1. Introduction

For cancer patients the presence of metastases dictates the staging assessment which in turn defines the appropriate treatment path selected. For gynecological malignancies, like ovarian carcinoma, it is of immense importance to differentiate between localized and metastatic disease status as that drastically affects management. The peritoneal cavity is the most affected site for ovarian carcinoma [1]. Exact delineation of cancer spread to the abdominal cavity dictates not only the role of neoadjuvant chemotherapy, but also the resectability/debulking potential [2]. Even after major operative resections, disease recurrence is a significant clinical problem for patients with both early and advanced stage ovarian cancer, with the peritoneum being the most common site of recurrence (>75%) [3]. It is presumed that micrometastatic sites already exist at the time of the initial operation, but traditional means of examination used for initial staging, e.g. including cross-sectional radiographic imaging and staging laparoscopy, are unable to completely identify them [4,5]. Therefore, a great need exists for imaging solutions that can identify early metastatic deposits in a label-free and minimally invasive manner.

For in situ, real time diagnosis, novel imaging modalities that offer metabolic and structural information at the cellular and subcellular level can be of great help, especially since these modalities are being progressively incorporated in probes and micro-endoscopes that allow intra-vital access to organs that lie deeper in our bodies [6–10]. Such a modality is two-photon microscopy (TPM) [11]. TPM is a promising non-invasive microscopy modality for subcellular resolution imaging of dense, optically scattering tissue due to the deeper penetration depth of the near-infrared excitation light and the minimized photodamage attributed to confined two-photon excitation to the focal volume [12]. As tissues contain molecules that fluoresce or scatter light naturally when excited under appropriate conditions, the intrinsic contrast from the tissue provides submicron resolution images with sufficient architectural details to allow noninvasive analysis of cellular morphology and extracellular matrix architecture, without the use of exogenous dyes or agents [13].

In this this study, endogenous signals from cellular and fibrillar collagen features, produced by two photon excited fluorescence and second harmonic generation (a coherent two-photon scattering process) were evaluated as potential markers for diagnostic tissue evaluation. As cancer invasion affects tissue architecture originally at the microscopic level, label free microscopy modalities that can sample early tissue changes have the potential to better detect otherwise occult cancer metastases. The goal of this project was to set the foundation for an in situ, real time image analysis technique that could provide automated evaluation and classification of healthy and diseased peritoneal tissues using label-free two-photon imaging.

2. Materials and methods

2.1 Tissue collection

Adult patients (>=18 years of age) who were evaluated at Lahey Hospital & Medical Center for a suspected or confirmed diagnosis of ovarian malignancy and who were planned to undergo open operative resection or biopsy of the malignancy as part of their routine treatment plan were recruited for the study. All study patients (N = 8) underwent open laparotomy as part of routine medical care. Post completion of all intra-abdominal procedures of the operation, biopsies of healthy parietal peritoneum (N = 8) and if present of peritoneal metastases (N = 4) were collected from each patient. Tissue sampling was conducted according to a protocol approved by the Lahey Clinic Institutional Review Board with written informed consent from all participants. All lesions were evaluated by a pathologist at Tufts Medical Center (E.M.G.), using standard hematoxylin and eosin histology. A statistical size calculation was not possible prior to the study initiation as this was the first investigational study in human samples. Sample sizes were set so as to secure validation of the statistical analysis assumptions.

2.2 Imaging

The tissues were imaged employing a multiphoton laser scanning microscope to generate intrinsic fluorescence and SHG images at 755 nm and 900 nm excitation respectively with signal emission collected at 460 ± 20 and 525 ± 25 nm. Laser light was focused on the sample using a 25x objective (0.9 NA / water-immersion), and neutral density filters were employed to achieve a power of 25–35 mW. At least 2 to 3 random fields per tissue were evaluated, reaching a total of 30 and 11 images for the healthy and metastatic biopsy tissue groups, respectively (512 × 512 pixels; 600-micron field of view; resolution of 1.17 microns per pixel). Imaging was focused within a depth of ∼20-100 microns from the mesothelial surface of the tissues.

2.3 Image processing and analysis

Image processing was performed in Matlab (MathWorks, Natick, MA). All acquired images were normalized for illumination power and detector gain. Image data were evaluated utilizing custom automated batch processing algorithms and image texture analysis techniques, namely gray level co-occurrence matrices (GLCM) [14]. GLCM texture analysis is a statistical method of examining quantitatively the arrangement of image intensities, by considering not only the pixels’ intensity values but also their spatial relationship. As such, texture analysis attempts to statistically quantify intuitive, visually perceived image qualities such as the difference in image areas’ luminance (e.g. image contrast) as a function of the spatial variation in pixel intensities. For example, an image filled with pixels having entirely similar intensity values to their neighboring pixels would have zero contrast. GLCM texture analysis considers the relation between two pixels at a time, called the reference and the neighbor pixel. Here, for each image pixel, 12 increasing pixel distances (1, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20 and 25 pixels) and 4 orientations (0, 45, 90 and 135 degrees) were evaluated. As each GLCM matrix was considered symmetric, information from each respective supplementary angle (e.g. 180, 225, 270 and 315 degrees) was simultaneously incorporated. Two textural features (contrast and correlation) were computed to evaluate the spatial distribution of intensity variations within each image, using 64 intensity level quantization. Contrast provides a measure of the intensity dissimilarities between neighboring pixels, while correlation provides a measure of pattern repetitiveness. Contrast analysis was performed in the full field of view image signal of the fluorescence and SHG channels, whereas correlation analysis was performed utilizing only areas characterized by fibrillar collagen SHG-positive image signal from the same channels respectively (Fig. 1(a), Appendix Fig. 5). To avoid zero intensity image gaps that can affect the statistical calculations and to enable automated full image analysis without manual region of interest selection, the image intensities contained within SHG-positive areas (Appendix Fig. 5(c)) were digitally cloned to fill any zero intensity image gaps from SHG-negative areas (Appendix Fig. 5(d)), using a custom automated digital object cloning (DOC) technique as described previously [15,16]. In brief, the segmented fluorescence or SHG channel image (Appendix Fig. 5(c)) is digitally shifted with a random displacement upon itself and its SHG-positive image intensity patterns are utilized to overwrite void regions (zero intensity pixels) created by the segmentation (Appendix Fig. 5(b-c)). This process is repeated iteratively until no zero intensity pixels remain (Appendix Fig. 5(d)). The random displacement secures that there is no user bias in preferentially selecting one region of the image to fill in the segmented gaps and the lack of pixel shuffling or foreground overwriting preserves the inherent textural heterogeneity contained within the image. SHG-positive image pixels were determined using the 900/460 channel where SHG signal dominated, by applying a threshold of 10% of the maximum image value to remove background. If a field contained cellular features, the 755(460) and 900(525) channels were utilized to identify those features by custom bandpass thresholding as described previously [15].

 

Fig. 1. Cancer invasion alters tissue component constitution A. Examples of two-photon label-free images capturing SHG (pseudocolored green) and autofluorescence (pseudocolored red) signals for healthy parietal peritoneum and ovarian metastatic tissue. Signal overlay reveals spatial overlap of collagen SHG and fluorescence in healthy tissues, whereas destruction of matrix components by cancerous cellular infiltration is observed in metastasis. Insert in ovarian metastasis fluorescence image highlights the presence of cellular clusters (white arrows), identified by dark nuclei and bright cytoplasmic features. B. Quantification of image area coverage by automated image feature segmentation as shown in A. *denotes significance at a = 0.05 C-D. Examples of histological H&E microscopic images for each tissue group (C-Healthy,D-Metastasis). Black arrows indicate mesothelial surface while red arrows highlight infiltrative cancerous cellular clusters in the metastatic biopsy. Mesothelial cells were not typically recognized in the examined sections.

Download Full Size | PDF

2.4 Statistics

The extracted image analysis outcomes were finally used for automated tissue classification using quadratic discriminant analysis, which does not rely on the assumption that the observations vary consistently across all classes. Before proceeding with the discriminant analysis, the mean values for each group and predictor variable were compared using univariate nested model t-tests. For each t-test, a nested design considered the lack of biological independence for the multiple optical fields derived from individual patient/tissues, with statistical group outcomes reported ultimately at the participant/patient level. Significance level for all statistics was set to α = 0.05. All statistical analyses were performed in JMP Pro 13 (SAS Institute, Cary, NC).

3. Results

3.1 Signals from extracellular matrix components dominate the two-photon image features generated from healthy parietal peritoneal tissues

Collagen fibers, the main constituents of the sub-mesothelial tissue stroma, naturally produce a two-photon (2P) scattering signal, named second harmonic generation (SHG) and an associated fluorescence signal created by the crosslinks between the collagen fibers (Fig. 1(a), 1st row). In healthy tissues the fibrillar SHG and fluorescence collagen signals spatially overlap (Fig. 1(a), 1st row) and typically dominate the image area (Fig. 1(a)). In contrast, in metastatic tissues extracellular collagen coverage decreases as infiltrative cellular populations appear and overtake the parenchyma (Fig. 1(a), 2nd row). We note that SHG is generated specifically by fibrillar structures, which lack a center of symmetry, and not cells. The contributions of the extracellular matrix components versus cellular contributions were quantified in an automated manner and show statistically different outcomes between healthy and metastatic tissues, in agreement with the visual observations (Fig. 1(a-b)) and histological findings (Fig. 1(c-d)).

To evaluate more quantitatively the divergent image patterns between healthy and diseased tissues, we proceeded with GLCM image texture analysis (Fig. 2). As healthy tissues display densely alternating bright (redder hues) and darker (bluer hues) repetitive patterns, mostly governed by the presence of the fibrillar collagen fibers, that span through large image areas (Fig. 2(a)), image contrast is expected to be high. On the contrary, in diseased tissues as collagen fibers and associated crosslinks are diminished effectuated by the cancerous infiltration, large image areas with decreased luminance and overall unvaried intensity patterns appear, which are expected to negatively affect image contrast values. Indeed, polar representation of extracted contrast values as a function of angle and distance (Fig. 2(a)) and mean contrast values averaged over all angles and examined distances (Fig. 2(b)) for each tissue type statistically corroborate the anticipated contrast decrease in metastatic tissues.

 

Fig. 2. Automated image contrast texture analysis. A. (top row) SHG and Fluorescence intensity images from healthy and metastatic parietal peritoneum. For all panels of top row, image heat maps reflect varying image intensities. A multicolor image heat map was employed here to better highlight visually the differences between the image intensities. Color and scale bars are the same for all images of the panel. (bottom row) Corresponding polar plots graphically depicting contrast values as a function of angle and distance. Hemi circle center corresponds to contrast values within 1-pixel distance and dotted concentric semicircles represent increasingly distances of 2, 4, 6, 8, 10, 12, 14, 16, 18, 20 and 25 pixels (1 pixel ≈1.2 microns). Plot heat maps reflect varying extracted contrast values. Color bars are the same for all plots of bottom row panel B. Mean contrast values averaged over all angles and distances for each tissue group. *denotes significance at a = 0.05

Download Full Size | PDF

3.2 Collagen fiber thickness and associated crosslink patterns are affected by cancerous infiltration in metastatic tissues

Based on these initial observations of the healthy matrix stromal destruction within the metastatic tumor tissues, we also sought to investigate whether fiber thickness and their associated crosslink patterns were also affected by cancerous infiltration. For that purpose, we utilized another textural feature, namely correlation. The correlation measures the linear dependency of intensity levels between the defined neighboring pixels [17]. High correlation outcomes translate to high predictability of pixel relationships, e.g. repetitive appearance of patterns [18]. Exploiting that property, if correlation is sequentially extracted for larger distances, the distance at which the correlation value declines to 50% of its initial value (D₅₀) may be used as an estimation of the average size of definable objects within the evaluated image area (Fig. 3(a)) [17,18]. As collagen fibers are typically longer than thicker, the D₅₀ value correlates with fiber thickness rather than length, as in that direction the correlation between neighboring pixels decreases sharply. Collagen associated areas were isolated using segmentation masks as shown in Fig. 1(a) and Appendix Fig. 5 and a custom digital cloning algorithm filled any zero intensity pixels with the original collagen-positive pixel information (Fig. 3(a), Appendix Fig. 5). The latter served a double purpose. First, it minimized large image gaps, as zero intensity pixels can erroneously affect the statistical robustness of the extracted textural features and further enabled automated full image evaluation without user defined regions of interest. Both aspects are important when considering algorithmic translation to a clinical environment for real-time, unbiased image evaluation.

 

Fig. 3. Automated image correlation texture analysis of collagen associated areas. A. (top row) Representative SHG and fluorescence intensity images from healthy and metastatic parietal peritoneum as shown in Fig. 2(a) but processed to contain only collagen-SHG related information. For all panels, image heat maps reflect varying image intensities. Color and scale bars are the same for all images of the panel. (bottom row) Corresponding correlation plots graphically depicting correlation values as a function of increasing distance (1, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20 and 25 pixels (1 pixel ≈1.2 microns). Asterisks and blue lines signify for each plot the distance where correlation declines to 50% of its initial value (D₅₀) for this set of images. B. Mean D₅₀ values from analysis of each type of image for the two tissue groups. *denotes significance at a = 0.05

Download Full Size | PDF

Correlation for the fibrils and associated crosslinks remained elevated for larger distances (Fig. 3(a-b)) in the healthy tissues with D₅₀ correlation values being at a ∼2-30 pixel range and a 12 pixel average (≈14 microns), well within the range of previously described collagen fiber thicknesses [19,20]. In metastatic tissues, correlation of collagen SHG and associated fluorescent crosslink signal, decreased more sharply, with statistically significant differences detected between the two tissue groups (Fig. 3(b)).

3.3 Morphological and textural outcomes enable high classification accuracy in classifying healthy versus metastatic peritoneal tissues as evaluated quantitatively by discriminant analysis

Lastly, to test the potential of the aforementioned metrics as quantitative diagnostic biomarkers, the extracted image coverage and textural statistics were utilized through quadratic discriminant analysis to classify in an automated manner the captured images for each tissue group. Quadratic discriminant analysis is a more flexible classifier than linear discriminant analysis and is advantageous when the covariance of the predictor variables is not common across all classes (Appendix Fig. 6) [21]. The model resulted in good spatial discrimination of the tissue groups (Fig. 4(a)) and quantitatively resulted in 97.5% accuracy (40/41 images correctly classified), 100% sensitivity (11/11 metastatic images correctly classified) and 96.6% specificity (29/30 healthy images correctly classified). All metastatic tissue images were classified as diseased, resulting in zero false negatives and a perfect score of sensitivity (Fig. 4(b)). As extraction of the quantitative image statistics had a computational cost of less than a second, textural analysis of intrinsic contrast images holds excellent translational potential for real time image classification in a clinical setting.

 

Fig. 4. A. Plot of the canonical QDA discriminant scores showing the separation of the healthy (black; N = 8 participants; 30 sampled areas) and diseased metastatic (red; N = 4; 11 sampled areas) tissues, on the basis of the morphological tissue metrics extracted from SHG and Fluorescence image analysis as shown in Fig. 1–3. Each point represents one sampled area. Outer line ellipses represent 50% of data coverage. Crosses display group means and inner line ellipses indicate 95% confidence intervals for the mean of each tissue group respectively. As the two groups differ significantly, the confidence ellipses do not intersect B. Original classification outcomes presented based on the comparison of the QDA model predictions for the healthy and metastatic disease groups with corresponding histopathological evaluations and extracted accuracy, sensitivity and specificity outcomes.

Download Full Size | PDF

4. Discussion and conclusions

The goal of this study was to set the foundation for an in situ, real time diagnostic analysis technique that could provide accurate classification of healthy parietal peritoneal and ovarian metastatic tissues. For this purpose, we acquired collagen-SHG and fluorescence intensity images in a label free manner via two-photon microscopy from healthy and metastatic human peritoneal tissue biopsies collected from ovarian cancer patients. We implemented GLCM textural image analysis algorithms along with discriminant analysis models to evaluate the extracted image parameters for automated, user-unbiased tissue diagnostics. We found that the organization of the fibrillar collagen in metastatic tissues was altered, most likely due to remodeling of the extracellular matrix during the parenchymal invasion from malignant cellular populations [22]. Matrix destruction led to decreased detected textural image contrast values. Collagen fiber thickness and associated fluorescence crosslinks were also decreased in metastatic ovarian tissues as evaluated by textural correlation analysis. When utilized by discriminant analysis, the extracted image parameters resulted in high classification accuracy (97.5%), sensitivity (100%) and specificity (96.6%).

These preliminary outcomes, although extracted from a limited ex vivo cohort of peritoneal biopsies, confirm that morphological optical differences exist between healthy and metastatic peritoneal tissues, suggestive of functional matrix changes since the structure and organization of collagen impacts significantly tissue stiffness [22,23]. Automated image analysis and statistical models can be implemented in near real time for accurate tissue classification. Future steps would target the evaluation of a greater image sample from a more extended patient population. More ample data pool would further allow the implementation of prospectively predictive models to further validate clinical translational potential. Nonetheless our results indicate that two-photon SHG and fluorescence imaging in combination with automated analytical approaches enables in near real time noninvasive, user unbiased tissue classification. As these analytical techniques can be coupled with rapidly advancing miniaturization of two-photon imaging systems to afford their use in clinical situations, they hold great potential in assisting surgical assessment of peritoneal tissues at the bedside.

Appendix

Appendix figures

 

Fig. 5. Automated image analysis steps for isolation of extracellular matrix associated features prior to textural correlation analysis. A. Example of unfiltered intensity image as shown in Fig. 3(a). This image will be utilized for textural contrast analysis B. Segmentation mask identifying ECM collagen positive associated features in green as also shown in Fig. 1(a). C. Segmented areas from raw image after application of segmentation mask. D. Final intensity image as shown in Fig. 3(a), after the automated digital object cloning (DOC) algorithm that utilizes the segmented areas in C. to fill any segmented voids produced by any non-ECM feature removal. This image will be finally utilized for the textural correlation analysis. These analytical steps were commonly used for isolation of the fibrillar collagen SHG-positive image signal in the SHG and fluorescence channels respectively. Scale bar (100 µm) is same for all images of figure and colorbar is same for images A,C and D.

Download Full Size | PDF

 

Fig. 6. Scatterplot Matrix reporting covariances for each classification group and each pair of covariates. Observations vary differentially across classes and covariate pairs.

Download Full Size | PDF

Funding

National Institute of Biomedical Imaging and Bioengineering (NIBIB) (NIBIB- R21 EB023498); Hellenic Medical Society of New York (Stavros Hartofilis fellowship); Society of American Gastrointestinal and Endoscopic Surgeons (SAGES) (SAGES 2013 Research Award).

Author contributions

D.P., I.G. and T.S. were responsible for conceiving and designing the study; T.B.S, K.M.R-C., V.W., and T.S. were involved in patient recruitment; E.M.G evaluated histopathologically the samples; D.P. performed the imaging experiments, algorithm development, analyses and data interpretation; D.P., I.G. and T.S. wrote the manuscript which was reviewed and edited by all co-authors.

Disclosures

The authors declare that there are no conflicts of interest related to this article

References

1. E. Lengyel, “Ovarian cancer development and metastasis,” Am. J. Pathol. 177(3), 1053–1064 (2010). [CrossRef]  

2. S. H. Chandrashekhara, G. S. Triveni, and R. Kumar, “Imaging of peritoneal deposits in ovarian cancer: A pictorial review,” World J Radiol 8(5), 513–517 (2016). [CrossRef]  

3. P. Amate, C. Huchon, A. L. Dessapt, C. Bensaid, J. Medioni, M. A. Le Frere Belda, A. S. Bats, and F. R. Lecuru, “Ovarian cancer: sites of recurrence,” Int J Gynecol Cancer 23(9), 1590–1596 (2013). [CrossRef]  

4. C. M. Patel, A. Sahdev, and R. H. Reznek, “CT, MRI and PET imaging in peritoneal malignancy,” Cancer Imaging 11(1), 123–139 (2011). [CrossRef]  

5. Y. Kodera, “Gastric cancer with minimal peritoneal metastasis: is this a sign to give up or to treat more aggressively?” Nagoya J Med Sci 75(1–2), 3–10 (2013).

6. T. Schnelldorfer, “Image-enhanced laparoscopy: a promising technology for detection of peritoneal micrometastases,” Surgery 151(3), 345–350 (2012). [CrossRef]  

7. G. Ducourthial, P. Leclerc, T. Mansuryan, M. Fabert, J. Brevier, R. Habert, F. Braud, R. Batrin, C. Vever-Bizet, G. Bourg-Heckly, L. Thiberville, A. Druilhe, A. Kudlinski, and F. Louradour, “Development of a real-time flexible multiphoton microendoscope for label-free imaging in a live animal,” Sci. Rep. 5(1), 18303 (2015). [CrossRef]  

8. C. M. Brown, D. R. Rivera, I. Pavlova, D. G. Ouzounov, W. O. Williams, S. Mohanan, W. W. Webb, and C. Xu, “In vivo imaging of unstained tissues using a compact and flexible multiphoton microendoscope,” J. Biomed. Opt. 17(4), 040505 (2012). [CrossRef]  

9. D. G. Ouzounov, D. R. Rivera, W. O. Williams, J. A. Stupinski, T. L. Southard, K. H. Hume, J. Bentley, R. S. Weiss, W. W. Webb, and C. Xu, “Dual modality endomicroscope with optical zoom capability,” Biomed. Opt. Express 4(9), 1494–1503 (2013). [CrossRef]  

10. D. R. Rivera, C. M. Brown, D. G. Ouzounov, W. W. Webb, and C. Xu, “Multifocal multiphoton endoscope,” Opt. Lett. 37(8), 1349–1351 (2012). [CrossRef]  

11. F. Helmchen and W. Denk, “Deep tissue two-photon microscopy,” Nat. Methods 2(12), 932–940 (2005). [CrossRef]  

12. P. T. So, C. Y. Dong, B. R. Masters, and K. M. Berland, “Two-photon excitation fluorescence microscopy,” Annu. Rev. Biomed. Eng. 2(1), 399–429 (2000). [CrossRef]  

13. I. Georgakoudi and K. P. Quinn, “Optical imaging using endogenous contrast to assess metabolic state,” Annu. Rev. Biomed. Eng. 14(1), 351–367 (2012). [CrossRef]  

14. R. M. Haralick, K. Shanmugam, and I. Dinstein, “Textural Features for Image Classification,” IEEE Trans. Syst., Man, Cybern. SMC-3(6), 610–621 (1973). [CrossRef]  

15. D. Pouli, M. Balu, C. A. Alonzo, Z. Liu, K. P. Quinn, F. Rius-Diaz, R. M. Harris, K. M. Kelly, B. J. Tromberg, and I. Georgakoudi, “Imaging mitochondrial dynamics in human skin reveals depth-dependent hypoxia and malignant potential for diagnosis,” Sci. Transl. Med. 8(367), 367ra169 (2016). [CrossRef]  

16. J. Xylas, K. P. Quinn, M. Hunter, and I. Georgakoudi, “Improved Fourier-based characterization of intracellular fractal features,” Opt. Express 20(21), 23442–23455 (2012). [CrossRef]  

17. M. Hall-Beyer, “GLCM Texture: A Tutorial v. 3.0 March 2017,” (2017).

18. J. Adur, H. F. Carvalho, C. L. Cesar, and V. H. Casco, “Nonlinear optical microscopy signal processing strategies in cancer,” Cancer Inf. 13, CIN.S12419 (2014). [CrossRef]  

19. T. Ushiki, “Collagen fibers, reticular fibers and elastic fibers. A comprehensive understanding from a morphological viewpoint,” Arch. Histol. Cytol. 65(2), 109–126 (2002). [CrossRef]  

20. S. Cheng, Y. Lu, Y. Li, L. Gao, H. Shen, and K. Song, “Hydrogen sulfide inhibits epithelial-mesenchymal transition in peritoneal mesothelial cells,” Sci. Rep. 8(1), 5863 (2018). [CrossRef]  

21. C. J. Huberty, S. Olejnik, and C. J. Huberty, Applied MANOVA and Discriminant Analysis, 2nd ed., Wiley Series in Probability and Statistics (Wiley-Interscience, 2006), pp. xxxiv.

22. T. R. Cox and J. T. Erler, “Remodeling and homeostasis of the extracellular matrix: implications for fibrotic diseases and cancer,” Dis. Models Mech. 4(2), 165–178 (2011). [CrossRef]  

23. M. A. Karsdal, M. J. Nielsen, J. M. Sand, K. Henriksen, F. Genovese, A. C. Bay-Jensen, V. Smith, J. I. Adamkewicz, C. Christiansen, and D. J. Leeming, “Extracellular matrix remodeling: the common denominator in connective tissue diseases. Possibilities for evaluation and current understanding of the matrix as more than a passive architecture, but a key player in tissue failure,” Assay Drug Dev. Technol. 11(2), 70–92 (2013). [CrossRef]