI is that it is contactless, does not damage samples, and can provide results almost in-real-time. Since the samples are not damaged by visible and near infrared light used for imaging, they can be further used normally, even after being imaged using HSI. In the scope of this study, it was thus possible to perform HSI and histological examination of same samples, enabling the comparison of the two. In contrast to histological examination, where samples have to be stained, HSI harnesses tissue native chromophores and can offer different contrasts without staining.
HSI was used to image whole abdominal walls of healthy subjects and subjects with chemically induced peritoneal fibrosis. Physiological and morphological changes of the abdominal wall among subjects were evaluated using statistical metrices of optical parameters obtained from the HSI data. Based on these metrices, subjects were successfully classified as diseased or healthy. HSI results were also compared with histology slides taken from the same samples to verify the results. Since HSI datacube is large (size of a single image can be more than 5 GB), a fast algorithm for data processing was needed. MonteCarlo multi layer (MCML) [10], although providing precise results, was not fast enough, even in the Nvidia CUDA GPGPU enhanced version [11]. To achieve fast processing of images that can provide results almost instantaneously, a modified version of Kubelka-Munk two flux model was used, with application specific modifications. With this model, fitting of a single point took less than a second, so the processing time for a whole image took approximately one hour.
2. Methods
2.1 Hyperspectral imaging
Hyperspectral images were acquired using a custom built push-broom HSI working in a reflectance geometry with a broadband LED light source covering the spectral range from $400~\mathrm {nm}-1000~\mathrm {nm}$[12,13]. The light source was configured to illuminate the whole sample homogeneously from two sides while using a combined diffusor/polarizer (Bolder Vision Optik, Inc., USA) in front of the LEDs. The imaging part of the system included an ImSpector V10E imaging spectrograph (Specim, Spectral Imaging Ltd, Finland), a 17 mm lens (Schneider Kreuznach Xenoplan 1.4/17-0903, Jos. Schneider Optische Werke GmbH, Germany) and 5.0MP monochrome CMOS camera (Blackfly S BFS-U3-51S5M-C, Flir Systems Inc., USA). The working distance of the objective was 29 cm. To mitigate specular reflections, a polarizer (Bolder Vision Optik, Inc., USA) was used in front of the objective in a cross-polarized configuration with the LED polarizers. Images were acquired with the resolution of $2048 \times 2448$ pixels in spectral and spatial dimensions of the spectrograph, respectively. The raw spectral resolution was approximately $0.5~\mathrm {nm}$ and one pixel on the detector corresponded to $0.06~\mathrm {mm}$ in the object plane. The field of view in the direction perpendicular to the scanning axis was $144~\mathrm {mm}$. The effective spectral and spatial resolutions of the system were $1~\mathrm {nm}$ and $250~\mathrm {\mu m}$, respectively, as evaluated by a gas discharge tube and spatial grids used for system calibration. To measure the 3D shape of examined objects, a laser 3D profilometry system was used. It consists of a laser line projector that projected a line on the sample at an oblique angle and a camera to detect intersection line deformation due to the surface shape. It reconstructed height maps from these deformations using the laser line triangulation technique [12,14–16]. The whole setup is represented schematically in Fig. 1. Data acquisition and system control was implemented in MATLAB (2016b, 2018b, MathWorks, USA).
Fig. 1. Hyperspectral imaging system: schematic representation (a) and a photograph of the system on site (b).
The work presented herein is at the stage of development and exploration, corresponding to IDEAL stages 2a and 2b [17].
2.2 Study protocol
The animal study was performed on 10 C57BL/6J male mice at 19 weeks of their age (Medical Experimental Centre, Ljubljana, Slovenia). The experiment was conducted in accordance with the national and EU legislation (Directive 2010/63 EU) and was approved by National Ethic Committee of Republic of Slovenia and the Administration of the Republic of Slovenia for Food Safety, Veterinary and Plant Protection (Licence No. 34401-13/2018/5). Animals were housed in open cages (Ehret, $825~\mathrm {cm}^2$ floor area) with autoclaved bedding (Lignocel 3/4, Germany) and enrichment (Mouse house, Tecniplast; paper towels; tunels), maintanance diet (Altromin 1324, Germany) and autoclaved water (bottles, Tecniplast) ad libitum. Animals were maintaned in a barrier system at $22-25~^{\circ }\mathrm {C}$, $55\pm 10 \%$ humidity, $12~\mathrm {h}$ light/dark cycle (light 7a.m.-7p.m.); SPF according to FELASA recommendations (QM diagnostics, Netherlands).
Peritonitis was induced in the pure model group by chlorhexidine gluconate (CHX) solution ($0.1\%$ CHX (Wako Pure Chemical Industries) in $15\%$ ethanol in phosphate buffered saline - PBS) injected intraperitonealy ($200~\mathrm {\mu L}$ per mouse) every second day for 7 days. Control group received PBS injected intraperitonealy every second day ($200~\mu \mathrm {L}$ per mouse) instead of CHX. During the disease induction in the treated model group, resolvin D1 (an omega 3-fatty acid metabolite dissolved in absolute ethanol ($100~\mathrm {ng/\mu L}$), Cayman Chemical, USA) was injected intraperitonealy ($100~\mathrm {ng}$ dissolved in PBS, injected as $100~\mathrm {\mu L}$ per mouse) every day for 7 days as a treatment strategy. Overview of the three groups is presented in Table 1. At the end of the study, mice were euthanized with $CO_2$. Immediately after the euthanasia, the fur and skin were removed from abdominal wall, and one half of the abdominal wall was excised, placed on a black rigid closed-cell foamed PVC (FOREX Classic, Airex AG, Switzerland) and imaged using HSI. Afterwards, samples were fixed in a standard $4\%$ buffered formaldehyde solution for histology. To avoid potential bias, the study was performed as a double blind study.
Table 1. Overview of mice groups, their roles in the experiment and protocol for each group. Whole cohort consisted of $n=10$ subjects. Column marked with # lists number of subjects in each group. i/p - intraperitoneally
2.3 Optical properties evaluated
To describe scattering and absorption in the sample and relate them to physiological properties, models based on the a-priori knowledge about the sample structure can be used to approximate absorption ($\mu _a$) and scattering ($\mu _s$) coefficients. The reduced scattering coefficient $\mu _s'=(1-g)\mu _s$, where $g$ is anisotropy of scattering, is approximated using a well known power law [18] in the form of
where $a$ and $b$ are parameters fitted by the model. Parameter $a$ is equal to the value of $\mu _s'$ at $500~\mathrm {nm}$, while scattering power $b$ describes wavelength dependence of the scattering coefficient and is related to the structure of the tissue.To model absorption, blood was included in the model as the most prominent chromophore. Absorption coefficient $\mu _a$ of the tissue can be written as a sum of products of pure constituent absorption coefficients and their volume fractions in the form [18]
where $BVF_D$ and $BVF_O$ are volume fractions of deoxygenated and oxygenated blood, respectively. From volume fractions of blood, total blood volume fraction $BVF=BVF_O+BVF_D$ was calculated. Data for blood absorption coefficient spectra was taken from Prahl [19] and is linearly related to molar extinction coefficient $\epsilon (\lambda )$ by $\mu _a(\lambda )=0.0054~\mathrm {M}~\epsilon (\lambda )$ [19]. In this paper, blood species volume fractions are used, as is common in the biomedical optics field. To convert between blood volume fractions and molar concentrations, relation $c_x = 0.0054~\mathrm {M}~BVF_x$ can be used.Besides blood, collagen is also present in the tissue. The absorption of the collagen is roughly exponential in shape and could cause crosstalk with scattering. Pure collagen absorption at $500~\mathrm {nm}$ is approximately $0.3~cm^{-1}$ [20], but its volume fraction is small (in rat muscle, dry content is approximately 6%) [21] and thus the absorption $\mu _a(500~nm)\approx 0.03~cm^{-1}$ is much lower than the contributions of scattering and blood absorption (at $5\%$ blood volume fraction). Thus effects of the collagen absorption were omitted from the model.
2.4 Two-flux semi-coupled model
To process the data, a model based on Kubelka-Munk and Beer-Lambert theory was developed. It processed the data for a single pixel in seconds, while accounting for non-ideal black substrate. The model is depicted schematically in Fig. 2.
Fig. 2. Quantities in the two flux semi-coupled model. Downward propagating flux $I_+$ and reverse propagating flux $I_-$ interact with the tissue as described by scattering parameter $S$ and absorption coefficient $\mu _a$. Axis $x$ is aligned with the direction of the downward propagating flux. Fluxes are determined by boundary conditions due to incoming light flux $I_0$ and substrate reflectance $R_S$.
Light propagating in the tissue is first written in terms of differential equations describing the attenuation of the incoming, downwards propagating flux $I_+$ and the reverse propagating flux $I_-$:
where $\mu _a$ is the absorption coefficient and $S$ is a parameter describing scattering. Please note, that in Eq. (4) the sign is inverted due to the $x$ axis direction. This model is simplified by assuming that the medium is not strongly scattering and that substrate reflectance is small ($R_s\ll 1$). This assumption insures that downwards propagating flux $I_+$ is much greater than the reverse flux $I_-$ ($I_+\gg I_-$) and that coupling $I_-\rightarrow I_+$ can be neglected.To solve the system of differential equations, boundary conditions are prescribed to the model. Flux $I_+$ at the top is equal to the incoming flux $I_+(0)=I_0$. Reflectance of the non ideally black substrate $R_s$ is considered in the form of boundary condition at the bottom $I_-(d)=I_+(d)\cdot R_s$.
Integrating the Eqs. (3) and (4), using the boundary conditions and expressing the model in terms of sample reflectance $R=I_-(0)/I_0$ the model equation used in this paper takes the form:
The scattering parameter $S$ is different from the scattering coefficient $\mu _s$, so a connection between them is needed to compare results with previously published values. It represent the loss of the photons from the flux due to the scattering. Since only backscattering events are considered as a scattering event for the two flux model, a simple rule-of-thumb relation can be defined as where $p_{BS}$ is defined as the backscattering coefficient obtained by integrating the phase function over backscattering angles. To approximate the tissue phase function, the Henyey-Greenstein phase function [22] has been used in this paper. For a single scattering value of $g=0.97$ found in literature [23], $p_{BS}=0.0064$ indicating that only a small fraction is backscattered.For simplicity and stability of the model, the number of free parameters in the model is further decreased by omitting refractive index and thus internal reflections within the tissue.
2.5 Data processing
Data processing was performed in sequential steps depicted in Fig. 3. The first step in the processing workflow was pre-processing, which prepared raw digital number data from the hyperspectral camera for further analysis. The data was corrected for the nonlinearity of the detector [13] and normalized to the reflectance values using white and dark references [12]. Since the white reference used for measurements was made from PTFE (Dastaflon d.o.o., Slovenia), a correction of the reference measured using a Spectralon(R) spectroscopic standard (Labsphere Inc., USA) was also performed. Profilometric data for the determination of sample thickness was aligned to the HSI cube using affine transformations in MATLAB. Finally, areas with insufficient signal to noise ratio (SNR) and specular reflections were removed as to not affect the automated analysis. They were identified by low reflectances and high reflectance with spectral artifacts respectively.
Fig. 3. Data processing workflow (a) and an example of a model fit (red line) over normalized experimental reflectance values (blue line) (b).
After the data was pre-processed, whole HSI images were fitted pixel by pixel to the model described in section 2.4. Fit stability was improved by fitting the negative logarithm of the reflectance $-\log R$. To approximate the wavelength dependence of the scattering parameter $S$, the power law (1) was used. Absorption of blood was described as a sum of oxygenated and deoxygenated blood following Eq. (2). Results of an example fit are presented in Fig. 3(b). At the end of the fitting step, maps of $a$ and $b$ as well volume fractions of both blood varieties for each corresponding pixels were returned along with their relative errors.
Fit output was post-processed to remove areas of unreliable fit that correspond to the areas with insufficient SNR not removed from analysis in pre-processing. They were identified by large relative error and extremal values of the parameters allowed in the fitting algorithm.
To classify subjects into groups, histograms of tissue properties over whole maps were calculated per subject per fitted parameter. To smooth the histograms, the data was fitted using the histfit function in MATLAB with the ’kernel’ parameter which fits a nonparametric kernel-smoothing distribution sampled in 100 points. These histograms were then used to calculate statistical metrices and distinguish between groups.
3. Results
3.1 Image statistics
First, maps of tissue properties were calculated using the model described in section 2.4. Figure 4 shows the output of the model fit to a healthy control group subject (S02). Areas with values close to the physical and physiological limits of parameters allowed during the fit were observed in parts of the sample. These areas also corresponded to larger relative errors, as can be seen in the lower row of graphs in brighter colors. An example of such an area is the thin central part of the sample that contained a large fraction of blood. These areas generally exhibited insufficient SNR and had to be removed from the analysis.
Fig. 4. Model parameters fit to a healthy subject (S02) in log10 scale over RGB projections of hyperspectral images with corresponding relative fit parameter errors in the bottom row (log10 scale).
Differences between healthy control subject in Fig. 4 and diseased subjects in Fig. 5 can be observed for both CHX and CHX+R group subjects. Values of $b$ are homogeneously larger in diseased subjects. Values of $a$ appear similar between groups, but spatial distribution is different; for diseased subjects areas of lower and higher scattering can be separated. Similarly, for diseased subjects blood vessels are more pronounced (visible as thin lines of increased $BVF_O$ and $BVF_D$). Between CHX and CHX+R group subjects no notable differences are observable from maps in Fig. 5.
Fig. 5. Model parameters fit for subjects in other experimental groups in log10 scale over RGB projections of hyperspectral images. Results for subject in CHX group (S06) are displayed in a), while results for a subject in CHX+R group (S10) are displayed in panel b).
To obtain an assessment of inflammation that can be compared to other techniques, histograms of fitted parameters were calculated for each subject to detect changes in their distributions between different groups. An example of histograms of scattering power $b$ are shown in Fig. 6. In the subjects of the healthy control group (PBS), the distributions had a sharply pronounced peak. In both model groups (CHX, CHX+R), the distribution peak was located at higher values of $b$ (corresponding to smaller scatterers) with shapes different than in the healthy subjects. Similar differences in distribution shapes were also observed for other parameters.
Fig. 6. Histograms of fitted scattering power $b$ over whole image for different study subjects in blue and a fitted kernel smoothing function shown as a red line. Subject codes and their groups are denoted in the histogram titles.
To better understand the differences in distribution shapes and provide a single numerical measure of tissue changes, statistical metrices of the histograms were calculated. They were chosen to detect the differences in peak position and distribution shape. Based on this criteria, selected metrices were full width at half maximum (FWHM) of the distribution, mode of the distribution (value with the highest frequency of occurrence), and distribution symmetry which was defined as
where $s$ is distribution symmetry, $m$ is the mode of the distribution, $n$ is the number of histogram bins, and $h_i$ is the value in i-th bin of the histogram. Values greater than one correspond to greater frequency of pixels with values smaller than mode and values less than one indicate the opposite.3.2 Classification
The ability of the method to separate the diseased subjects from the healthy ones was verified using the Mann-Whitney equivalent Wilcoxon test in MATLAB (ranksum). Statistical metrices were calculated for each subject and the significance of the metric differences between the groups were computed using the test. Results are presented in Table 2, showing acceptable significance in at least one distribution parameter for most physiological and optical properties. As a threshold for statistical significance, a p-value of $p<0.1$ was selected due to the low sample size and large variations in subject response that were also confirmed by the histological examination.
Table 2. Significance of tissue parameters for different subgroups performed using Wilcoxon rank sum test. p-values have been rounded to 2 decimal places. Compared subgroups are: PBS control group P, CHX pure model group C and CHX + resolvin D1 treated model group R. p-values lower than 0.1 are bolded in the table. The dataset size is $n=10$. $BVF$ - total blood volume fraction, $BVF_D$ - deoxygenated BVF, $BVF_O$ - oxygenated BVF, $a$ - scattering at $500~\mathrm {nm}$ and $b$ - scattering power.
Based on the statistical analysis, the metrices that could differentiate the diseased subjects from the healthy were identified as: mode of oxygenated blood volume fraction, FWHM of total blood volume fraction, symmetry of $a$ distribution and mode of $b$ as seen in Table 2.
Volume fraction of oxygenated blood had a high statistical significance ($p=0.02$) for the mode of the distribution, with markedly lower fractions for the control group subjects, as seen in Fig. 7. There was no obvious difference in fractions between the treated and the untreated disease model groups ($p=0.86$). Increased oxygenated blood fraction was related histologically to samples with more pronounced inflammation and necrosis. Please note, that due to the nature of the study protocol involving the euthanasia using $CO_2$, oxygenated blood fraction should not be understood directly in the way of active dynamic processes as it would be in live tissues.
Fig. 7. Mode of oxygenated blood volume fraction distribution (a) and corresponding boxplots for the three experimental groups (b).
The FWHM of the total blood volume fraction distribution exhibited larger values for the control group subjects ($p=0.03$), as can be seen in Fig. 8. This indicates a broader histogram and consequently more areas with fractions lower and higher than the mode of the distribution than in the model groups subjects. In these subjects, blood is thus more equally distributed over the whole sample. Differences between model groups were not statistically significant ($p=0.63$).
Fig. 8. FWHM of total blood volume fraction distribution (a) and corresponding boxplots for the three experimental groups (b).
Values for the scattering parameter $S$ were approximately $1.1~cm^{-1}$ at $500~nm$. For a rough estimate, given a value of $g(500~nm)\approx 0.97$ for a mouse muscle in the literature [23], Eq. (6) can be used to estimate the scattering coefficient value $\mu _s(500~nm)\approx 172~cm^{-1}$, which is in the range of $90~\mathrm {cm}^{-1} - 300~\mathrm {cm}^{-1}$ reported by other groups [23,24]. It is important to note that $g$ of our samples was not known and that the dispersion of existing measurements reported in literature is large due to the dependance on sample type and preparation.
Fig. 9. Symmetry $s$ of the scattering parameter at $500~\mathrm {nm}$ $a$ (a) and corresponding boxplots for the three experimental groups (b).
Fig. 10. Mode of the scattering power $b$ for subjects included in the study (a) and corresponding boxplots for the three experimental groups (b).
Due to the structural changes accompanying the inflammation, such as edema, fibrosis, and inflammatory cells infiltration, a difference in $a$ between the control and treated groups is expected, with higher scattering in the diseased parts. For $a$, statistical significance of the distribution symmetry ($p=0.02$) […] was observed.
Scattering power $b$ resulted in the values around $1$. This value agrees with the results and spread of values reported in the literature [18,23].
A statistically significant change ($p=0.07$) was observed in the mode of scattering power $b$ which is related to the size of the particles scattering light (larger values towards $b\approx 4$ indicating particles smaller than the wavelength, i.e. Rayleigh scattering limit). The mode of $b$ in Fig. 10 shows higher frequency of areas with smaller $b$ in the control group subjects indicating that larger particles are dominant.
Careful examination of histograms in Fig. 6 also reveals a difference in distribution of $b$ across samples between control and diseased subjects. The shape of the distribution of $b$ for the healthy control subjects has a prominent single peak and gradually decays towards $b=4$. In the subjects from the model groups, a second peak appears at higher values, which could be related histologicaly to inflammation and fibrosis. This change of monomodal to bimodal distribution could be in part due to the presence of inflammatory cells, decay of necrotic cells, increased concentration of collagen due to the fibrosis and also due to the separation of muscle fibers caused by edema.
3.3 Comparison to histology
To verify the results obtained by HSI, histological examination of the same abdominal walls was performed by an experienced pathologist. Histological examination revealed no changes of the peritoneal membrane in the control group (Fig. 11(a)) but moderate to severe inflammation of the peritoneal membrane and underlying tissue in both CHX treated groups (i.e. CHX and CHX+R). Histologically, the intensity of infiltration of mononuclear and polymorphonuclear inflammatory cells, the depth of injury and tissue degradation (necrosis), and the amount of intercellular fluid (edema) in both CHX treated groups were evaluated (Fig. 11(b-d)). However, no significant differences were found in the severity of inflammation between the CHX and CHX+R groups, which is in agreement with the HSI results.
Fig. 11. Histological sections of normal and damaged mouse abdominal wall (hematoxylin eosin stain, 40x magnification). (a) A normal abdominal wall showing a longitudinal section of skeletal muscles covered by a thin layer of parietal peritoneum (arrow). (b) An abdominal wall with damaged inner layer of skeletal muscles (bracket). There are necrotic muscle cells infiltrated with inflammatory cells and increased amount of intercellular fluid (edema). (c,d) An abdominal wall with damaged inner and outer layer of muscle cells (brackets). Muscle cells are necrotic, surrounded by increased amount of intercellular fluid and inflammatory cells. A vein (asterix) is dilated and filled with erythrocytes. Scale bar length equals $100~\mathrm {\mu m}$. Black arrow on the right indicates from which side of the tissue HSI imaging was performed.
Fig. 12. Histological section of normal and damaged parts of a mouse abdominal wall discriminating collagen fibers (blue) from muscles (red) (Masson Trichrome stain, 100x magnification). (a) A normal part of the abdominal wall showing scarce amount of collagen fibers under the parietal peritoneum (blue arrow). (b) A part of an abdominal wall with a damaged inner layer of skeletal muscles and a higher amount of collagen fibers on both sides of the wall and around the area of injection (blue coloured areas/stripes). Scale bar length equals $100~\mathrm {\mu m}$. Black arrow on the right indicates from which side of the tissue HSI imaging was performed.
Representative histological pictures of all three groups are shown in Fig. 11. Panel (a) shows morphological characteristics of abdominal wall and peritoneal membrane in healthy subjects (S01), while panel (b) shows changes observed in CHX+R group (S08), where pronounced edema and infiltration of inflammatory cells on the peritoneal side of the abdominal wall is seen. Skeletal muscles denoted with curly brackets indicate muscle cells in the process of degradation (necrotic) and surrounded with the interstitial fluid, which is seen as empty spaces between cells. Panels (c) and (d) show changes observed in the CHX group (S04), where necrotic muscle cells and infiltration of inflammatory cells and edema are seen as widening of the subperitoneal space and edema.
In addition, increased amount of collagen fibers (fibrosis) was also observed in subjects in both CHX treated groups. However, the intensity of fibrosis varied among subjects in both CHX treated groups (Fig. 12). Panel (a) shows part of a normal abdominal wall in a subject from the CHX+R group (S09), where only a stripe of collagen fibers is seen (blue colored). Increased production of collagen fibers seen as a thicker stripe of blue colored fibers in the treated model group subject (S10) is shown in panel (b).
The increased mode of $BVF_O$ was related to subjects exhibiting advanced inflammation and necrosis. As an example, in subjects S08 and S05 exhibiting the highest $BVF_O$ mode (Fig. 7), the most prominent necrosis was confirmed histologically.
In general, healthy subjects exhibited larger total blood volume fraction (BVF) distribution FWHM (Fig. 8), indicating that the concentration of blood varies within the imaged area unlike CHX animals, where perfusion is increased evenly. An exceptionally high BVF FWHM in subject S09 can be attributed to localized bleeding, as observed in histology.
HSI measured asymmetry of scattering parameter at $500~\mathrm {nm}$ $a$ distribution was substantially increased in subjects S06 and S09 (Fig. 9). Histologically, these subjects exhibited changes localized to a single side of the linea alba and localized focal changes, respectively.
The mode of the scattering power $b$ distribution was highest for subjects S10 and S08 (Fig. 10). Histologically, S10 exhibited pronounced infiltration and fibrosis, while in subject S08 abundant necrosis and edema were present. Changes in $b$ could be attributed to the increased concentration of smaller inflammatory cells and collagen fibers in subject S10 and changes in scattering due to the increased separation between the muscle cells caused by the edema in subject S08.
In general, subject S09 exhibited values closer to the healthy population (with the exception of $a$ symmetry), which is in agreement with localized changes observed histologically; the majority of the sample from subject S09 was, in fact, histologically closest to healthy control group.
A good association was observed between the histological findings and the calculated statistical metrics from HSI data. The subjects exhibiting extreme values in the statistical metrices in HSI also exhibit extreme histological changes. Thus, HSI is not only able to differentiate between diseased and healthy subjects, but can also estimate the degree of change in the tissues of the diseased subjects.
4. Discussion
In this study, the capability to detect peritonitis and fibrosis in mouse models of peritoneal fibrosis using the hyperspectral imaging was demonstrated, exhibiting high statistical significance (Table 2). The values for total blood volume fraction for the diseased and healthy animals were in range of $\approx 5\%$, which agrees well with the standard physiological volume fractions. The scattering parameter at $500~\mathrm {nm}$ $a$, as measured in this study, translates to values of $\mu _s(\lambda =500~\mathrm {nm})\approx 172~\mathrm {cm^{-1}}$, which is in agreement with the values reported in the literature. It is important to note that this result depends on sample itself and its preparation, so values in literature exhibit a large spread. The same is true for scattering power $b$, which was approximately $1$, and also agrees with previous studies of mice tissues.
HSI findings were also confirmed by histology, showing that HSI is sensitive to inflammation and related changes, such as necrosis, infiltration of inflammatory cells, edema, and fibrosis. No difference between the treated and untreated model groups were deemed statistically significant according to the HSI data, which was also confirmed by the histology. To support the interpretation of the imaging results, a simple MCML on CUDA [11] simulation was performed to evaluate the imaging penetration depth. Sample thickness, as measured by 3D laser profilometry, ranged from few hundred microns to about $1.5~\mathrm {mm}$ at the thickest parts along the linea alba. For simulations, values for absorption, scattering and anisotropy at $500~\mathrm {nm}$ of rat muscle ($\mu _a=1.17~\mathrm {cm^{-1}}$, $\mu _s=89~\mathrm {cm^{-1}}$ and $g=0.903$) and a typical refractive index of animal tissues ($n=1.4$) were obtained from literature [24]. The MC simulation results show that even at the greatest thickness, total transmittance through the simulated sample was more than 30%, meaning that some light can reach through the whole sample thus imaging the whole depth. This also agrees with the observed translucency of the samples. Although this indicates that all parts of the sample are probed, changes near the imaged surface of the sample have greater effect on the observed spectra. All samples were imaged from peritoneal side, as indicated by the arrows in Figs. 11 and 12, where changes are more pronounced.
Statistically significant changes between healthy and diseased groups were observed in oxygenated blood volume fraction. These changes are not yet completely understood from a medical point of view, due to the novelty of the technique. Nonetheless, observation of progressed necrosis in subjects exhibiting highest observed modes of oxygenated hemoglobin volume fractions indicates that these effects could be due to the changes in blood circulation or different local chemical environments. No significant changes in mode values for total blood volume fractions were observed, which could be due to the variability between subjects. There is, however, a difference in blood distribution as indicated by statistically significant lower total blood volume fraction FWHM in diseased subjects. These could be related to locally increased blood perfusion in the diseased subjects. These effects will have to be studied in future in greater detail to offer a good explanation of results observed in this study.
The light transport model presented in this work was successfully used to extract physiological and morphological parameters. Although some assumptions were made while constructing the model, it offers a stable fit performance on most of the samples and fast processing time in the range of seconds per image pixel. Thus some accuracy is sacrificed to obtain a fast and stable method. The model could be improved by rigorously treating boundary conditions stemming from refractive index mismatch but it would introduce another free parameter and decrease the stability of the fit. By omitting the refractive index mismatch, multiple internal reflections are not included in the model. For mismatch of refractive indexes of biological tissues $n_1\approx 1.4$ in air $n_2=1.0$, a relative error of $\approx 3\%$ is expected under normal incidence, as given by Fresnel equations.
The main challenge in the study was posed by the sample nature. Murine abdominal wall samples are very thin ($100~\mathrm {\mu m}$ - $1~\mathrm {mm}$) and therefore they appear translucent. Thus, the amount of the backscattered light that is detected is small, causing low signal to noise ratio with no characteristic spectral features. Since samples are thin and covered with membranes, specular reflection can be present, despite using crossed polarizers. Areas with increased specular reflection correspond to the curved parts of the sample, where polarization can change upon reflection. Additionally, interference patterns were observed during the measurements on some parts of the samples, which could be attributed to membranes covering the sample acting as an interferometer. A similar effect called wing interference pattern (WIP) was observed in insect wings, where structure of the wing is covered with a membrane [25].
The presented hyperspectral imaging of the animal models could also benefit human patients by improving the understanding of the disease progression and treatment outcomes e.g. in PD. Currently, biopsy is a golden standard to evaluate peritoneal changes in humans, which can be limited by size and sampling positions of the evaluated regions [26,27]. Although it was recently demonstrated that an ultrasound based examination of peritoneal membrane thickness is possible [28], it is inaccurate and requires experienced personnel. Therefore, alternative techniques are still needed. The presented imaging approach could be translated to human subjects. Due to the different thickness and optical properties of human tissues compared to the animal tissues, the light propagation model would have to be modified and evaluated. Since animal models are designed so that both morphological and physiological changes are similar to those observed in humans, changes of optical properties should be similar as well. Other more accurate models of light propagation could also be used, such as spectrally-constrained steady-state diffuse reflectance [29] using MCML [11] to solve the inverse problem.
Another problem of the translation is light penetration depth. Human peritoneum is thicker compared to mouse peritoneum ($\approx 100~\mathrm {\mu m}$ [28] vs few $\mathrm {\mu m}$ in mice as seen from histology in our study), therefore in humans light is reflected mostly from the peritoneum while in mice it penetrates to the underlying tissues. A useful consequence is, that although only changes close to the surface would be detected in humans, they have a greater effect on the recorded spectra due to the higher probability of tissue-light interactions along this longer path.
A possible approach to clinically perform hyperspectral imaging of human peritoneum would be endoscopic imaging during the PD catheter replacement procedure. Histology in Figs. 11 and 5 shows that changes are most pronounced near the peritoneal tissue surface, which would be accessible for imaging with a hyperspectral imaging endoscope [30] during the procedure.
To verify and test the approach before clinical application, a set of measurements on tissue phantoms with known optical properties would have to be performed. To mimic the optical properties and structure of peritoneum as well as underlying muscles, a layered, solid tissue phantom could be used [31] with addition of purified hemoglobin to mimic blood.
5. Conclusion
It was shown that hyperspectral imaging in combination with a custom analytical model tailored to the application can provide statistically significant discrimination between a healthy murine group and a group with induced peritonitis ($p<0.02$). The most significant changes detected in the physiology of the diseased subjects were in oxygenated blood volume fractions, information not offered by the traditional examination techniques. Furthermore, morphological changes in tissue structure due to the changes associated with the inflammation and fibrosis are discernible through evaluation of the scattering related parameters. HSI findings were compared to histology, where most of the HSI observed changes could be related to pathologies observed in tissue sections.
Hyperspectral imaging shows great promise in assisting researchers and medical personnel in studies where traditional examination techniques are limited either by spatial resolution or limited area coverage. It can offer an effective mezoscopic scale tool that could guide other examination techniques in the future while offering additional insight into properties of the tissues, such as blood volume fraction and scattering.
Funding
Javna Agencija za Raziskovalno Dejavnost RS (J2-8171, P1-0389, P3-0054, P3-0314).
Acknowledgments
Authors would like to acknowledge the help of the Slovenian Society for Rheumatology development.
Disclosures
The authors declare that there are no conflicts of interest related to this article.
References
1. Y. Tomino, “Mechanisms and interventions in peritoneal fibrosis,” Clin. Exp. Nephrol. 16(1), 109–114 (2012). [CrossRef]
2. R.-P. Lee, C.-J. Lee, Y.-M. Subeq, and B.-G. Hsu, “A model of chlorhexidine digluconate-induced peritoneal fibrosis in rats,” Tzu Chi Med. J. 24(3), 108–115 (2012). [CrossRef]
3. Y. Ito, H. Kinashi, T. Katsuno, Y. Suzuki, and M. Mizuno, “Peritonitis-induced peritoneal injury models for research in peritoneal dialysis review of infectious and non-infectious models,” Ren. Replace. Ther. 3(1), 16 (2017). [CrossRef]
4. Y. Ishii, T. Sawada, A. Shimizu, T. Tojimbara, I. Nakajima, S. Fuchinoue, and S. Teraoka, “An experimental sclerosing encapsulating peritonitis model in mice,” Nephrol., Dial., Transplant. 16(6), 1262–1266 (2001). [CrossRef]
5. H. Suga, S. Teraoka, K. Ota, S. Komemushi, S. Furutani, S. Yamauchi, and S. Margolin, “Preventive effect of pirfenidone against experimental sclerosing peritonitis in rats,” Exp. Toxicol. Pathol. 47(4), 287–291 (1995). [CrossRef]
6. A. Gowen, C. Odonnell, P. Cullen, G. Downey, and J. Frias, “Hyperspectral imaging - an emerging process analytical tool for food quality and safety control,” Trends Food Sci. Technol. 18(12), 590–598 (2007). [CrossRef]
7. C. Balas, G. Epitropou, A. Tsapras, and N. Hadjinicolaou, “Hyperspectral imaging and spectral classification for pigment identification and mapping in paintings by el greco and his workshop,” Multimed. Tools Appl. 77(8), 9737–9751 (2018). [CrossRef]
8. M. Govender, K. Chetty, and H. Bulcock, “A review of hyperspectral remote sensing and its application in vegetation and water resource studies,” Water SA 33, 2 (2009). [CrossRef]
9. G. Lu and B. Fei, “Medical hyperspectral imaging: a review,” J. Biomed. Opt. 19(1), 010901 (2014). [CrossRef]
10. L. Wang, S. L. Jacques, and L. Zheng, “MCML–monte carlo modeling of light transport in multi-layered tissues,” Comput. Methods Programs Biomed. 47(2), 131–146 (1995). [CrossRef]
11. E. Alerstam, T. Svensson, and S. Andersson-Engels, “Parallel computing with graphics processing units for high-speed monte carlo simulation of photon migration,” J. Biomed. Opt. 13(6), 060504 (2008). [CrossRef]
12. L. Rogelj, U. Pavlovčič, J. Stergar, M. Jezeršek, U. Simončič, and M. Milanič, “Curvature and height corrections of hyperspectral images using built-in 3d laser profilometry,” Appl. Opt. 58(32), 9002–9012 (2019). [CrossRef]
13. R. Dolenec, L. Rogelj, J. Stergar, and M. Milanic, “Modular multi-wavelength led based light source for hyperspectral imaging,” in Novel Biophotonics Techniques and Applications V, vol. 11075 (International Society for Optics and Photonics, 2019), p. 110751M.
14. L. Rogelj, U. Pavlovčič, M. Jezeršek, M. Milanič, and U. Simončič, “Reducing object curvature and height variation effects in hyperspectral images,” in Diffuse Optical Spectroscopy and Imaging VII, vol. 11074 (International Society for Optics and Photonics, 2019), p. 110742E.
15. U. Pavlovčič, J. Diaci, J. Možina, and M. Jezeršek, “Wound perimeter, area, and volume measurement based on laser 3d and color acquisition,” BioMed. Eng. OnLine 14(1), 39 (2015). [CrossRef]
16. M. Jezersek and J. Mozina, “A laser anamorph profilometer,” Strojniski Vestnik 49, 76–89 (2003).
17. P. McCulloch, D. G. Altman, W. B. Campbell, D. R. Flum, P. Glasziou, J. C. Marshall, and J. Nicholl, “No surgical innovation without evaluation: the IDEAL recommendations,” Lancet 374(9695), 1105–1112 (2009). [CrossRef]
18. S. L. Jacques, “Optical properties of biological tissues: a review,” Phys. Med. Biol. 58(11), R37–R61 (2013). [CrossRef]
19. S. Prahl, “Tabulated molar extinction coefficient for hemoglobin in water,” https://omlc.org/spectra/hemoglobin/summary.html [accessed 2. 8. 2019].
20. S. K. V. Sekar, I. Bargigia, A. D. Mora, P. Taroni, A. Ruggeri, A. Tosi, A. Pifferi, and A. Farina, “Diffuse optical characterization of collagen absorption from 500 to 1700 nm,” J. Biomed. Opt. 22(1), 015006 (2017). [CrossRef]
21. R. E. Neuman and M. A. Logan, “The determination of collagen and elastin in tissues,” J. Biol. Chem. 186, 549–556 (1950).
22. L. G. Henyey and J. L. Greenstein, “Diffuse radiation in the galaxy,” Astrophys. J. 93, 70–83 (1941). [CrossRef]
23. G. Hall, S. L. Jacques, K. W. Eliceiri, and P. J. Campagnola, “Goniometric measurements of thick tissue using monte carlo simulations to obtain the single scattering anisotropy coefficient,” Biomed. Opt. Express 3(11), 2707–2719 (2012). [CrossRef]
24. A. N. Bashkatov, E. A. Genina, and V. V. Tuchin, “Optical properties of skin, subcutaneous, and muscle tissues: a review,” J. Innovative Opt. Health Sci. 04(01), 9–38 (2011). [CrossRef]
25. E. Shevtsova, C. Hansson, D. H. Janzen, and J. Kjarandsen, “Stable structural color patterns displayed on transparent insect wings,” Proc. Natl. Acad. Sci. U. S. A. 108(2), 668–673 (2011). [CrossRef]
26. T. Mizumasa, K. Honda, S. Aoki, C. Hamada, M. Miyazaki, Y. Ito, Y. Tanno, T. Nakano, and M. Nakayama, “Proposal of peritoneal biopsy procedures for patients undergoing peritoneal dialysis,” Ren. Replace. Ther. 6(1), 8 (2020). [CrossRef]
27. K. Kawanishi, K. Honda, and C. Hamada, “Recommendations for pathological diagnosis on biopsy samples from peritoneal dialysis patients,” Pleura peritoneum 2(1), 3–15 (2017). [CrossRef]
28. A. C. Abrahams, A. Dendooven, J. W. van Der Veer, R. Wientjes, R. J. Toorop, R. L. Bleys, A. P. Hendrickx, M. S. van Leeuwen, Q. G. de Lussanet, M. C. Verhaar, G. Stapper, and T. Q. Nguyen, “Direct comparison of the thickness of the parietal peritoneum using peritoneal biopsy and ultrasonography of the abdominal wall in patients treated with peritoneal dialysis,” Peritoneal Dial. Int. 39(5), 455–464 (2019). [CrossRef]
29. A. J. Welch and M. J. Van Gemert, (eds.) Optical-thermal Response of Laser-irradiated Tissue (Springer, 2011), 2nd ed.
30. R. T. Kester, N. Bedard, L. S. Gao, and T. S. Tkaczyk, “Real-time snapshot hyperspectral imaging endoscope,” J. Biomed. Opt. 16(5), 056005 (2011). [CrossRef]
31. S. K. V. Sekar, A. Pacheco, P. Martella, H. Li, P. Lanka, A. Pifferi, and S. Andersson-Engels, “Solid phantom recipe for diffuse optics in biophotonics applications: a step towards anatomically correct 3D tissue phantoms,” Biomed. Opt. Express 10(4), 2090–2100 (2019). [CrossRef]
