Unilocular adipocyte and lipid tracer for immunofluorescent images

Abstract

Adipose tissue is a highly dynamic endocrine organ that serves as the body’s primary energy reservoir through the storage and mobilization of lipids. Adipocyte cellular size has been recognized as an indicator of cellular status; hypertrophic adipocytes are more prone to insulin resistance and the secretion of pro-inflammatory cytokines. Thus, the size and number of lipids is important to consider both in the clinic with a biopsy and when developing disease models and regenerative tissue constructs. Tools available to analyze adipocyte size are finely tuned for hematoxylin-eosin images and tend to be challenged by confocal derived z-stack images which contain intensity gradients. Therefore, ImageJ manual analysis is the commonly utilized tool to measure these images. With there being heterogeneity in different researcher’s analytical approach when conducted manually, the MATLAB script, PixCell, was developed to reduce the subjectivity and time involved in adipocyte size analysis. Given its stepwise thresholding and masking steps, PixCell retains on average a >80% accuracy when tested on excised human adipose tissue, adipocyte-laden collagen gels, and lipoaspirate seeded silk scaffolds. PixCell is able to consistently detect and measure lipids within regions of varying pixel intensities. This makes PixCell an appealing tool for use in both the clinical and pre-clinical setting, while greatly enhancing and streamlining the user experience of analyzing lipid sizes.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-024-80613-w.

Introduction

Adipose tissue is a highly dynamic endocrine organ that serves as the body’s primary energy reservoir through the storage and mobilization of free fatty acids and triglycerides in the form of intracellular lipid droplets¹. Adipocytes alter their size in order to accommodate natural fluctuations in energy intake and expenditure. In an energy demanding state, homeostatic control takes over to shift the tissue into a lipolytic state whereby triglycerides are hydrolyzed into their constituents: free fatty acids and glycerol². With approximately 95% of a white adipocyte being composed of the lipid droplet, hydrolysis of the lipid droplet results in a reduction in adipocyte cell size^3,4. However, expansion of adipose tissue is also possible through adipocyte hypertrophy (enlargement) or hyperplasia (formation of new adipocytes)^5,6. Adipocyte hypertrophy is characterized by an increase in adipocyte cell size through triglyceride synthesis. Notably, this occurs through stimuli such as insulin. Upon insulin release, lipoprotein lipase will be synthesized and bind to endothelial cells to break down plasma triglycerides into free fatty acids and glycerol for storage in adipocytes^7,8. If the demand for storage exceeds the available space within existing adipocytes, hyperplasia occurs where stem cells and preadipocytes proliferate and differentiate into new mature adipocytes in order to accommodate the surplus lipids⁹. In this instance, small multilocular lipids will merge to form one unilocular neutral lipid that is able to increase or decrease in size through lipogenesis or lipolysis, respectively. Importantly, hyperplasia has been marked as the “healthier” form of tissue expansion, while hypertrophied adipocytes have been correlated to inflammation and insulin resistance^10,11 indicating the importance of adipocyte size with the development and progression of related pathologies. Adipose tissue related pathologies are studied both clinically and preclinically. When assessing the cellularity and morphology of the adipocytes from a given individual in the clinic, biopsies are generally obtained to undergo histology via staining with hematoxylin-eosin (H&E) to display proteins and nuclear material¹². Current tools available to assess adipocyte size, like QuPath and AdipoCount, focus primarily on analyzing these H&E images^13,14. However, with the advancement of technology much more is being performed pre-clinically not only to study patient-derived samples, but also to develop adipose models for the purposes of studying obesity and adipose related pathologies, like adipose fibrosis, in addition to the development of tissue-engineered regenerative constructs^15–21. With many of these engineered models being three-dimensional in order to maintain the morphology of mature adipocytes and the native architecture of the tissue^19,20,22,23, confocal and multiphoton microscopy are commonly used to visualize these three-dimensional constructs²⁴. These imaging modalities allow for optical sectioning of thick samples, reducing the interference of out of focus light and enhancing the visualization clarity into a three-dimensional construct²⁴.

There are complexities specific to confocal immunofluorescent imaging that make analysis with programs such as QuPath and AdipoCount more challenging. In particular, these programs are finely tuned for H&E images whereby the brightest portion of the cell is the cell membrane. However, this is not the case in immunofluorescent imaging where there are intensity gradients within the cells along with variability from cell to cell²⁵. Further, H&E data is obtained from a slice of tissue (5–10 μm), typically paraffin-embedded which is melted along with the lipids, leaving a void, that is then imaged using brightfield microscopy. However, in maximum intensity projections of z-stacked images, volumetric data is compressed onto a two-dimensional plane. While this maximum intensity projection does not retain the depth information from the z-axis, it does provide indirect clues about the spatial distribution of the brightest structures along the z-axis. The brightest pixels within the maximum intensity projection indicate the relative position of the most intense features within the volume²⁶. Thus, depth is inferred from this spatial arrangement (with brighter features being closer to the viewer)^27,28. Analysis methods for these images must be able to identify and measure across a range of pixel intensities. Current tools, outlined in Table 1, struggle to appropriately analyze data found in the foreground or background of an image given that the presence of shadows or imperfections within a maximum intensity image disrupt the accuracy of the program. Thus, open-source software, ImageJ, is commonly utilized to manually count and measure the lipid sizes within these projected images^29,30. The drawback of this manual approach is apparent as it is time consuming while also adding a layer of bias into analysis. Here, we present a Matlab script capable of analyzing lipid droplet sizes from patient-derived samples and adipose tissue models imaged using confocal microscopy.

Table 1.

Current approaches available for adipocyte lipid size analysis.

Current Approaches	Strengths	Drawbacks	Citations
ImageJ Manual Analysis	• Open Source • Widely available and commonly used • User can apply perceptual restoration on incomplete adipocytes and manually delineate boundaries	• Bias between different users • Manual analysis takes more time than more automated approaches	13,29,30,31
AdipoQ	• Compatible with open-source ImageJ software • Comprehensive user guide assists with parameter adjustment • Adaptable from H&E to fluorescently labelled cells • Versatile and customizable based on user needs • Manual deletion capabilities	• Requires parameter optimization • High resolution with strong lipid borders is necessary • Parameter adjustment may be necessary on different images from the same data set	32
CellProfiler Pipeline	• Open Access • High Throughput • Performs very well on multilocular adipocytes	• Does not perform as well in dense culture regions or when lipids are fused to larger objects • Parameter adjustment may be necessary on different images from the same data set	33
QuPath	• Can handle large data files • Optimized for H&E images • Manual deletion capabilities within the software	• Requires manual adjustment of many parameters adding subjectivity • Not optimized for immunofluorescent images	29,34
Adipocyte Tools	• Plugin for ImageJ allows for increased accessibility • Thorough explanation and guide • Designed for H&E slide analysis	• Over-segments • Shadows and overlap cause interference	29,34
AdipoCount	• Highly specialized for H&E to detect outlines of lipid voids	• Designed specifically for H&E acquired images where adipocytes and lipids are voids	13,14
AdipoSoft	• Open source and compatible with ImageJ	• Designed for H&E where adipocytes and lipids are voids • Under-segments	35

Materials and methods

Immunohistochemical sample preparations

Subcutaneous adipose tissue was harvested from panniculectomies, or abdominoplasties performed at the University of Pittsburgh Medical Center (UPMC) with approval by the University of Pittsburgh Institutional Review Board (IRB No. 0511186). Tissue was manually dissected from the skin and either set aside for bulk tissue fixation or pulse blended in a Ninja Blender until the tissue resembled lipoaspirate. The tissue was then used for either seeding silk scaffolds or seeding into collagen gels.

Silk-based adipose tissue engineered constructs

Established protocols were followed in order to prepare adipose tissue-engineered models with silk scaffolds¹⁹. In short, whole Bombyx Mori cocoons (OliverTwistsFibres, Durham, UK) were cut into pieces and subsequently degummed by boiling for 30 min in a 0.02 M Sodium Carbonate solution (Na₂CO₃) (Sigma-Aldrich, St. Louis, MO, USA). The silk fibroin was rinsed and allowed to dry overnight at room temperature. Dry fibers were dissolved in a 9.3 M Lithium Bromide (LiBr) solution (Sigma-Aldrich, St. Louis, MO, USA) at 60 °C for 4 h. The solution was transferred to a dialysis cassette (Thermo Fisher Scientific, Waltham, MA, USA) and stirred in milli-Q water for 48 h with 6 intermittent water changes. Upon completion, the silk solution was centrifuged twice at 4800 rpm for 20 min to ensure solution purity. Aqueous silk was lyophilized and then dissolved in a 17% hexafluoroisopropanol (HFIP) solution (Sigma-Aldrich, St. Louis, MO, USA) overnight. In order to ensure consistent scaffold porosity, Sodium Chloride (NaCl) (Sigma-Aldrich, St. Louis, MO, USA) crystals were sifted to obtain crystals with diameters between 500 and 600 μm. The silk solution was then poured over the NaCl crystals within LDPE sample containers (Kartell Labware, Noviglio, MI, Italy). These containers were sealed for 24 h to allow for the silk permeation followed by opening the containers and allowing the solution to dry for an additional 24 h. The dried scaffolds were placed in methanol (PHARMCO-AAPER, Brookfield, CT, USA) for 24 h to induce β-sheet formation. After methanol annealing, the scaffolds were dried in a chemical hood for 24 h to allow for methanol evaporation. To leach the NaCl porogen, the scaffolds were rinsed for 2–3 days under continuous stirring. Finally, the scaffolds were cut into cylinders of 2 mm height and 4 mm diameter and autoclaved in distilled water for use in sterile applications.

Prior to seeding scaffolds with adipose tissue, the distilled water in the autoclaved silk scaffolds was replaced with fresh cell culture media (DMEM, high glucose, 10% FBS, 1% Penicillin/ Streptomycin (Thermo Fisher Scientific, Waltham, MA, USA)) and incubated at 4°C overnight. In order to seed the silk scaffolds with adipose tissue, scaffolds were removed from culture media, aspirated, and then submerged in the lipoaspirate. To promote cell attachment, this mixture was then incubated at 37 °C in the cell culture incubator for one hour. Seeded scaffolds were then removed from the lipoaspirate, placed in a 48-well tissue culture plate (Greiner Bio-One, Monroe, NC, USA), and then once again incubated for 1 h at 37°C before adding 1 mL of fresh medium to the seeded scaffolds. Media was refreshed twice per week for the duration of the experiment.

Seeding collagen gels with adipocytes

Liquified adipose tissue was placed in a collagenase solution composed of 1% BSA (Thermo Fisher Scientific, Waltham, MA, USA) and 0.1% Collagenase (Thermo Fisher Scientific, Waltham, MA, USA) within PBS (Thermo Fisher Scientific, Waltham, MA, USA). Liquified adipose tissue was combined with the collagenase solution at a 1:1 ratio and incubated at 37°C for 1 h. This collagenase solution is used in order to break down the extracellular matrix to release the adipocytes for future culture. Once incubated, the solution was then centrifuged at 300 xg for 5 min to phase separate the stromal vascular fraction (SVF), adipocytes, and an oil layer. The oil layer was aspirated off prior to the adipocytes being gently sifted through both a 1 mm and 350 μm sieve to remove any clumped extracellular matrix holding stromal cells. Isolated adipocytes were then centrifuged at 300xg for 5 min in order to remove any additional oil following filtration. PureCol (3 mg/mL) (Advanced BioMatrix, Carlsbad, CA, USA) was prepared following the manufacturer’s instructions whereby 80% of the total is PureCol, 10% is 10x DMEM, and 10% is 0.1 M NaOH to correct the pH. Isolated adipocytes were mixed with the collagen at a ratio of 0.25 mL adipocytes per mL of collagen. 500 μL of cell suspension was added to each well of a 48-well plate and then allowed to crosslink for one h at 37°C. Fresh cell culture media was added to each well with media being refreshed every other day for the duration of the experiment.

Inducing adipocyte hypertrophy on an adipose on a chip platform

Huff et al. recently utilized the Micronit platform (Enschede, Netherlands) to generate a fat-on-a-chip (FOAC) model with mature human adipocytes²¹. Briefly, mature adipocytes were mixed with HyStem hyaluronic acid (HA) (Advanced Biomatrix, Carlsbad, CA) at a 1:1 ratio prior to seeding on the Micronit chip and connection to the perfusion system which provided constant high-glucose media to the cells for four days. Chips were fixed on day 0 and day 4 to assess (via immunofluorescence and confocal microscopy) the fluctuation in lipid sizes from continuous nutrition²¹. Images from this experiment (6 images from two experiments) were provided and reused from Huff et al. for quantification by PixCell²¹.

Immunocytochemistry and confocal imaging

All samples (bulk tissue and adipose tissue seeded into silk scaffolds and collagen gels) were fixed with neutral buffered formalin (Sigma-Aldrich, St. Louis, MO, USA). at room temperature for 30 min followed by 3, 3-min PBS washes. Samples were treated with 0.1% Triton-X-100 for 15 min before being stained for 1 h at room temperature with either BODIPY(1:4000) (Thermo Fisher Scientific, Waltham, MA, USA) or AdipoRed (1:35) (Lonza, Durham, NC, USA) in order to visualize the lipids and Phalloidin-488 or Phalloidin-555 (Thermo Fisher Scientific, Waltham, MA, USA) to visualize f-actin. Z-stack scans were taken after staining using either a Zeiss LSM or Nikon Confocal Microscope. Imaging settings were maintained within trials. Images were preprocessed in opensource ImageJ Software³¹. Maximum intensity z-projections of the lipid channel (BODIPY/AdipoRed) were acquired through the following steps: ‘Image’ ◊ ‘Stacks’◊ ‘Z-project’ ◊ ‘Max Intensity’.

External data acquisition

To ensure reproducibility and generalizability of PixCell in other labs, we obtained confocal images of adipose tissue (in the form of z-stacks) that were acquired at the Kaplan Lab at Tufts University and the Ghobrial Lab at the Dana Farber Institute. Permission to use this unpublished data was granted by the investigators.

Image processing using PixCell

PixCell (https://github.com/rabbottlab/pixcell/, version 1) was developed within Matlab and requires the most recent (MATLAB_R2023) version in order to be run. In order to run the software, the following toolboxes must be installed: imaging toolbox, image processing toolbox, the computer vision toolbox, and the mapping toolbox. Upon loading PixCell into the desired directory, the user hits “Run,” and will be prompted to select an image for analysis. This image must be in the same directory as PixCell for the processing to run smoothly. Also, it is recommended that the user select a .tif file of the maximum intensity projection of the lipid channel as the image to analyze. While the average intensity can also be processed, there would be less cells processed in comparison to if the maximum intensity was processed. While other file formats are also accepted (.png, .jpg, .gif), please be aware that they will require manual entry of the pixel/micron ratio. After image selection it is necessary to follow any prompted instructions (like inputting the pixel/micron ratio of the image), wait completely until the code is done running (approximately 10 to 15 s), and then the user can reference the directory for the processed data. All images were analyzed using the default PixCell settings. The data will contain a folder for your image containing labelled lipid images alongside an Excel file containing the corresponding diameters and areas of the processed image. It is suggested that the user follow post-processing steps to ensure appropriate accuracy.

Image analysis using AdipoQ

Recently, Sieckmann et al. developed AdipoQ which consisted of two plugins, “AdipoQ Preparator” and “AdipoQ Analyzer”. Both plugins were downloaded from the repository (https://github.com/hansenjn/AdipoQ, version updated in 2022) and installed according to the provided user guide. Maximum intensity projections of z-stacked images of lipoaspirate seeded scaffolds (6 images from 3 separate trials) were preprocessed with the AdipoQ Preparator within ImageJ. Importantly, the setting to manually enter preferences (default settings: cultured cells) were used as they are designed to capture fluorescently labelled lipids. Recommended settings were retained as they worked effectively on a majority of our images. Settings can be seen in Supplementary Figure S7. After processing, AdipoQ Preparator provides a .tif file containing the segmented image that is then loaded and quantification undertaken with the AdipoQ Analyzer. Once the image is loaded into ImageJ and AdipoQ Analyzer run, preferences are manually entered into the GUI. Notably, the Minimum particle size is set to 20, the “Increase range for connecting adipocytes” box is checked, and “particles touching x or y borders” are excluded as is done in PixCell. All parameters were held constant when analyzing all of the images. All captured areas were converted to diameters using the area of a circle formula (A=πr²).

Image analysis using CellProfiler

Adomschick et al. outlined a pipeline for CellProfiler 4.2.6 (https://cellprofiler.org/) to quantify lipid droplet sizes³³. Their provided pipeline was utilized, and step-by-step protocol followed using maximum intensity projections of z-stacked images of lipoaspirate seeded scaffolds (6 images from 3 separate trials). With the images to be analyzed only including the lipid channel (BODIPY) the “ColorToGray” and “IdentifyPrimaryObjects” modules designated for the DAPI channel was disabled in CellProfiler. No modifications were necessary in the “ColorToGray” module for the BODIPY channel; however, settings were altered in the “IdentifyPrimaryObjects” module for the BODIPY channel to optimize the program for the size of the unilocular lipids. Notably, the “Typical diameter of, objects, in pixel units (Min, Max)” setting was altered to keep the minimum at 5 μm and maximum at 300 μm, which are the parameters within PixCell to acquire all of the lipids. Additionally, it was found that the “Method to distinguish clumped objects” and the “Method to draw dividing lines between clumped objects” needed to be set to “Shape” rather than “Intensity” in order to minimize the split of a unilocular lipid into multiple lipids (Figure S5). All parameters were held constant when analyzing all of the images. All captured areas were converted to diameters using the area of a circle formula (A=πr²).

Manual image analysis

Image analysis is manually conducted through loading the desired maximum intensity projection (6 images from 3 separate trials) into ImageJ, ensuring the scale is appropriately set, and using the “Freehand Selections” capability to manually trace an adipocyte. The area of this selection is then analyzed by selecting “Analyze” → “Measure” and the selection then turned into a permanent drawing by selecting “Edit” → “Draw”. This is repeated until all decipherable lipids are traced. Importantly, partial lipids residing along the X and Y axis are excluded from tracing to be consistent with the other automated approaches. Additionally, partial lipids were completed through referencing the corresponding z-stack. All captured areas were converted to diameters using the area of a circle formula (A=πr²).

Post-processing image analysis

For all images it is recommended that the user inspect the labelling accuracy for their image. PixCell accuracy was determined through the analysis and inspection of >2000 adipocytes per adipose platform (i.e. excised human adipose tissue (5 images from 2 trials), adipocyte-laden collagen gels (5 images from 2 trials), and lipoaspirate seeded scaffolds (5 images from 2 trials)). Every processed image was inspected for proper labeling and unique cellular identity, with reference to the adipocyte-only Excel sheet and the original image. In instances of an inappropriately labelled adipocyte, the cell was manually measured in ImageJ if the lipid was visible with reference to the z-stack. This manual measurement replaced that of the PixCell measurement. In instances where the lipid was not fully visible after reference to the z-stack the PixCell lipid measurement was discarded, and the lipid deemed as unmeasurable. Additionally, images that underwent comparison to AdipoQ and CellProfiler underwent pixel intensity mapping in MATLAB. Images were loaded, grayed (using the im2gray function), and then pseudo-colored with the indicated multicolor look-up table.

Statistical analysis

Data was analyzed and displayed with GraphPad Prism 9 Software (GraphPad, San Diego, CA, USA). All PixCell and manually corrected diameters were input and analyzed with GraphPad Prism to automatically establish bins to develop the frequency distributions for the percentages. Before assessing significant differences between PixCell measurements and the measurements after manual correction, the Kolmogorov-Smirnov test was performed to evaluate the normality of the data distribution. In instances where it was not Gaussian (as in the measurements obtained from the excised human adipose tissue and the hypertrophy measurements) a non-parametric test (the Kolmogorov-Smirnov test) was utilized to determine statistical differences between groups. Finally, a one-way ANOVA with Tukey’s multiple comparison test was used to assess statistical differences in the accuracy of the different platforms (excised adipose tissue, adipocyte-laden collagen gels, and lipoaspirate seeded silk scaffolds).

Results

A maximum z-projection takes the highest intensity pixels from each slice of a z-stack and merges them such that they form a two-dimensional representation of the three-dimensional structure imaged through z-stacks. This maximum intensity projection of the lipids is converted into a greyscale image to allow for thresholding with the imbinarize function. Initial thresholding steps, through this function, are applied in order to establish the difference in foreground and background through binarization of the image. In all images tested, the background is darker than the foreground. In order to maximize the number of full adipocytes captured, incomplete adipocytes on the boundary edges are excluded. As there is a stark difference in what is an object versus what is the background, boundaries of these objects can be established. As shown in Fig. 1 these objects are predominantly high intensity lipids. The Regionprops function allows for characteristics of the objects in black and white to be extracted. Characteristics can be defined by the individual, where the imperative ones for this algorithm included area, centroid (x, y), and both the maximum and minimum axis length. One unique feature of lipids is their apparent circularity, due to the aggregation of nonpolar elements³⁶. Drawing on this circularity allows for lipids to be separated from artifacts while also approximating each lipid’s area when a part of it is hidden by objects in the foreground. This algorithm applies a perfect circle to the coordinates of these entities, with the average of the maximum and minimum axis length being defined as the diameter of the circle. The centroid coordinates and non-scaled diameter are utilized to apply these circles for visualization while data is scaled and stored in a table. It should be noted that part of this process entails the deletion of entities with scaled diameters less than 5 microns (to account for artifacts due to resolution) and greater than 300 microns, as this is accepted as the maximum diameter for human adipocytes³⁷. Further, converted diameters between 5 and 20 microns are stored separately for the user to assess as extracellular lipids, if desired. What is then taken to be higher intensity adipocytes are then masked with black pixels to allow for features with less intensity to be isolated and stored as well. These adipocytes require the application of locally adaptive thresholding on the original image concatenated with the masked image. This threshold is calculated based on the local mean intensity of the neighboring pixels. This approach excludes any individual based error from when a user sets the threshold as is sometimes conducted in ImageJ³³. Through this approach, bias from manual thresholding is avoided. Unique cellular identity between high-pass and low-pass data is established through the UniqueTol function in MATLAB which only allows for the existence of unique elements within a given tolerance. In this application, the coordinates of the centroid are compared with each other. If the absolute elemental difference is less than or equal to 0.05, then the adipocyte is considered a duplicate and deleted from the main data set. This value is set to account for the compact nature of adipocytes and the slight overlap of lipids when capturing a z-stack.

Fig. 1.

Representation of the steps that PixCell takes in order to process a lipid image. 1) A maximum intensity image is selected for input, 2) The maximum intensity projection is greyscaled, 3) This greyed image undergoes a first round of thresholding in order to capture the high intensity cells, 4) These high intensity cells are masked with black pixels and merged onto the original image before 5) undergoing an additional thresholding to capture any lipids missed in the first pass, which are predominantly the lower intensity lipids. 6) The two captures are compared with only unique lipids being saved (duplicates excluded), and 7) This data is outputted as an image and corresponding Excel sheet for the user to conduct any post-processing. In the masking step, high intensity lipids are shown in green and the lower intensity lipids, purple. Additionally, intermittent thresholding results (High Intensity Capture and Low intensity Capture) are shown in red while final (post-duplicate removal of high and low intensity capture comparison) adipocytes captured are shown in blue.

Whole mount imaging of adipose tissue samples was performed on excised adipose tissue given the inclusion of the native tissue architecture and lipid structure. The compact nature of adipose tissue is demonstrated in Fig. 2A, with each adipocyte in close proximity to its neighbors. 10 maximum intensity projections of excised adipose tissue were processed using PixCell (see Figure S1 for individual images alongside their PixCell Segmentation and frequency distribution before and after manual correction), with Fig. 2B showing an average representation of PixCell’s results whereby the final adipocytes’ centroid and radii are overlaid onto the original image, making each adipocyte a circle. The accuracy of this overlay was manually determined through the evaluation of > 2000 adipocytes from the processed maximum intensity projections. It should be noted that if perturbations occur that affect the circularity of the adipocyte, accuracy might be affected. An adipocyte whose PixCell overlay did not appropriately match up with its original outline was marked for manual correction. With the PixCell overlay having a numerical label corresponding to its number in the outputted Excel sheet, the PixCell adipocyte diameter can be manually corrected in these post-processing steps. The incorrect PixCell adipocytes were manually measured in ImageJ if they were deemed measurable. This was done through the application of an appropriately sized circle on the incorrect PixCell adipocyte, and the extraction of the corresponding (converted) area and diameter by pressing “Analyze” and then “Measure” in ImageJ. Figure 2C shows the histogram of adipocyte diameters before and after manual correction. After manual correction there are less adipocytes with a diameter of < 60 μm and more adipocytes with a diameter of > 60 μm. It can be seen that there is 10% increase in the percentage of adipocytes measured in the 70 μm bin after manual correction. This did end up statistically increasing the overall average mean (Fig. 2D).

Fig. 2.

Excised Adipose Tissue was stained with BODIPY or AdipoRed (green = lipids), z-stacks acquired via confocal imaging, and maximum intensity projections (5 images from two separate patients) (A) utilized for PixCell processing. PixCell outputs a visualization of adipocytes it processed (B) along with an Excel file of labelled adipocyte diameters (with approximately 200 lipid diameters per image). Following manual correction in ImageJ, the histogram of adipocyte diameters (C), and average diameters (D) was compared. Error bars represent standard deviation relative to the displayed mean. Scale bar = 100 μm. *** indicates a p-value < 0.001.

Despite an accuracy of 85.51% (Fig. 3B), the remaining 14.49% lipids that were inappropriately measured were carefully inspected in order to determine instances that may challenge PixCell. In instances where a minimal portion of the adipocyte edge was visible (Fig. 3A), PixCell registered the object, but underapproximated its diameter (Fig. 3C). If the adipocyte was not discernable alongside the corresponding z-stack (going up and down through the z-stack to find the boundary of the adipocyte) the adipocyte was deemed unmeasurable and manually deleted from the data set. Another case that was noted was when PixCell was unable to recognize the individuality of the objects, resulting in the overlaid circle being accounted by two or more lipids (Fig. 3D). These were treated as individuals during their manual correction. Finally, artifacts, such as nonspecific binding or background signals, were detected by PixCell and required manual deletion (Fig. 3E).

Fig. 3.

PixCell generated overlay (A) of a representative image of excised human adipose tissue stained with BODIPY. Despite an 85.51% average accuracy (calculated from 5 images from two separate patients), PixCell was inaccurate (B) when only a small portion of the lipid edge is shown (making it unmeasurable) (C), when there are shadows from the foreground or unclear boundaries between cells (D), and when an artifact such as non-specific staining interfered (E).

With the rise of adipose tissue-engineered constructs for in vitro drug testing and disease modeling¹⁵, it is crucial that all adipocyte tracing software be compatible and show reproducible results with images of lipids in various in vitro platforms. Thus, PixCell was not only tested on human excised adipose tissue, but also on liquefied adipose tissue seeded on silk scaffolds and isolated adipocytes seeded into collagen gels. These platforms have been widely used to develop models for the study of adipose tissue targeted drugs and diseases^{17–19,38–39}. To seed the cells into the gels and onto the scaffolds, the matrix of the adipose tissue must be broken down, such that the cells can then conform to their new environment. Adipose tissue is liquified to allow for the cells to penetrate into the pores of the silk scaffold. This liquified adipose tissue resembles lipoaspirate in its consistency as well as its cellular contents (unilocular adipocytes, multilocular adipocytes, adipose stromal cells, and endothelial cells)¹⁹. However, in order to seed adipocytes into collagen gels, the architecture must be further broken up through enzymatic digestion⁴⁰. Once the adipocytes are released from their matrix, they can be seeded into collagen gels. These processing steps are apparent when assessing their corresponding PixCell processed immunocytochemistry images (Fig. 4B). Importantly, adipocyte-laden collagen gels contain more space between the adipocytes compared to excised adipose tissue and liquified adipose tissue. Given that adipocyte overlap (Fig. 3B and C) is one of the reasons for PixCell errors, this may explain why adipocyte-laden collagen gels have the highest PixCell accuracy (Fig. 4A and Figure S2/S3 for individual images alongside their PixCell segmentation and frequency distribution before and after manual correction). While utilizing silk scaffolds with liquified adipose tissue allows for the retention of the host cells, silk autofluorescence complicates the usage of an automated lipid quantifier due to its recognition as an object and thus interpretation as a lipid by PixCell (Figure S4). This influenced the accuracy of PixCell on adipose-laden silk scaffolds and must be addressed in manual correction (Fig. 4A). Similar to the frequency distribution of adipocytes from excised adipose tissue (Fig. 2C), there are less adipocytes under the diameter of 40 μm after manual correction than what is measured by PixCell in both adipose tissue models assessed for accuracy (Fig. 4C, D) displaying PixCell’s slight diameter underapproximation. Lipoaspirate seeded scaffolds experience a maximal number of adipocytes in the 70 μm bin (Fig. 4C) similar to what is seen in the excised adipose frequency distribution (~60 μm maxima) (Fig. 2C), while adipocyte-laden collagen gels experience the most adipocytes in the 50 μm diameter bin (Fig. 4D).

Fig. 4.

PixCell Accuracy (calculated from 5 images from 2 separate trials and displayed as the mean and standard deviation) (A) on maximum intensity projections of confocal-imaged and lipid stained (green) Excised Human Adipose Tissue, Adipocyte-laden Collagen Gels, and Lipoaspirate Seeded Scaffolds (B). Scale Bar = 100 μm. The frequency distribution before and after manual correction of PixCell measurements of lipoaspirate seeded scaffolds (C) and Adipocyte-laden collagen gels (D) indicates an increase in the number of lipids >50 μm and a decrease in the lipids <40 μm after manual correction. * Indicates a p-value < 0.05 and ** indicates a p-value < 0.01.

With lipoaspirate seeded scaffolds experiencing the lowest accuracy in PixCell (Fig. 4A), 6 maximum intensity projections from 3 separate trials of lipoaspirate seeded scaffolds were analyzed by PixCell, AdipoQ³², and the CellProfiler pipeline generated by Adomschick et al.³³. Images were also analyzed manually in ImageJ to compare both PixCell and other software against. Collectively, there are a range of differences in the frequency distribution of the measurements from each software (Fig. 5A). The maximal frequency (18.45%) of the manual measurement’s diameters resides in the 40 μm bin, PixCell’s peak (22.66%) is in the 30 μm bin, CellProfiler has 18.78% of the diameters in the 35 μm bin, and AdipoQ has 18.06% of the diameters in the 30 μm bin (Fig. 5A). These discrepancies are better comprehended by examining the performance of each software on different images within each trial. Figure 5B delineates the segmentation outcomes from the 4 analytical methods applied to 3 separate images from trial #1, highlighting the variability in segmentation. Visually, there are noticeable differences in the segmentation efficacy between the different images (Fig. 5B). While AdipoQ seemingly segments the adipocytes very well in image 1, it was not as effective in image 3 and 5. It is plausible that this could be due to the troubles caused by regional pixel intensity disparities. Figure S6 shows the intensity maps of images 1, 3, and 5. When there are consistent intensities within a region (shown in image 1), AdipoQ performs well, but when there are drastic intensity changes, the adipocytes are not captured as effectively, as evidenced in image 5 and somewhat in image 3. AdipoQ did not detect as many lower intensity adipocytes as the CellProfiler pipeline and PixCell (Figure S6, S9, S10, S11). Despite detecting adipocytes in the lower intensity regions, the CellProfiler pipeline had tendencies to either under-segment or over-segment the adipocytes, sometimes within the same image (Fig. 5B). For instance, image 3 exhibits an instance where two adipocytes were detected as one (under-segmented) and instances where one true adipocyte is broken into two segmented adipocytes (over-segmented) (Fig. 5B). While PixCell also encounters such issues (Fig. 3), it more consistently captures adipocytes across varying intensity regions and images without any script adjustments (Fig. 5B). The violin plots shown in Fig. 5C display the diameter distributions for each image and analytical method (all of the violin plots for each image are included in Figure S8-11). With visual perception aided by a companion z-stack, the manual measurements are the most thorough and are what we use as a point of reference for the more automated methods. In comparing both the spread and distribution of the individual AdipoQ, CellProfiler, and PixCell measurements to that of the manual measurements, there are some notable similarities and differences. For instance, the plots from images 4, 5, and 6 are relatively consistent between CellProfiler, PixCell, and the manual measurements with just CellProfiler displaying an elevated maxima in comparison to both PixCell and the manual measurements (Fig. 5C). Differences in the spread of the data are definitely seen when comparing AdipoQ to the manual measurements with images 2, 3, and 6 having elevated maximas, whereas the majority of the data is less than 30 μm in images 4 and 5. It seems that PixCell performs similarly across different images from different trials (Figure S9, S10, S11), indicating its potential as a reliable lipid tracer.

Fig. 5.

Methods for adipocyte analysis (AdipoQ, CellProfiler, and PixCell) are compared to the benchmark, manual analysis. The frequency distribution (A) of the measurements from each method show a range of differences best understood by looking at the visual results from each method with CellProfiler indicating individual lipid droplets with individual colors (B). BODIPY (green = lipids). Scale Bar = 100 μm. The performance of the method varies between images from the same trial, which is reflected in the violin plots (C) showing the lipid diameter distribution results for each image with each method. Data is re-displayed in supplemental figures (S8 and S9) amongst their related images and replotted in S17 with all 3 trials.

With PixCell’s ability to identify and measure adipocytes from both a range of platforms and within regions of varying intensities, it was then imperative to assess the possibility for PixCell to be used in characterizing adipose tissue from lean and obese individuals. Huff et al. provided blinded images acquired on adipose samples before (day 0) and after (day 4) induction of adipocyte hypertrophy²¹. Figure 6A shows representative images before and after hypertrophy induction alongside their PixCell generated image (see Figure S12 and S13 for all of the images alongside their PixCell segmentation from trial 1 and trial 2 before and after the induction of hypertrophy on a fat on a chip model). After manual correction of the PixCell measurements it can be seen that there is no statistical difference between the mean of the PixCell measurements and the manually corrected PixCell measurements from day 0 and day 4 images (Fig. 6B). However, it should be noted that after manually correcting the day 4 measurements the average adipocyte diameter went up (Fig. 6B). The frequency distribution of this data (Fig. 6C) shows an increase in the frequency of adipocytes in the 30 and 40 μm bins after manually correcting the day 0 measurements. After manual correction of the day 4 measurements the distribution is held relatively constant besides an apparent increase in the frequency of >100 μm adipocytes. Taken together, this shows that PixCell has a slight tendency to underestimate hypertrophic adipocytes. However, after the induction of hypertrophy, PixCell detects both a reduction in the frequency of adipocytes less than 50 μm and an increase in the frequency of >70 μm (Fig. 6C) which leads to a statistically significant increase in the average adipocyte diameter (Fig. 6B) which is a characteristic of hypertrophic adipocytes.

Fig. 6.

Immunofluorescent images of adipocytes (green) within a fat on a chip system, before and after the induction of hypertrophy were analyzed by PixCell (A). Scale bar = 200 μm. Results were scrutinized and manually corrected following previously established criteria and means and standard deviations displayed (B) * indicates a p-value < 0.05 and **** indicates a p-value < 0.0001 from the Kolmogorov-Smirnov non-parametric test. The frequency distributions of all computed results are shown (C).

In looking at the data with all of the lipids (Fig. 7C) it is apparent that there are many lipids’ diameters that are below 20 μm. Additionally, there has been a recent interest in the involvement of extracellular lipid droplets in a disease state and as a means of intercellular communication^41,42. It was seen that these smaller lipids were not contained within any cell membrane⁴³. With it being accepted that an adipocyte’s diameter is above 20 μm³⁷and the upper limit for an extracellular lipid is 15.43 μm⁴³, these lipids were separated from the adipocyte data in both the high pass and low pass steps. These lipids were compared to each other to allow for unique lipid identity, but were not compared to the adipocyte centers, as it is not uncommon for these lipid droplets to be on top of an adipocyte (Fig. 7A) but did not have to be (Fig. 7B). These lipids were stored in a separate Excel sheet and each of their centroids and diameters were utilized to make an extracellular lipid figure (Fig. 7D) for each data set. Additionally, PixCell exports both an Excel sheet and a figure with all lipids and exclusively the extracellular lipids if that is of interest to the user (Fig. 7D).

Fig. 7.

Lipids (green color) not contained within the bounds of the adipocyte cell membrane are measured by PixCell. Adipocyte-laden collagen gels (scale bar = 100 μm) (A) displayed smaller extracellular lipids (scale bar = 50 μm) (B). PixCell outputs lipids less than 20 microns in diameter (C) in a separate Excel and image file (D) for inspection by the user.

Discussion

With H&E staining having been well established for over a century, both preclinical and clinical analysis has relied heavily on this imaging modality⁴⁴. However, since Hematoxylin targets acidic structures and Eosin binds nonspecifically to proteins, there is a limitation of how much information can be obtained from these micrographs^45,46. Further, the visualization of adipose tissue with this technique requires the dissolution of the lipids, with lipid area being inferred from the cell membrane outline. With the field of tissue engineering and regenerative medicine greatly advancing, so too are the methods for analyzing these engineered tissue constructs. Both the advancement of staining and imaging techniques have allowed for the visualization of more complex structures. For example, immunocytochemistry has allowed for the visualization and characterization of adipocytes throughout their course of differentiation, with early stages being marked by the expression of Pref-1 and CEBP and later stages, perilipin⁴⁷. These cells also undergo drastic morphology changes throughout the course of their differentiation, going from fibroblastic-like cells to multilocular immature adipocytes, and then unilocular adipocytes when driven towards the white adipose tissue lineage⁴⁸. This knowledge has been obtained through immunofluorescence multiplexing, whereby multiple antigens are targeted in conjunction with lipophilic dyes in order to trace lipid morphology and growth over the course of differentiation⁴⁹. Additionally, mature adipocytes are inherently three-dimensional structures, and when combined with three-dimensional culture platforms, benefit from the usage of confocal microscopy to image through the construct and interpret the lipids’ sizes and shapes⁵⁰. The ability to pair these histology and imaging forms has greatly informed research into obesity, diabetes, and cardiovascular disease. However, limitations exist in the ability to analyze the data obtained from these images as adipocyte software is predominantly tailored to processing H&E images (Table 1).

PixCell was developed with this advancement of staining and imaging in mind. It utilizes maximum intensity projections of fluorescently stained lipids acquired from confocal imaging. Considering how these images were acquired by imaging through the z-plane and then generating a maximum intensity projection, there are complexities specific to this type of image not found in H&E images. PixCell employs a masking approach in order to capture adipocytes in the background and foreground. While there is recognition of these adipocytes, PixCell accuracy is limited in instances where there is maximal adipocyte overlap. If there is a limited portion of the rounded edge of the adipocyte visible in the maximum intensity projection, there is an increased likelihood that PixCell either underestimates the adipocyte diameter or merges this lipid with its neighboring adipocyte. Despite there being statistical differences in the accuracy between the different three-dimensional platforms (Fig. 4A), the standard deviation is relatively small for the accuracy within each platform indicating consistency between the accuracy of PixCell labeling of images from the same adipose platform. It is apparent that artifacts, such as oil and silk autofluorescence (Figure S4) will interfere with PixCell’s accuracy; thus, the user must take this into consideration while imaging or must perform manual correction. It is recommended that the user have many cells in a high-resolution image to reduce the percentage of cells that are incorrectly labelled. Further, it is recommended that the user consider the number of cells on the edge of the image as those will most likely be omitted from processing. Despite these limitations, PixCell showed consistent accuracy when labelling and measuring adipocytes from several ex vivo tissue and in vitro tissue-engineered platforms while also being able to correctly identify and measure adipocytes before and after the induction of hypertrophy (Fig. 6). Additionally, PixCell can be used on adipose images obtained from other labs (Figure S15), indicating its generalizability. Overall, PixCell provides a more user-friendly approach to acquiring lipid measurements while also reducing the potential bias that is involved in obtaining manual ImageJ measurements.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Author contributions

E.K.J. wrote the main manuscript text and made the figures. T.D. wrote the code with input from N.A.M., E.K.J., and R.D.A. All authors reviewed the manuscript.”

Data availability

The matlab script is included as a supplement to this manuscript (see attached) and is available at the repository https://github.com/rabbottlab/pixcell/ under an MIT license. All datasets generated and analyzed during the current study are available in the github repository.

Declarations

Competing interests

The authors declare no competing interests.