Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2024 Jan 1.
Published in final edited form as: J Cell Physiol. 2022 Dec 11;238(1):274–284. doi: 10.1002/jcp.30927

Topography of pleural epithelial structure enabled by en face isolation and machine learning

Betty S Liu 1, Cristian D Valenzuela 1, Katherine L Mentzer 2, Willi L Wagner 3, Hassan A Khalil 1, Zi Chen 1, Maximilian Ackermann 4, Steven J Mentzer 1,*
PMCID: PMC9845181  NIHMSID: NIHMS1852293  PMID: 36502471

Abstract

Pleural epithelial adaptations to mechanical stress are relevant to both normal lung function and parenchymal lung diseases. Assessing regional differences in mechanical stress, however, has been complicated by the nonlinear stress-strain properties of the lung and the large displacements with ventilation. Moreover, there is no reliable method of isolating pleural epithelium for structural studies. To define the topographic variation in pleural structure, we developed a method of en face harvest of murine pleural epithelium. Silver-stain was used to highlight cell borders and facilitate imaging with light microscopy. Machine learning and watershed segmentation were used to define the cell area and cell perimeter of the isolated pleural epithelial cells. In the deflated lung at residual volume, the pleural epithelial cells were significantly larger in the apex (624±247 um2) than in basilar regions of the lung (471±119 um2) (p<.001). The distortion of apical epithelial cells was consistent with a vertical gradient of pleural pressures. To assess epithelial changes with inflation, the pleura was studied at total lung capacity. The average epithelial cell area increased 57% and the average perimeter increased 27% between residual volume and total lung capacity. The increase in lung volume was less than half the percent change predicted by uniform or isotropic expansion of the lung. We conclude that the structured analysis of pleural epithelial cells complements studies of pulmonary microstructure and provides useful insights into the regional distribution of mechanical stresses in the lung.

Introduction

Epithelial layers provide crucial barrier function by lining the surfaces of organs throughout the body. In normal circumstances, epithelial tissues are classified by the shape and function of their cells (Davidson, 2012); however, epithelial structure is notably difficult to image in situ. The epithelium is a single layer of densely packed cells contoured to the organ surface. Epithelial structures become even more complex when responding to developmental growth or surface injury.

Single-cell segmentation of epithelium, the process that identifies and outlines individual cells in an image, is one of the most difficult tasks in biologic image analysis (Akram, Kannala, Eklund, & Heikkilä, 2016; Van Valen et al., 2016). Conventional image segmentation uses watershed algorithms (Jo, Han, Kim, Lee, & Yang, 2021; Kandel et al., 2021; Wang et al., 2019). Watershed algorithms identify gray-level inflection as a boundary even if the change is subtle. The limitations of watershed approaches are apparent in analyzing complex or noisy images. Noisy images produce over-segmentation or irregular boundaries. The addition of markers to control over-segmentation has complicated the process while producing marginal benefits (Litjens et al., 2017; Liu, Wolterink, Brune, & Veldhuis, 2021)

Recent attempts to address the problem of single-cell segmentation have applied transfer learning algorithms that 'pre-train' the neural network on a more general dataset (Chan, Samala, Hadjiiski, & Zhou, 2020; Deo, 2015). For example, the neural network may be trained on general features of epithelial cells before being 'fine-tuned' to more specialized features of specific epithelial tissues. This approach has been useful because of the limited availability of en face epithelial images. The transfer learning algorithm can be trained to learn and predict outlines by using both intensity and shape features. A basic strategy of transfer learning is to classify individual pixels into intracellular, boundary, and background categories. But even transfer learning algorithms are challenged by images of epithelial cells—that is, complex images with densely packed neighboring cells.

In this report, pleural epithelial cell boundaries were defined by silver-stain. Silver-stain provided a discrete high-contrast signal that could be applied in situ. After en face isolation, the stained pleural epithelium was imaged by light microscopy and analyzed by both watershed and transfer learning neural network (TLNN) algorithms. The data provided useful insights into pleural stress distributions reflected in the structural adaptations of pleural epithelium.

Methods

Animals.

Male and female mice, 8- to 10-week-old wild-type C57BL/6 ( Jackson Laboratory, Bar Harbor, ME USA), were anesthetized before euthanasia. The care and nurturing of the animals was consistent with guidelines of the American Association for Accreditation of Laboratory Animal Care (Bethesda, MD) and approved by the Brigham & Women’s Institutional Animal Care and Use Committee.

In situ silver-staining.

Freshly harvested lung was gently rinsed with phosphate-buffered saline and incubated in 5% D-glucose (Gibco Laboratories, Grand Island, NY, USA) for 3 minutes. The lung surface was treated with 0.4% Silver Nitrate (Sigma-Aldrich, Saint Louis, MO, USA) for 30 seconds, submerged briefly in 5% D-glucose solution prior to before exposure to 254 nm UV light (Thermoscientific, Waltham, MA, USA) for 60 seconds.

Pectin.

The citrus pectins used in this study were obtained from a commercial source (Cargill, Minneapolis, MN, USA). The characterization of the high methoxyl citrus pectin has been detailed elsewhere (Zheng et al., 2021). Briefly, the proportion of galacturonic acid residues in the methyl ester form determined the degree of methoxylation. The high-methoxyl pectins (HMP) were defined as a greater than 70% degree of methoxylation. The pectin powder was stored in low humidity at 25°C. To create the pectin film used in isolation, the pectin powder was dissolved at 25°C by a step-wise increase in added water to avoid undissolved powder (Panchev, Slavov, Nikolova, & Kovacheva, 2010). The complete dissolution of the pectin was achieved by a high-shear 10,000rpm rotor-stator mixer (L5M-A, Silverson, East Longmeadow, MA USA). The dissolved pectin was poured into a standard mold for further studies. The pectin films were allowed to equilibrate to ambient 40% relative humidity.

En face harvest.

After general anesthesia, tracheal intubation and euthanasia by exsanguination, the lungs were inflated with 3 consecutive “total lung capacity” perturbations, involving a 3 second ramp to 30 cmH2O plateau pressures. A midline thoracoabdominal incision exposed the lung block. A median sternotomy facilitated exposure of bilateral lungs, heart, and trachea. In sequence, the left atrium, right ventricle, and inferior vena cava were incised. A 22G olive-tipped cannula was inserted through the right ventricle into the pulmonary artery, and the lungs underwent a vascular flush of 20 cc of phosphate-buffered saline. For maximally inflated (TLC) lungs, a 1-ml syringe was attached to the angiocatheter and injected with 3% agarose, low gelling temperature (Sigma-Aldrich, Saint Louis, MO, USA), at 42°C in the volume of average TLC. The tissue was allowed to cool until solidified prior to staining and pleural epithelial isolation.

The freshly harvested lung was placed on a gel-phase high-methoxyl citrus pectin film. After 20 seconds of development at room temperature, the lung was peeled off at an angle of 120 degrees with a steady rate of 2mm per second. A thin film of PBS was maintained on the specimen to prevent dehydration. After isolation, connective tissue was often identified on the parenchymal surface. This tissue was removed (“thinned”) using a soft bristle artists brush. The pectin was then mounted, tissue-side down, on a poly-L-Lysine (Sigma Aldrich, Saint Louis, MO, USA) coated glass slide. The slide was then submerged in phosphate-buffered saline for 60 minutes on a shaker to allow pectin to dissolve. The slides were then washed 3 times, fixed with −20 °C acetone, and mounted with DAPI-containing medium (Vector Laboratories, Burlingame, CA, USA).

Fluorescence immunohistochemistry.

The murine lung samples were fixed in 4% paraformaldehyde in PBS at 4°C for 24 h. After 24 h, the specimens were submerged in O.C.T. compound and frozen in a mixture of acetone and dry ice. The O.C.T. blocks were kept at −80°C for 24 h prior to cryosectioning. Cryostat sections were obtained from lung specimens embedded in O.C.T. compound, and snap frozen. After warming the slide to 27°C, the sections emmersed in acetone at 4°C. The slides were washed with PBS buffer and blocked with 10% goat serum in PBS for 30 min. The slides were treated with primary and secondary antibody. The slides were incubated with each antibody for 1 hour at 27°C, washed three times, counterstained with Hoechst 33342 (Sigma-Aldrich, St. Louis, MO, USA) for 15 min and mounted using VectaShield mounting media (Vector Laboratories, Inc., Burlingame, CA, USA).

Antibodies.

The primary antibodies used in fluorescence immunohistochemistry were obtained from commercial sources: anti-WT-1 (ab15249)(Abcam, Cambridge, UK) rabbit polyclonal primary antibody; anti-E-cadherin (13-1900)(Thermo Fisher Scientific, Waltham USA) rat monoclonal antibody; anti-Vimentin (ab24525)(Abcam) chicken polyclonal primary antibody; and anti-Mesothelin (250519)(Abbiotec, Escondido, CA USA) rabbit polyclonal antibody. The cross-absorbed detection (secondary) antibodies were also obtained from commercial sources: goat anti-rabbit Texas-red (T2767)(Thermo Fisher); goat anti-chicken FITC (PA1-28794)(Thermo Fisher); and goat anti-rat Texas-red (1010-07)(Southern Biotechnology, Birmingham AL, USA). Primary antibodies were used at a 5-fold saturating concentration. Secondary antibodies were titered to minimize nonspecific background staining; negative controls were stained with detection antibody alone.

Grid confocal microscopy.

As previously described (Lee et al., 2009), the images were acquired with the optical system equipped with an Optigrid (Qioptiq, Fairport, NY, USA) controlled by Volocity 5.0 (Perkins-Elmer, Waltham, MA, USA) software. Image data were processed with a Dell Precision 390 Workstation with dual Xeon processors, 16Gb RAM and a NVIDIA Quadro FX 3450 graphics card (NVIDIA, Santa Clara, CA, USA). After distance calibration, including Z dimension, image stacks were acquired using the MetaMorph 7.10. The image stacks were processed with standard filters including the Best Focus application.

Lung pleural surfaces.

Sampling of the lung surface was informed by the airway architecture. The surface regions were defined by the corresponding bronchial segments (Wallau, Schmitz, & Perry, 2000): Apex, cranial bronchi E; Base, terminal bronchi Tl+TM; Medial, medial bronchi M1-3; Lateral, lateral bronchi L1-3, Dorsal, dorsal bronchi D1-3; Ventral, ventral bronchi V1-3; Cardiac, entire cardiac lobe.

Live/Dead Assay.

A commercial Live/Dead Assay Kit (L3224) (ThermoFisher Scientific, Waltham, MA USA) was used. After aliquoting, a final 2 uM calcein AM and 4 uM EthD-1 working solution was used. The tissue was incubated for 30 minutes on an orbital shaker for 30 min at 25 °C. The tissue was washed with PBS and mounted on a glass slide in PBS. The tissue was imaged on a TE-2000 imaging system with differential fluorescent image acquisition (calcein, excitation 485/20 emission 530/25 and ethidium excitation 530/25 emission 645/40). In addition to their viability, the isolated pleural epithelium expressed a mesothelial phenotype (WT-1+, mesothelin+, E-cadherin+) (Valenzuela et al., 2017).

Pleural epithelial cell image acquisition.

The epithelial cells were imaged using a Nikon Eclipse TE2000 inverted epifluorescence microscope using Nikon Plan Apo 10 and Plan Fluor 20 objectives. An X-Cite (EXFO, Vanier, Canada) 120-W metal halide light source and a liquid light guide were used to illuminate the en face samples. Excitation and emission filters (Chroma, Rockingham, VT) in separate LEP motorized filter wheels were controlled by a MAC5000 controller (Ludl, Hawthorne, NY) and MetaMorph 7.10 software (Molecular Devices, Downingtown, PA USA). The 14-bit fluorescent images were digitally recorded with an electron multiplier CCD (EMCCD) camera (C9100-02, Hamamatsu, Japan). Epithelial image stacks, systematically imaging the epithelial sample, were reconstructed into an integrated montage.

Watershed algorithm.

The algoithm used commands from the MetaMorph 7.10 (Molecular Devices) image analysis software. The original silver-stained epithelial image was processed using a series of filters that systematically flatten background regions and reduce background noise. The Flatten background command was used to correct for gradated or uneven background regions of the original silver-stained image. The Invert command was used to invert bright/dark intensities within their full range. The Binarize command was used to create a binary (black and white) image mask. The Remove single pixel command reduced background noise. The commands Close-open and Open-close were used to sequentially smooth the dark (close) and light (open) objects using a 5 pixel circular shape filter. The Dilate command expanded bright objects using local MAX intensity prior to applying the Invert and Dilate commands.The Dilate command was used to connect objects which were discontinuous or to fill "holes" in objects. The images prior to and after the Dilate command were compared by Logical XOR; that is, an operation that determined if either of the pixels in the first two binary planes were "turned on" (white). TheWatershed command was used with the Dilate image as marker and Flattened image as source image to obtain single pixel outline. At the last step, the contrast was inverted to produce a dark boundary.

Transfer learning neural network (TLNN).

Several transfer learning neural network algorithms were compared to our watershed method of segmentation. One of the neural network algorithms was widely available as a commercial site (www.biodock.com). BioDock provided adequate boundary segmentation and convenient image management. The silver-stained images, acquired with light microscopy, were uploaded into the neural network platforms. The analyzed epithelial images were processed with applied autoencoders; that is, custom neural network architectures consisting of a trained encoder that created a low-dimension latent space representation of the input image and a decoder designed to transform the data back from the latent space representation to the original image. Because of the limited number of available pleural epithelial images, our pre-training sets necessarily included an expanded corpus of general medical images. The model was then fine-tuned with an optimized subset of application-specific labeled images relevant to creating segmented masks corresponding to cell boundaries.

Statistical analysis.

The statistical analysis was based on measurements in at least three different samples. The unpaired Student’s t test for samples of unequal variances was used to calculate statistical significance. The data was expressed as mean ± one standard deviation. The significance level for the sample distribution was defined as p<.01.

Results

Staining and isolating pleural epithelium.

In situ silver-staining of the lung surface minimized processing artifact while maximizing boundary contrast. To facilitate the examination of sliver-stained samples, we developed an en face epithelial isolation technique (Fig. 1A-D). The biopolymer called pectin—previously shown to tightly adhere to the pleural glycocalyx (Servais et al., 2018)—was gently applied to the pleural surface. After a 5 second development time a steady peel force applied to the pectin resulted in separation of the tissue subjacent to the pleural epithelium. The technique produced large continuous sheets of pleural epithelium with rare discontinuities (Fig. 1E, arrows). Skillful application of the technique produced mostly viable cells assessed by a fluorescence live/dead (calcein/ethidium) assay (Fig. 1F-G). Light and fluorescent microscopic examination of isolated pleural epithelium demonstrated silver-stained boundaries (Fig. 2). The staining was nearly complete but punctuated by occastional gaps in the epithelial cell boundaries (Fig 2C,D). Confocal microscopy optical sections demonstrated en face tissue thickness less than 25 um (Fig. 3A) and confirmed the expression of pleural epithelial antigens WT-1 (Walker et al., 1994), E-cadherin, vimentin and mesothelin (Hassan et al., 2010)(Fig. 3B-E).

Figure 1.

Figure 1.

Open in a new tab

Isolation of the pleural epithelium. A) A freshly harvested murine lung, inflated to a desired volume, was placed on a gel-phase translucent pectin film to maximize contact area. B) After a brief interval at room temperature, the lung was peeled off the pectin at a peel angle of 120 degrees and an approximate rate of 2 mm/s. C) Pectolyase was added to the tissue to dissolve the pectin film. D) The pleural epithelium was mounted to glass slide for imaging. E) Scanning electron microscopy of a representative sample of isolated pleural epithelium prior to mounting. Note the rare discontinuities in the pleura (arrows). F) Live/dead assay with calcein (green) and ethidium (red) fluorescence. G) The majority of the pleural cells were calcein positive and ethidium negative indicating cell viability (p<.001). N=3 mice; mean=633 cells/mouse.

Figure 2.

Figure 2.

Open in a new tab

Silver-stained pleural epithelium isolated en face. A) Pleural epithelium image obtained at 60x magnification and montaged by a MetaMorph (Molecular Devices) stitching algorithm to obtain a continuous sheet. B) The autofluorescence of the silver-stained viewed with 590 nm long pass filter after DAPI counterstain (Vectorshield, Vector Laboratories, Newark, CA USA). C) Portion of the visceral pleura with region of higher magnification (D) demonstrating discrete discontinuities in the cell boundaries (arrow). E) Silver-staining of other regions of the pleura demonstrate contiguous borders even at higher magnification (F). Bar = 50 um.

Figure 3.

Figure 3.

Open in a new tab

Structured illumination confocal microscopy of the isolated pleural epithelium. The isolated en face tissues were mounted on an optical surface without tissue frames or pinning. The tissue was stained with specific primary and detection secondary antibodies (see Methods). After calibration of the Z-position linear encoder (Ludl, Hawthorne, New York), typically 20 optical sections were acquired as Optigrid confocal images, thresholded and processed with standard Metamorph 7.10 filters. A) The 3D image stacks, analyzed using the Kymograph function, consistently demonstrating en face pleural tissue thickness less than 25 um. The Best Focus function was applied to produce representative images of pleural epithelial tissue antigen expression: B) WT-1 (Conner, Cibas, Hornick, & Qian, 2014), C) E-cadherin, D) vimentin and E) mesothelin (Kachali, Eltoum, Horton, & Chhieng, 2006). Negative controls, consisting of secondary antibody alone, demonstrated only scant nonspecific staining.

Watershed algorithms.

The pleural epithelial images were studied using traditional watershed algorithms; that is, segmentation methods that do not require prior knowledge of the images. Our watershed algorithms applied filters that sharpened boundary information and removed non-boundary noise. The application of our algorithm was effective with silver-stained images because of the high contrast and limited noise (Fig. 4). The noisier the image—a common scenario in fluorescence microscopic images—the more likely the algorithm produced over-segmentation and irregular boundaries.

Figure 4.

Figure 4.

Open in a new tab

Defining pleural cell boundaries using a watershed algorithm. The algoithm used commands from the MetaMorph 7.10 (Molecular Devices) image analysis software. A) The Flatten background command was used to correct for gradated or uneven background regions of the original silver-stained image. B) The Invert command was used to invert bright/dark intensities within their full range. C). The Binarize command was used to create a binary (black and white) image mask. D) The Remove single pixel command reduced background noise. The commands Close-open and Open-close were used to sequentially smooth the dark (close) and light (open) objects using a 5 pixel circular shape filter. F) The Dilate command expanded bright objects using local MAX intensity prior to applying the Invert and Dilate commands (G).The Dilate command was used to connect objects which were discontinuous or to fill "holes" in objects. The images prior to and after the Dilate command were compared by Logical XOR; that is, an operation that determined if either of the pixels in the first two binary planes were "turned on" (white). H) Watershed command was used with the Dilate image as marker and Flattened image as source image to obtain single pixel outline. I) Finally, the contrast was inverted to produce a dark boundary.

Transfer learning neural networks.

Transfer learning neural networks (TLNN) provide a general approach to image segmentation and object detection. Here, the TLNN algorithm was pre-trained on a more general dataset to learn both intensity and shape features. With limited specific training, TLNN identified more than 75% of the epithelial cells but occasionally failed with neighboring cells with incomplete boundaries, background noise and overstaining (Fig. 5, Raw data). Early errors in segmentation were addressed by parameter tuning in our validation set (Fig 5, TLNN version 1) with substantial performance improvement in our test set (Fig 5, TLNN version 2). After parameter tuning, there was no statistical difference between TLNN and watershed algorithms in our analysis of cell number, cell area, cell perimeter and shape factor (p>.05).

Figure 5.

Figure 5.

Open in a new tab

Comparison of TLNN and watershed segmentation. Raw data reflecting challenging image patterns reflecting air bubbles (A), incomplete border (B), background noise (C) and over-staining (D). The initial validation set (TLNN version 1) was inferior to watershed segmentation in both cell boundary and artifact identification. With additional parameter tuning (TLNN version 2), the test set matched watershed segmentation with nearly-identical measures of cell number, cell area and cell perimeter. Cell area extending to the margins of the image were excluded from the analysis. N=1000 cells compared. In TLNN version 1, the cell border verticies are denoted with a blue dot.

Geometry of epithelial structure at resting volumes.

To evaluate the structural geometry of the lung surface, pleural epithelium was systematically sampled over the surface of deflated lungs. Each planar sample was oriented with respect to apicobasal and mediolateral axes and the cell centroid was assigned an X-Y coordinate. Over 1000 pleural epithelial cells in N=40 lungs were analyzed by TLNN. Morphometric analysis showed an increase in epithelial cell area (Fig. 6A, p<.05) and cell perimeter (Fig. 6B, p<.05) at the apex of the lung with a significant increase in cell area was apparent within 3 mm of the lung apex (Fig 6, asterisk).

Figure 6.

Figure 6.

Open in a new tab

Topography of the static pleural epithelium. In deflated lungs, the pleural epithelium was selectively sampled on 7 regions of the lung surface. The epithelial cell area (A) and cell perimeter (B) was significantly greater at the apex of the lung than at the base (p<.01, asterisk). The surface regions were defined by the corresponding bronchial segments (Wallau et al., 2000): The inset is a modified diagram of bronchial segments: A, apex; B, base; C, cardiac; D, dorsal; L, lateral and V, ventral. Data shown N=12 mice.

Geometry of epithelial structure with inflation.

To address the long-standing problem of how pleural epithelium adapts to dynamic changes in lung volume, we studied pleural epithelium as derived from mid-portion of the lungs at residual volume (maximally deflated) and total lung capacity (maximally inflated). Microscopic examination of the pleural epithelium demonstrated an epithelial cell surface area that was larger at TLC (Fig. 7A) than RV (Fig 7B). Using our TLNN segmentation strategy, the change in average cell surface was 56.7% (Fig. 7C,D). The change in average cell perimeter was 26.8% (Fig. 7E,F). Of note, the lung volume change from RV to TLC was an increase of 207%.

Figure 7.

Figure 7.

Open in a new tab

Topography of the inflated pleural epithelium. Pleural epithelial cell samples were studied at various distances from the apex at total lung capacity (TLC) (A) and at residual volume (RV) (B). Black bar = 50 um. C) Scattergrams demonstrate cell area at TLC (black dots) and RV (gray dots). D) Cell area is summarized by 95% confidence bands reflecting TLC (black bands) and RV (gray bands). E) Scattergrams demonstrate cell perimeter at TLC (black dots) and RV (gray dots). F) Cell perimeter is summarized by 95% confidence bands reflecting TLC (black bands) and RV (gray bands).

Discussion

In this report, we defined the topography of pleural epithelial structure in the mouse lung. The pleural epithelial cells were 1) labeled in situ, 2) isolated as an en face surface, 3) imaged with light microscopy and 4) analyzed using watershed and TLNN algorithms. This approach documented the static and dynamic topographic variation in pleural epithelial cell size and cell perimeter. We conclude that the structural analysis of pleural epithelial cells complements studies of pulmonary microstructure and provides useful insights into the regional distribution of mechanical stresses in the lung.

Long-recognized as influencing both normal and pathologic function (Macklem, 1978; Mead, 1961), regional differences in lung mechanical stresses have been difficult to assess because of the nonlinear stress-strain properties of the lung parenchyma and the large displacements with ventilation (West, 1977). Classic experiments, including snap-freezing intact animals (Glazier, Hughes, Maloney, & West, 1967), have focused on distortion of subpleural alveoli as a reflection of tissue strain. Here, we used the geometry of pleural epithelial structure to map the regional differences in stress. In adult mice, the apical portion of the lung had significantly greater cell area and perimeter than other regions of the lung. The inference from these observations is that the apex is exposed to significantly larger expanding stresses than other regions of the lung.

A conventional understanding of cellular physiology suggests that cell size is not fixed, but responds to external factors. Structural and metabolic plasticity is important for adaptation to changing environments (Miettinen, Caldez, Kaldis, & Björklund, 2017). In particular, the pleura is an epithelial surface adapted to dramatic changes in lung volume and pleural pressures. Assuming that for normal changes in lung volume, the lungs expand uniformly and isotropically, we predicted that the percentage expansion of area on the surface of the lung would equal two-thirds of the percentage change in lung volume. The 207% increase in lung volume observed experimentally was associated with a 56.7% increase in epithelial surface area. Thus, the change in surface area was significantly less than the 136% change predicted by isotropic expansion. This finding updates prior surface area estimates using optical markers (Lehr, Butler, Westerman, Zatz, & Drazen, 1985).

En face harvest facilitated the morphometry of the pleural epithelium. Previous hautchen (“thin skin”’) en face monolayer preparation techniques have typically focused on the isolation of endothelial cells (Fornas & Fortea, 1987; Hirsch, Martino, Orr, White, & Chisolm, 1980) or visceral mesothelial cells (Raftery, 1979; Riese, Freudenberg, & Haas, 1978). These methods have involved surgical microdissection, enzymatic treatments as well as tissue fixation to frames and slides. With few exceptions, the technical limitations of these approaches have included unpredictable cell loss, incomplete stripping, disruption of the monolayer, and wrinkling of the specimen (Hirsch et al., 1980). These complications have precluded the effective analysis of epithelial structure. In contrast to these approaches, the application of the pectin bioadhesive described here enabled the harvest of broad regions of the intact pleural epithelium. The harvest method uniquely preserved the spatial distribution and the structure of pleural cells.

Silver-stain is most frequently used as a nonspecific protein stain in polyacrylamide gels (Oakley, Kirsch, & Morris, 1980) or tissue stain in histology (Newman & Jasani, 1998). Silver-stain also has the remarkable, and as yet unexplained, ability to define endothelial cell boundaries (Hirata, Baluk, Fujiwara, & McDonald, 1995). We extended the application of silver-staining to the pleural epithelium. The silver-stain profile was remarkably similar to the labeling observed with endothelial cells (Hirata et al., 1995; Zand, Underwood, Nunnari, Majno, & Joris, 1982). The primary advantages of silver-stain was in situ labeling prior to en face harvest. Further, the silver-stain did not photobleach—a potential limitation of fluorescent dyes. Stable signal isolation was particularly valuable during the development of our image analysis algorithms. The disadvantage of silver-stain was the occasional unexplained gaps and blurry boundaries. An explanation for these observations will likely depend upon an improved understanding of the mechanisms regulating silver deposition.

The principal benefit of machine learning applied to the pleural epithelium was time efficiency. Our supervised learning approach was restricted to cell identification in normal adult pleural epithelium; that is, the efficient identification of normal cell boundaries. This approach allowed us to rapidly analyze detailed morphometry of hundreds of cells in the en face samples. In addition to defining normal pleural epithelium, the present work provides a foundation for machine learning applied to more complex epithelial structure. Epithelial tissues can exhibit significant plasticity due to cell division, extrusion and intercalation (Guillot & Lecuit, 2013). Further, these tissues can demonstrate dramatic global changes in epithelial structure related to collective cell migration (Spatarelu et al., 2019), cellular unjamming transition (Park, Atia, Mitchel, Fredberg, & Butler, 2016) and epithelial-mesenchymal transition (Levayer & Lecuit, 2008). In the future, we anticipate that unsupervised learning with convolutional neural network deep learning algorithms will significantly contribute to our understanding of these processes.

Acknowledgments

Supported in part by NIH Grant HL134229, CA009535, HL and the German Research Foundation (SFB1066).

Abbreviations:

RV

residual volume

SEM

scanning electron microscopy

TLC

total lung capacity

TLNN

transfer learning neural network

References

  1. Akram SU, Kannala J, Eklund L, & Heikkilä J (2016). Cell segmentation proposal network for microscopy image analysis. In Deep Learning and Data Labeling for Medical Applications (pp. 21–29). Athens: Springer International Publishing. [Google Scholar]
  2. Chan HP, Samala RK, Hadjiiski LM, & Zhou C (2020). Deep learning in medical image analysis. Adv. Exp. Med. Biol, 1213, 3–21. doi: 10.1007/978-3-030-33128-3_1 [DOI] [PMC free article] [PubMed] [Google Scholar]
  3. Conner JR, Cibas ES, Hornick JL, & Qian XH (2014). Wilms Tumor 1/Cytokeratin Dual-Color Immunostaining Reveals Distinctive Staining Patterns in Metastatic Melanoma, Metastatic Carcinoma, and Mesothelial Cells in Pleural Fluids An Effective First-Line Test for the Workup of Malignant Effusions. Cancer Cytopathology, 122(8), 586–595. doi: 10.1002/cncy.21439 [DOI] [PubMed] [Google Scholar]
  4. Davidson LA (2012). Epithelial machines that shape the embryo. Trends Cell Biol, 22(2), 82–87. doi: 10.1016/j.tcb.2011.10.005 [DOI] [PMC free article] [PubMed] [Google Scholar]
  5. Deo RC (2015). Machine learning in medicine. Circulation, 132(20), 1920–1930. doi: 10.1161/circulationaha.115.001593 [DOI] [PMC free article] [PubMed] [Google Scholar]
  6. Fornas E, & Fortea A (1987). Autoradiography of endothelium in whole rat aorta by a new method. Atherosclerosis, 66(1-2), 95–98. doi: 10.1016/0021-9150(87)90183-3 [DOI] [PubMed] [Google Scholar]
  7. Guillot C, & Lecuit T (2013). Mechanics of epithelial tissue homeostasis and morphogenesis. Science, 340(6137), 1185–1189. doi: 10.1126/science.1235249 [DOI] [PubMed] [Google Scholar]
  8. Hassan R, Schweizer C, Lu KF, Schuler B, Remaley AT, Weil SC, & Pastan I (2010). Inhibition of mesothelin-CA-125 interaction in patients with mesothelioma by the anti-mesothelin monoclonal antibody MORAb-009: Implications for cancer therapy. Lung Cancer, 68(3), 455–459. doi: 10.1016/j.lungcan.2009.07.016 [DOI] [PMC free article] [PubMed] [Google Scholar]
  9. Hirata A, Baluk P, Fujiwara T, & McDonald DM (1995). Location of focal silver staining at endothelial gaps in inflamed venules examined by scanning electron-microscopy. Am. J. Physiol.-Lung Cell. Mol. Physiol, 269(3), L403–L418. [DOI] [PubMed] [Google Scholar]
  10. Hirsch EZ, Martino W, Orr CH, White H, & Chisolm GM (1980). A simple rapid method for the preparation of en-face endothelial (Hautchen) monolayers from rat and rabbit aortas. Atherosclerosis, 37(4), 539–548. doi: 10.1016/0021-9150(80)90061-1 [DOI] [PubMed] [Google Scholar]
  11. Jo H, Han J, Kim YS, Lee Y, & Yang S (2021). A novel method for effective cell segmentation and tracking in phase contrast microscopic images. Sensors (Basel), 21(10). doi: 10.3390/s21103516 [DOI] [PMC free article] [PubMed] [Google Scholar]
  12. Kachali C, Eltoum I, Horton D, & Chhieng DC (2006). Use of mesothelin as a marker for mesothelial cells in cytologic specimens. Seminars in Diagnostic Pathology, 23(1), 20–24. doi: 10.1053/j.semdp.2006.06.006 [DOI] [PubMed] [Google Scholar]
  13. Kandel ME, Kim E, Lee YJ, Tracy G, Chung HJ, & Popescu G (2021). Multiscale assay of unlabeled neurite dynamics using phase imaging with computational specificity. ACS Sens, 6(5), 1864–1874. doi: 10.1021/acssensors.1c00100 [DOI] [PMC free article] [PubMed] [Google Scholar]
  14. Lee GS, Miele LF, Turhan A, Lin M, Hanidziar D, Konerding MA, & Mentzer SJ (2009). Spatial calibration of structured illumination fluorescence microscopy using capillary tissue phantoms. Microsc Res Tech, 72(2), 85–92. doi: 10.1002/jemt.20647 [DOI] [PMC free article] [PubMed] [Google Scholar]
  15. Lehr JL, Butler JP, Westerman PA, Zatz SL, & Drazen JM (1985). Photographic measurement of pleural surface motion during lung oscillation. J. Appl. Physiol, 59(2), 623–633. [DOI] [PubMed] [Google Scholar]
  16. Levayer R, & Lecuit T (2008). Breaking down EMT. Nat. Cell Biol, 10(7), 757–759. doi: 10.1038/ncb0708-757 [DOI] [PubMed] [Google Scholar]
  17. Litjens G, Kooi T, Bejnordi BE, Setio AAA, Ciompi F, Ghafoorian M, … Sánchez CI (2017). A survey on deep learning in medical image analysis. Med. Image Anal, 42, 60–88. doi: 10.1016/j.media.2017.07.005 [DOI] [PubMed] [Google Scholar]
  18. Liu L, Wolterink JM, Brune C, & Veldhuis RNJ (2021). Anatomy-aided deep learning for medical image segmentation: a review. Phys. Med. Biol, 66(11). doi: 10.1088/1361-6560/abfbf4 [DOI] [PubMed] [Google Scholar]
  19. Miettinen TP, Caldez MJ, Kaldis P, & Björklund M (2017). Cell size control – a mechanism for maintaining fitness and function. BioEssays, 39(9), 1700058. doi: 10.1002/bies.201700058 [DOI] [PubMed] [Google Scholar]
  20. Newman GR, & Jasani B (1998). Silver development in microscopy and bioanalysis: a new versatile formulation for modern needs. Histochem. J, 30(9), 635–645. doi: 10.1023/a:1003404128497 [DOI] [PubMed] [Google Scholar]
  21. Oakley BR, Kirsch DR, & Morris NR (1980). A simplified ultrasensitive silver stain for detecting proteins in polyacrylamide gels. Anal.Biochem, 105, 361–363. [DOI] [PubMed] [Google Scholar]
  22. Panchev IN, Slavov A, Nikolova K, & Kovacheva D (2010). On the water-sorption properties of pectin. Food Hydrocoll, 24(8), 763–769. doi: 10.1016/j.foodhyd.2010.04.002 [DOI] [Google Scholar]
  23. Park J-A, Atia L, Mitchel JA, Fredberg JJ, & Butler JP (2016). Collective migration and cell jamming in asthma, cancer and development. J. Cell Sci, 129(18), 3375–3383. doi: 10.1242/jcs.187922 [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. Raftery AT (1979). Regeneration of peritoneum - fibrinolytic study. J. Anat, 129(OCT), 659–664. [PMC free article] [PubMed] [Google Scholar]
  25. Riese KH, Freudenberg N, & Haas W (1978). En face preparation methods for investigation of endothelia and mesothelia. Pathol. Res. Pract, 162(3), 327–336. [DOI] [PubMed] [Google Scholar]
  26. Servais AB, Kienzle A, Valenzuela CD, Ysasi AB, Wagner WL, Tsuda A, … Mentzer SJ (2018). Structural heteropolysaccharide adhesion to the glycocalyx of visceral mesothelium. Tissue Eng. Part A, 24, 199–206. doi: 10.1089/ten.TEA.2017.0042 [DOI] [PMC free article] [PubMed] [Google Scholar]
  27. Spatarelu C-P, Zhang H, Dung Trung N, Han X, Liu R, Guo Q, … Chen Z (2019). Biomechanics of collective cell migration in cancer progression: experimental and computational methods. ACS Biomater. Sci. Eng, 5(8), 3766–3787. doi: 10.1021/acsbiomaterials.8b01428 [DOI] [PMC free article] [PubMed] [Google Scholar]
  28. Valenzuela CD, Servais AB, Kienzle A, Ysasi AB, Wagner WL, Ackermann M, … Mentzer SJ (2017). Harvest and imaging of pleural mesothelium en face. FASEB J, 31. [Google Scholar]
  29. Van Valen DA, Kudo T, Lane KM, Macklin DN, Quach NT, DeFelice MM, … Covert MW (2016). Deep Learning Automates the Quantitative Analysis of Individual Cells in Live-Cell Imaging Experiments. PLoS Comput. Biol, 12(11), e1005177. doi: 10.1371/journal.pcbi.1005177 [DOI] [PMC free article] [PubMed] [Google Scholar]
  30. Walker C, Rutten F, Yuan XQ, Pass H, Mew DM, & Everitt J (1994). Wilms-tumor suppressor gene-expression in rat and human mesothelioma. Cancer Res, 54(12), 3101–3106. [PubMed] [Google Scholar]
  31. Wallau BR, Schmitz A, & Perry SF (2000). Lung morphology in rodents (Mammalia, Rodentia) and its implications for systematics. J. Morphol, 246(3), 228–248. doi: [DOI] [PubMed] [Google Scholar]
  32. Wang W, Taft DA, Chen YJ, Zhang J, Wallace CT, Xu M, … Xing J (2019). Learn to segment single cells with deep distance estimator and deep cell detector. Comput. Biol. Med, 108, 133–141. doi: 10.1016/j.compbiomed.2019.04.006 [DOI] [PMC free article] [PubMed] [Google Scholar]
  33. Zand T, Underwood JM, Nunnari JJ, Majno G, & Joris I (1982). Endothelium and silver lines - an electron-microscopic study. Virchows Arch. A-Pathol. Anat. Histopathol, 395(2), 133–144. doi: 10.1007/bf00429607 [DOI] [PubMed] [Google Scholar]
  34. Zheng Y, Pierce A, Wagner WL, Khalil H, Chen Z, Servais AB, … Mentzer SJ (2021). Functional adhesion of pectin biopolymers to the lung visceral pleura. Polymers, 13, 2976. [DOI] [PMC free article] [PubMed] [Google Scholar]

RESOURCES