Potential conflict of interest: A.J.D. is a member of the clinical development team of Omnyx (www.omnyx.com), a corporation developing a complete digital pathology solution. A.L. serves as a consultant for Carl Zeiss MicroImaging, LLC.
Supported by: National Institutes of Health (NIH) R01-AI-081678, N01-AI-15416, U01-A1-077867, P01-A1-064343 (all to A.J.D.), R01-EB-005157, R01-CA-135509 (all to B.R. and W.M.L.) and an S-IDEA grant W81XWH-07-1-0325 (to B.R).
Routine light microscopy identifies two distinct epithelial cell populations in normal human livers: hepatocytes and biliary epithelial cells (BECs). Considerable epithelial diversity, however, arises during disease states when a variety of hepatocyte-BEC hybrid cells appear. This has been attributed to activation and differentiation of putative hepatic progenitor cells (HPC) residing in the canals of Hering and/or metaplasia of preexisting mature epithelial cells. A novel analytic approach consisting of multiplex labeling, high-resolution whole-slide imaging (WSI), and automated image analysis was used to determine if more complex epithelial cell phenotypes preexist in normal adult human livers, which might provide an alternative explanation for disease-induced epithelial diversity. “Virtually digested” WSI enabled quantitative cytometric analyses of individual cells displayed in a variety of formats (e.g., scatterplots) while still tethered to the WSI and tissue structure. We employed biomarkers specifically associated with mature epithelial forms (HNF4α for hepatocytes, CK19 and HNF1β for BEC) and explored for the presence of cells with hybrid biomarker phenotypes. The results showed abundant hybrid cells in portal bile duct BEC, canals of Hering, and immediate periportal hepatocytes. These bipotential cells likely serve as a reservoir for the epithelial diversity of ductular reactions, appearance of hepatocytes in bile ducts, and the rapid and fluid transition of BEC to hepatocytes, and vice versa. Conclusion: Novel imaging and computational tools enable increased information extraction from tissue samples and quantify the considerable preexistent hybrid epithelial diversity in normal human liver. This computationally enabled tissue analysis approach offers much broader potential beyond the results presented here. (HEPATOLOGY 2013)
In normal human livers, routine conventional histology divides epithelial cells into hepatocytes and biliary epithelial cells (BECs). A wide variety of hepatocyte-BEC transitional phenotypes, however, quickly appear in diseased livers. Cells with transitional phenotypes are thought to arise from proliferation and differentiation of hepatic progenitor cells (HPCs)1, 2 and/or from metaplasia of mature hepatocytes and BEC.3 Otherwise, typical hepatocytes can be found in portal tract bile ducts with no connection to lobular-based hepatocytes4 and the liver responds “intelligently” to various insults by rapidly producing more BEC and/or hepatocytes, as needed.5, 6 Understanding the complexity of these wound repair responses using traditional light and electron microscopy limits the type of data that can be extracted from tissue samples.5, 6
Indeed, the “value proposition” of conventional histology is in question because of various shortcomings: tissue biopsies are invasive, expensive, potentially morbidity- and mortality-producing, an unpleasant patient experience, subject to sampling error, and interpretations are prone to bias, subjectivity, and interobserver variability7 compared to less invasive and potentially more informative monitoring techniques, such as whole genome sequencing and peripheral blood and fluid messenger RNA (mRNA) arrays and proteomics/metabolomics.8, 9
Conversely, conventional histopathology quickly provides a wealth of irreplaceable data about structural integrity, spatial and temporal relationships, and rare events/cells. Only a tiny fraction of information is being harvested from tissue slides primarily because data extraction is dependent on a restricted staining repertoire and manual observation. The challenge, therefore, is to develop a modern replacement to the traditional histopathologic approach.
Dramatic advances in robotics, digital imaging, and computing have spawned the “-omics” revolution that challenges routine histopathology for improved tissue utilization. Comprehensive examination of mRNA, protein, and metabolite expression data can be used as powerful screening tools to query disease susceptibility, pathophysiology, and prognosis. These same innovations, however, are also revolutionizing histopathology. High-resolution whole-slide image (WSI) scanners now enable pathologists to view routinely prepared and stained slides on a computer screen instead of a microscope. Pathologists can then team with hardware and software engineers, mathematicians, and image analysis experts to greatly increase the value equation for histopathology in an era when there is increasing pressure to diagnose and monitor liver diseases noninvasively.8
We recently reported on the power of combining WSI, multiplex nanoparticle quantum dot staining, and automated image analysis to envisage and analyze multiple protein labels on a single slide to reveal biological mechanisms.7, 10-16 We use this novel approach to study liver epithelial diversity in formalin-fixed, paraffin-embedded, normal human liver tissue. In this report, automated quantitative data collection that described cell numbers, types, and nuclear/cytoplasmic analyte expression was spatially tethered to the tissue architecture. This enabled us to illustrate and locate preexisting diversity within BECs and hepatocytes in normal human liver, including the transition zone between these cell types in the canals of Hering (CH). This might lead to a better understanding of the considerable diversity that quickly appears during disease states.5
BEC, biliary epithelial cell; CD31, cluster of differentiation 31; CH, canal of Hering; EC, endothelial cell; HNF, hepatocyte nuclear factor; HPC, hepatic progenitor cell; ROI, region of interest; SMA, smooth muscle actin; SMC, smooth muscle cell; WSI, whole slide image.
Materials and Methods
Tissue Sources and Multiplex Quantum Dot Immunolabeling.
Four-μm sections from eight formalin-fixed paraffin-embedded and one frozen liver tissues were used for the analyses (Supporting Table 1; Institutional Review Board protocol 0404010, University of Pittsburgh). Before study inclusion, hematoxylin and eosin (H&E) slides from each case were reviewed. Tissues were prepared and stained as described.7, 10 Liver and putative hepatic progenitor cell (HPC) or CH characterization used the following panels (see Supporting Table 2 for antibody listing): panel A (β-catenin [β-cat], cytokeratin 19 [CK19], α-smooth muscle actin [αSMA], and CD31); and panel B (CK19, hepatocytes nuclear factor [HNF] 1β, and HNF4α).
High-Resolution Whole-Slide Scanning and Image Analysis.
Slides were imaged with a Mirax MIDI WSI scanner equipped with a Plan-Apochromat 40×/.95N.A. objective lens, AxioCam MRm digital CCD camera (Carl Zeiss, Jena, Germany) and specifically selected excitation/emission Qdot filters (Omega Optical, Brattleboro, VT) as described.7, 10 Additional supporting 3D wide-field imagery was created using a robotic AxioImager M1 microscope (Carl Zeiss, Gottingen, Germany) equipped with various dry and oil-immersion high NA objectives and ZEN imaging software (Carl Zeiss, Jena, Germany) for microscope control and image visualization.
Pixel-based image analytics was performed using ImageJ (http://rsbweb.nih.gov/ij/); tissue-tethered cytometry analysis utilized FARSIGHT (http://www.farsight-toolkit.org/wiki/FARSIGHT_Toolkit) and IAE-NearCYTE (http://nearcyte.org). The tissue-tethered cytometry software packages are designed to delineate all cells, define cell types by user classifications, and quantify analyte expression. A total of 20 regions of interest (ROIs) were randomly chosen that centered on portal tracts or central veins (Supporting Fig. 1A). Each ROI was exported into individual grayscale JPGs without compression and imported into FARSIGHT tissue cytometry7, 10 and ImageJ 1.45s for pixel analysis. Cell data obtained from FARSIGHT using panel A (β-cat/CK19/αSMA/CD31/DAPI) staining was classified, analyzed, and sorted by using an active training system (candidates are user-selected) based on nuclear size, elongation, CK19 intensity, β-catenin intensity, β-catenin total signal, CD31 intensity, and αSMA intensity. For certain expression patterns, data were selected and exported into a common delineated format for review with Microsoft Excel [e.g., CK19 = IF(CK19>1000, 1, 0), β-cat = IF(AND(nucleus size>23.4μm2, β-cat total>15000, β-cat surrounding>3, CK19 = negative, CD31 = negative, αSMA = negative), 1, 0), CD31 = IF(AND(CD31 average>48, CD31 surrounding>3, Nucleus size<57.2μm2, CK19 = negative, αSMA = negative), 1, 0), αSMA = IF(AND(αSMA average>38, CD31 = negative), 1, 0). 1 = positive, 0 = negative]. IAE-NearCYTE provides both pixel and cytometry features and was used to localize double and triple positivity on panel B (CK19/HNF1β/HNF4α/DAPI)-stained WSIs. Thresholds for fluorophore positivity were set manually and visual conformation was done after sorting to ensure specificity.
The statistical analysis methods to determine data sensitivity and significance were the t test, Mann-Whitney U test, and one-way analysis of variance (ANOVA) test all performed using Sigma Stat v. 11.0 (Aspire Software International, Ashburn, VA).
Comparison of Conventional Tissue Morphometry to Tissue-Tethered Cytometry.
Conventional image analysis of liver tissues (e.g., ImageJ, IAE-NearCYTE) has relied primarily on pixel-based determinations of percent tissue area occupied by a single analyte, channel, or stain assayed on individually captured microscopic fields (Supporting Figs. 1, 2). Examples include trichrome for fibrosis9 or absence of staining for fat.17 Pixel-based analyses are powerful, but unable to easily provide information about cell-specific physical characteristics (size, shape, location) or complex data from multiple analytes, or social interactions. Common open source software (e.g., ImageJ) is rich in functionality for routinely captured static images but does not easily accommodate WSI.
Cell-based image analysis (e.g., FARSIGHT and IAE-NearCYTE) is a higher-level image analysis approach based on grouping of similarly colored pixels into biologically meaningful structures, such as cells and/or parts thereof. Each nucleus can serve as the nidus for cell-associated nuclear and/or cytoplasmic analyte(s) (protein, DNA, mRNA) assays (Supporting Fig. 3). This enables identification of complex specific cell types based on Boolean logic relationships among multiple characteristics. For example, hepatocytes can be identified as cells with a relatively large (>23 μm2) round nucleus surrounded by β-catenin staining within a distance of 10 μm from the nucleus and negative CK19 staining (i.e., β-cateninfar/CK19-), whereas BECs are defined as smaller CK19+ cells. Cell-based approaches also enable the collection of data regarding location (X,Y) for 2D thin sections and Z planar addresses for thick sections, nuclear and cytoplasmic physical attributes (e.g., size, shape, and orientation characteristics), and nuclear and/or cytoplasmic analyte expression. Data collection can be followed by more sophisticated queries of social relationship.
Cell-based approaches also enable “tissue-tethered cytometry.” This refers to an ability to “virtually digest” the WSI. Each cell, regardless of size, shape, location, or phenotypic complexity, can be isolated and displayed in various formats. Examples include traditional and multidimensional scatterplots, whiskerplots, and signaling pathway schemes derived from covariance relationships (data not shown). Importantly, individual cells in either scatterplots or WSI are tethered to the exact same cell on the complementary display. The observer can easily transition between displays to assess the cell from informational perspectives.
To distinguish between the two approaches, 10 portal tracts and 10 perivenular ROIs were selected randomly from panel A (CK19/β-catenin/CD31/αSMA/DAPI)-stained high-resolution (40×) WSI images to determine the relative proportions of cell types in two separate livers (Supporting Table 1, Supporting Fig. 1A,B). As expected, αSMA+ cells were overrepresented and BECs were found only in portal/periportal ROIs compared to perivenular regions. FARSIGHT-generated data for hepatocytes, BEC, endothelial cell (EC), and smooth muscle cell (SMC) (Fig. 1A) sorted from one liver (total 20 ROIs) yielded 539/18,875 (2.86%) BECs; 9,153/18,875 (48.5%) hepatocytes; 1,093/18,875 (5.79%) EC; 669/18,875 (3.54%) SMC, and 7,421/18,875 (39%) unclassified cells (e.g., nerve cells, leukocytes, quiescent stellate cells, etc.). These results compared favorably with previously reported studies using tissue digestion and point-counting studies conducted on normal murine and human livers.18-22 Although difficult to compare, the current method is probably more sensitive and accurate because the WSIs were created with a high-resolution objective (Supporting Fig. 1B) and neither tissue disruption nor digestion, which can destroy and/or exclude cells, is needed. Because tissue-tethered cytometry enables harvesting of complex quantitative cell-specific data it was selected for all remaining analyses.
Normal Human Liver Tissue Cytometry.
Quantitative cytometric data generated from tissue-tethered cytometry on various cell populations from normal adult livers confirmed previous observations using more laborious techniques (Supporting Table 1 and Fig. 1). For example, hepatocyte nuclei are significantly larger (Fig. 1B) than all other liver cell types and perivenular hepatocyte nuclei are larger and more often binucleate (7.7 ± 1.8% versus 6.4 ± 1.0%; P < 0.001; Mann-Whitney t test, Fig. 1D) than periportal hepatocytes (Fig. 1B) (reviewed23). BEC nuclei lining large septal bile ducts (80-150 μm diameter) were also significantly larger than BEC nuclei lining small bile ducts from the same liver (<25 μm diameter; Fig. 1C), as previously shown in rodent whole liver digestion studies.24
The record of individual cell X-Y coordinates enabled a diagrammatic reconstruction of histological structure (Fig. 1D), which can be used for: (1) quick visual inspection quality control of cell identification and sorting characteristics; (2) social relationship between cell types; and (3) easy identification of specific cell types or rare events/cells within the context of tissue structure. For example, the number of nearest neighbors at predetermined radii from each hepatocyte (Fig. 1E left and right upper, yellow lines) can be easily calculated. Based on a training set optimal distance of 35 μm, nearest neighbor calculation showed that hepatocytes with larger nuclei (>100 μm2) showed fewer nearest neighbors (more widely separated) than hepatocytes with smaller nuclei (<100 μm2; Fig. 1E; 4.7 ± 2.1 versus 5.5 ± 2.2; P < 0.001; Mann-Whitney t test). Because hepatocyte nuclear and cytoplasmic sizes are directly proportional, more distant neighbors for larger nuclei are expected. Clustering of smaller cells, however, is not widely appreciated. This could be related either to polarization of nuclei to nearby edges of cells or small cells, or both. Studies are under way to further investigate this finding.
Reassured that software-generated data on WSI reproduced previously accepted and verified results for hepatocyte and BEC sizes and their distribution, we more closely examined CK19+ BEC using CK19/β-catclose/αSMA/CD31-stained WSI. Scatterplots generated from portal/periportal-based ROI consistently showed a population of CK19weak cells. Taking advantage of “tissue-tethered cytometry,” we queried the location of these CK19weak cells; a majority was located immediately adjacent to periportal hepatocytes. These characteristics correspond to accepted CH definitions,25, 26 which are thought to contain putative HPCs.
We next manually selected 10 portal/periportal ROIs, harvested cytologic characteristics of all software screened and human verified CK19weak cells that fulfilled location characteristics of CH cells. A total of 12 putative CH cells fulfilled these characteristics. These software-screened and human-selected CH cells or putative HPC characteristics were compared to otherwise typical mature BEC lining portal tract bile ducts and midzonal hepatocytes. CH cells showed significantly lower CK19 expression (CK19low; Fig. 2A) and identical β-catenin expression (β-cateninclose; Fig. 2B) compared to typical mature BEC lining portal tract bile ducts. Nuclear size, however, was intermediate between typical BEC and hepatocytes (Fig. 2C). CH cells also had a very high nuclear:cytoplasmic (N:C) ratio and a close relationship to CD31+ sinusoidal EC (Fig. 2D).
If the software-generated phenotypic characterization of CH cells is accurate, we should be able to use this technique and identify “rare event” CH cells in new image sets based on training set characteristics. Results from a representative experiment are shown in Fig. 2E-H. CK19low candidate cells were first gated (Fig. 2E) and then sorted according to their boundary profile (0 = rare touching neighbors; 0.07 = more touching neighbors similar to typical BEC lining bile ducts) and β-catenin intensity (Fig. 2F). Visual inspection of software-identified putative HPC in new tissue sections confirmed localization to CH (green arrows in Fig. 2G and high magnification in Fig. 2H). In contrast, high boundary = 0.07 CK19+ cells were located within otherwise typical bile ducts (red arrowheads in Fig. 2G). The same analysis of three separate livers using CK19low as the only discriminating criterion yielded a specificity of 0.72 ± 0.13 for software-selected CH cells when compared to human identification, which is CK19+ cells in periportal parenchyma adjacent to hepatocytes.25 When low boundary scores were combined with the CK19low criterion, the specificity for software-selected CH compared to human increased to 0.92 ± 0.02 (t test, P = 0.046).
We next examined CH cells acquired in multifocal planar wide-field images using panel A-stained (CK19/β-cat/CD31/αSMA/DAPI) 20-μm-thick sections to further examine CK19 expression in CH cells (Fig. 3). The multifocal imagery was created using a 100× objective lens scans on an AxioImager M1 microscope (Carl Zeiss, Gottingen, Germany) equipped with software control for autofocus and field stitching to create a seamless wide-field representation. The resultant image stack was processed by way of deconvolution methods using Zen software (Carl Zeiss, Munich, Germany) to render a 3D composite (Rotation in 3D can be viewed at: http://youtu.be/YGbyy9WoXz8). CH CK19low cells (asterisk in Fig. 3) showed a cuboidal shape and weaker CK19 expression (white arrowhead in lower left panel) at the cell border immediately adjacent to the hepatocyte (white arrowheads in Fig. 3). Serial virtual step sectioning is shown in Supporting Fig. 4 (Digital movie examples of the CH-Hepatocyte junction at 40× and 100× magnification can be seen at: (40×) http://youtu.be/JWyz4h9IUvk; (100×) http://youtu.be/ww99LiX0N0E).
Transcription Factor Expression in Liver Epithelial Cells.
Hepatocyte-associated nuclear transcription factor HNF427 and biliary associated transcription factor HNF1β28, 29 are essential for development and phenotypic maintenance of hepatocytes and mature BEC, respectively. We next determined whether any hybrid or bipotential (HNF1β+/HNF4α+) cells existed in bile ducts and periportal regions of normal adult human livers. Portal/periportal ROIs from five normal human livers were subjected to analysis and the results displayed on multiparameter scatterplots showing nuclear size and signal intensity for CK19, HNF1β, and HNF4α (Fig. 4A). Tissue-tethered cytometry showed that most CK19+/HNF1βstrong/HNF4α-cells with small nuclei mapped back to otherwise typical BEC lining portal tract bile ducts (cells “d” in Fig. 4B), as expected. Most CK19-/HNF1β-/HNF4αstrong cells with large nuclei localized to otherwise typical mature hepatocytes (cells “e” in Fig. 4B), as expected.
Two distinct populations of HNF4α+/HNF1β+ cells were identifiable in the X = HNF4α, Y = HNF1β scatterplot in Fig. 4A: (1) HNF1βstrong/HNF4αweak with a small nucleus (BEC-type) that showed phenotypic characteristics of BEC on routine light microscopy and (2) HNF1βweak/HNF4αstrong with a larger nucleus (hepatocyte-type) that showed phenotypic characteristics of hepatocytes on routine light microscopy. Transcription factor localization was verified using single immunoperoxidase labeling of frozen sections of normal human livers (Supporting Fig. 5). The observation that the quantitatively dominant transcription factor controlled the routine light microscopic phenotype of cells substantiated the validity of this approach: HNF4α-dominant cells appeared as hepatocytes and HNF1β-dominant cells appeared as BEC.
Tissue tethering was used to localize the various HNF1β+/HNF4α+ populations. CK19+/HNF1βhigh/HNF4αlow BEC-type cells (white cells “a” in Fig. 4B) localized to CH, but were rare. HNF1βlow/HNF4αhigh hepatocyte-type cells (yellow cells “b and c” in Fig. 4B) localized to the interface zone of normal livers and two different morphological subtypes of CK19-/HNF1βweak/HNF4αstrong hepatocyte type cells were observed among periportal hepatocytes: one showed an oval-shaped intermediate-sized nucleus (“b” in Fig. 4B); the other contained a large round typical hepatocyte nucleus (“c” in Fig. 4B). Among the various cell types at the interface zone, the rarest cell was CK19+/HNF1β-/HNF4α+ cell (Navigation of the WSI using toggle switches to turn off/on various analytes can be viewed at: http://youtu.be/WldIOlZfQVM).
Various HNF1β+/HNF4α+ cell types showed phenotypic and analyte expression characteristics intermediate between otherwise typical BEC lining portal bile ducts and hepatocytes (Fig. 4C). For example, HNF1βhigh/HNF4αlow BEC-type cells showed lower CK19 expression than mature BECs, as expected. CK19-/HNF1βweak/HNF4αstrong hepatocyte-type cells expressed significantly higher HNF1β than mature hepatocytes and significantly higher HNF4α expression than mature BEC (Fig. 4C). Cells most strongly positive for both HNF1β+ and HNF4α+ often showed hepatocyte features with an oval nucleus, but lacked CK19 staining and were close to CH cells or terminal bile ducts, but a direct connection could not always be confirmed in thick sections (Fig. 4D). A graphic reconstruction summarizing our results is shown in Fig. 5.
HNF1β28, 29 and HNF4α (reviewed27) are responsible for development and maintenance of mature BEC and hepatocyte phenotypes, respectively. In agreement with previous single marker studies, HNF1β can be expressed by periportal hepatocytes28, 29 and HNF4α can be expressed by occasional BEC.30 A novel workflow with multiplex labeling, however, enabled us to show that the quantitatively dominant transcription most strongly influenced the routine histopathologic appearance of the cells. Indeed, multiplex labeling, WSI creation, and automated image analysis enabled us to identify and characterize diverse epithelial populations that show transitional cytometric characteristics and analyte expression, including coexpression of HNF1β and HNF4α: (1) CK19+/HNF1βhigh/HNF4αlow BEC-type cells; (2) CK19weak/HNF1βhigh/HNF4αlow BEC/CH-type cells; (3) CK19-/HNF1βweak/HNF4αstrong hepatocyte-type cells with an oval nucleus; and (4) CK19-/HNF1βweak/HNF4αstrong hepatocyte-type cells with a large round nucleus (identical to mature hepatocytes). This extensive, but difficult to visualize with conventional histology, population of cells with phenotypic and biomarker (including transcription factor) expressions of a hybrid nature between hepatocytes and biliary epithelial cells are positioned over a relatively broad area from small portal-based bile ducts to otherwise typical periportal hepatocytes.
Previous studies in rodents show that progenitor cells (i.e., “oval” cells) arise from BEC to provide hepatocytes when regeneration of the liver needs to occur under conditions in which hepatocyte proliferation is inhibited.31 Similar conclusions have been reached in cirrhotic human liver tissue samples where hepatocytes are thought to be derived by terminal bile ducts/CH harboring putative progenitor cells.32-34 Conversely, periportal hepatocytes can give rise to BEC when BECs are incapable of regenerating.35 The phenotypic transitions between hepatic epithelial cells, or hybrid hepatobiliary cells, which are experimentally well-characterized in rodent liver biology, also occur in human liver during ductular reactions1-3, 5, 26, 36-38 and when otherwise typical hepatocytes appear intermixed among BEC in portal tract bile ducts.4
The location of these hybrid transitional cells in portal bile ducts, CH, and immediate periportal hepatocytes of normal liver coincides with the niche of potential stem cells described in rodent liver.39 Like their studies, we cannot exclude the possibility that bipotential periportal HNF1β+/HNF4α+ hepatocytes are derived from CH cells39 given their often close proximity (Fig. 4). However, given the preexistence of such hybrid cells in the normal liver and the abundance of hybrid epithelial phenotypes in diseased human livers, we postulate that the former expand to give rise to the latter when the liver microenvironment is disrupted. It is likely that latent hybrid cells play an important role during extreme regenerative situations when a need exists for facultative stem cells to rescue the regenerative failure of one or the other liver epithelial cell compartments. As such, they are likely to play an important role in understanding the pathogenesis of human liver disease.
Use of digital imaging and computational image analysis in liver pathology, to date, has been largely limited to capture single microscopic fields followed by pixel-based determination of fat quantities17 and fibrosis areas.9 This approach, however, has a low value proposition: it is impractical and time-consuming for marginal improvements in information extraction.9 Multiplex labeling techniques to identify complex cell phenotypes that also express target proteins of interest are also impractical, expensive, and inconvenient with traditional imaging methods40 because of: (1) reliance on traditional fluorophores with inherent drawbacks; (2) expensive and inconvenient fluorescent microscopes; and (3) dependence on tedious image capturing steps and subjective interpretation. Spatially overlapping signals derived from brightfield chromogenic multiplexing are harder to separate and quantify unless multispectral imaging is used.41
In this study we introduced a workflow that combines high-resolution digital imaging, robotics, computing, nanotechnology, and software “toolkits” that enable pathologists and researchers to extract more biologically significant cellular information from tissue samples than is currently achieved by human analysis alone. “Tissue cytometry” was first described nearly 10 years ago,42 but was only recently applied for tissue analysis, spurred by convergent advances in computer processing, imaging equipment, and staining materials/protocols that made the overall process more practical (reviewed7, 43).
Particular strengths of tissue-tethered cytometry include the ability to: (1) collect detailed quantitative physical (e.g., nuclear size, location, etc.) and analyte (protein, DNA, RNA) data on thousands of cells; (2) “virtually digest” the tissue while retaining structural context; and (3) represent the data in a variety of formats (e.g., multidimensional scatterplots, diagrams, signaling pathways, heat maps, and clustering diagrams) that might be more easy to interpret than the image itself or in conjunction with the image; (4) map social relationships among cells; and (5) identify or screen for rare event candidates followed by human quality assurance review (e.g., putative HPC; Figs. 2, 4).
The workflow technology described herein is being routinely applied for basic science and clinical trial research purposes,7, 10-16 but limitations exist for routine clinical implementation: (1) the serial, and time-consuming, nature of the staining; (2) uneven tissue staining; (3) long scan time required for creation of multiplex digital images; (4) the large amount of data generated; and (5) the need to train pathologists. Automatic nucleus segmentation is still a challenge when cell nuclei overlap in thick or inflamed tissue sections or when DAPI signal intensity varies across hepatocytes and stromal cells. Although common methods for nuclear segmentation histogram-based, clustering-based, and entropy-based algorithms44 are often used, recent higher specificity model-based computational methods are becoming practical because of improved computational power. Software interfaces to WSI are not yet platform agnostic, as such hybrid methods using multiple tools still require specialty informatics techniques to be used to repackage and import imagery data into the various programs.44
We thank the staff of the Research Histology Service from the Thomas E. Starzl Transplantation Institute, especially Lisa Chedwick, and the Roysam Laboratory, and Dr. William M. Lee of the Department of Medicine, Abramson Cancer Center, University of Pennsylvania. We also thank Dr. Stephen Strom from Karolinska Institutet and Hospital.