PodoCount: A Robust, Fully Automated, Whole-Slide Podocyte Quantification Tool.

Introduction: Podocyte depletion is a histomorphologic indicator of glomerular injury and predicts clinical outcomes. Podocyte estimation methods or podometrics are semiquantitative, technically involved, and laborious. Implementation of high-throughput podometrics in experimental and clinical workflows necessitates an automated podometrics pipeline. Recognizing that computational image analysis offers a robust approach to study cell and tissue structure, we developed and validated PodoCount (a computational tool for automated podocyte quantification in immunohistochemically labeled tissues) using a diverse data set.
Methods: Whole-slide images (WSIs) of tissues immunostained with a podocyte nuclear marker and periodic acid-Schiff counterstain were acquired. The data set consisted of murine whole kidney sections (n = 135) from 6 disease models and human kidney biopsy specimens from patients with diabetic nephropathy (DN) (n = 45). Within segmented glomeruli, podocytes were extracted and image analysis was applied to compute measures of podocyte depletion and nuclear morphometry. Computational performance evaluation and statistical testing were performed to validate podometric and associated image features. PodoCount was disbursed as an open-source, cloud-based computational tool.
Results: PodoCount produced highly accurate podocyte quantification when benchmarked against existing methods. Podocyte nuclear profiles were identified with 0.98 accuracy and segmented with 0.85 sensitivity and 0.99 specificity. Errors in podocyte count were bounded by 1 podocyte per glomerulus. Podocyte-specific image features were found to be significant predictors of disease state, proteinuria, and clinical outcome.
Conclusion: PodoCount offers high-performance podocyte quantitation in diverse murine disease models and in human kidney biopsy specimens. Resultant features offer significant correlation with associated metadata and outcome. Our cloud-based tool will provide end users with a standardized approach for automated podometrics from gigapixel-sized WSIs.

Chronic kidney disease is a state of prolonged and progressive reduction in kidney function that may evolve to end-stage kidney disease (ESKD). Driven by increasingly prevalent conditions with high incidence (e.g., diabetes, hypertension), chronic kidney disease accounts for unprecedented mortality and socioeconomic burden., To mitigate this, biomedical initiatives aim to identify disease-related biomarkers with improved precision for early detection and intervention. For glomerular diseases, some biomarker studies have focused on podocytes, highly specialized epithelial cells that maintain the kidney filtration barrier. The podocyte depletion hypothesis proposes3, 4, 5, 6, 7, 8 that podocyte loss in the setting of podocytopathic injury from hyperfiltration, hyperglycemia, or hypertension is an early determinant of proteinuria and glomerulosclerosis.9, 10, 11 Thus, podocyte enumeration offers a measurable indicator of irreversible glomerular injury and therapeutic success in these states.

Unfortunately, existing podocyte estimation methods or podometrics are semiquantitative and not scalable or offer morphologic assessment.,,12, 13, 14, 15, 16 Furthermore, podocyte identification on routine and special stains viewed under brightfield microscopy remains difficult. In a recent study, Venkatareddy et al. present a novel methodology for podometric estimation, including podocyte count and density from a single histologic section. Recognizing that the number of nuclear profiles per glomerulus cross-section is not a true estimate of podocyte count, this single-section method applied podocyte nuclear labeling, manual annotation, image analysis software, and stereological equations to arrive at a correction factor (CF) which modulates podocyte count and volume density estimates based on section thickness. Designed for single glomerulus quantitation from image patches, this estimation method lacks scalability to the WSI context. To achieve big-data podocyte studies that facilitate early detection and intervention, accurate and automated methods for brightfield whole-slide podocyte quantification must be established. Therefore, we developed PodoCount (an automated podometric tool for single-section estimation from WSIs) and used it to evaluate WSIs of kidney sections immunostained with a podocyte marker (see the workflow in Supplementary Figure S1).

Methods

Data Sources

Human data collection followed protocols approved by the Institutional Review Board at the Seoul National University College of Medicine (H-1812-159-998), Seoul, Republic of Korea. All experiments were performed according to federal guidelines and regulations. Animal studies were performed in accordance with protocols approved by the Institutional Animal Care and Use Committee at the Georgetown University, National Institutes of Health, University of Minnesota, and Johns Hopkins University, are consistent with federal guidelines and regulations, and are in accordance with recommendations of the American Veterinary Medical Association guidelines on euthanasia.

Description of Murine and Human Data

We analyzed 135 whole kidney sections from 135 mice across 6 kidney disease models (Figure 1a,c) and 45 kidney biopsy specimens from patients with DN (Figure 1b,d), as detailed in the subsequent sections. These multi-institutional, murine and human, male and female data feature highly variable sample preparation, staining, imaging, and pathology, comprising a diverse data set to assess robustness and reproducibility (Supplementary Figure S2).

Figure 1

Summary of data sets. The image data set contains light microscopic images of kidney tissues from 6 mouse models of glomerular disease and 5 stages of human DN. (a) The murine cohort was composed of tissues from 135 mice with control and diseased specimens for each model. Two distinct models of type II diabetes mellitus were studied (db/db and KKAy). The SAND intervention (saline, angiotensin II, uninephrectomy, and deoxycortisone) models postadaptive FSGS (FSGS [SAND]). Samples from SAND, HIVAN, and Progeroid syndrome models included male and female mice; those from the db/db, KKAy, and Aging mouse models consisted only of males. (b) The human DN study consisted of 45 patients (n = 35 male and n = 10 female subjects). Representative glomerular p57-PAS image from (c) each mouse model and (d) each Tervaert stage of the human DN cohort. DN, diabetic nephropathy; FSGS (SAND), a postadaptive model of FSGS, focal segmental glomerular sclerosis; HIVAN, HIV-associated nephropathy.

Models 1 and 2 (Diabetes)

There were 2 models of type 2 diabetes mellitus used (db/db and KKAy). Model 1 consisted of db/db (leptin receptor mutation) mice on BKS background (n = 4) that developed spontaneous/congenital disease and were compared with wild type (WT) BKS mice (n = 6). Model 2 consisted of the KKAy mice (n = 10) (described previously) that developed spontaneous diabetes of polygenic origin and were compared with WT mice (n = 8).

Model 3 (Focal Segmental Glomerular Sclerosis)

A postadaptive form of focal segmental glomerular sclerosis (FSGS) was induced in WT FVB/N mice (n = 8) by a combination of 4 interventions (SAND) (0.9% saline drinking water, angiotensin II infusion [osmotic pump], unilateral nephrectomy, and deoxycorticosterone [subcutaneous pellet],) and were compared with untreated mice (n = 11).

Model 4 (HIV-Associated Nephropathy)

Tg26 mice (gag-pol-deleted HIV-1 genome25) (n = 13) on an FVB/N × 129S F1 background with a collapsing glomerulopathy were compared with littermate controls (n = 11).

Model 5 (Aging)

In this aging model, 21-month-old C57BL/6 male mice (n = 5) were compared with 4-month-old controls (n = 6), both obtained from the National Institute on Aging rodent colony.

Model 6 (Progeria)

An Ercc1 progeroid mouse model (n = 30) and WT littermate controls (n = 20) (15–18 weeks old) on a C57BL/6J:FVB/N f1 background were used for this study., Mice were bred and genotyped as previously described.

Human Cohort

Human tissues consisted of needle biopsy samples from human patients with type 2 diabetes mellitus (n = 45) (biopsied 2011–2017 at the Seoul National University Hospital, which were collected from the Seoul National University Human Biobank). Biopsies were graded by a renal pathologist based on the Tervaert classification scheme. Clinical metadata including serum creatinine and estimated glomerular filtration rate (eGFR) were measured at the time of biopsy and at 1 year and 2 years postbiopsy. For the purposes of our study, progression to ESKD within 2 years after biopsy was the primary end point. Patients with type 1 diabetes mellitus and additional glomerulonephritis and patients without need for antidiabetic agents were excluded from the cohort.

Sample Preparation and Imaging

All samples were formalin-fixed, paraffin-embedded tissues cut at 2 μm in thickness. Podocyte nuclei were immunohistochemically labeled for p57kip2, a marker of podocyte terminal differentiation (catalog number ab75974, Abcam, Cambridge, United Kingdom), and detected with horseradish peroxidase and diaminobenzidine chromogen substrate (catalog number RU-HRP1000, Diagnostic BioSystems, Pleasanton, CA, and catalog number BSB0018A, Bio SB, Santa Barbara, CA, respectively). A periodic acid–Schiff poststain was applied without hematoxylin counterstain.

Brightfield WSIs were captured using an Aperio AT2 microscope (Leica Microsystems, Buffalo Grove, IL) or a NanoZoomer S360 slide scanner equipped with a 40× objective (Hamamatsu Photonics, Bridgewater, NJ). The full image data set consisted of WSIs of 135 whole murine kidney sections and 45 DN biopsy specimens, all from discrete mice and human participants (Figure 1).

Whole-Slide Segmentation of Renal Parenchyma

Images of podocytes and renal tissue compartments were extracted from WSIs through image segmentation by selecting image regions of interest based on differences in color, texture, and shape. Segmented structures included whole tissue sections, glomerular boundaries (glomerular capillaries, Bowman’s space, and Bowman’s capsule), and podocyte nuclei. Select image processing techniques that require a more extensive explanation are italicized and defined in the glossary of terms in Supplementary Table S1. Structural segmentation involved several sequential steps. First, a global mean-based threshold was applied to segment the tissue section from the WSI background. Glomerular boundaries were then detected using our published Human-AI-Loop tool, a convolutional neural network developed for WSI segmentation. For WSIs with >1 tissue section, tissue boundary segmentations were used to partition the whole-slide glomerular population into groups linked to their respective tissue section. Immunohistochemically positive podocyte nuclei were then segmented from detected glomerular units using stain deconvolution and local mean-based thresholding. Morphologic image processing techniques, including hole filling, size exclusion, and marker-controlled watershed,31, 32, 33 were applied to refine segmentations and separate overlapping nuclei.

Computational performance of WSI segmentation was evaluated. Tissue section and glomerulus boundaries were manually annotated in 12 randomly selected WSIs, equally sampled from each murine data set. A stratified randomized sampling was performed, wherein for each cohort and disease state, samples (or images) were given a number and then selected via the random number generator in MATLAB (“rand” command). Per WSI, automated segmentations were compared pixel wise against manual ground truth. To quantify tissue and glomerulus boundary segmentation performance, the pixel-wise sensitivity, specificity, precision, and accuracy were calculated across WSIs, defined as,where TP, true positive; TN, true negative; FP, false positive; and FN, false negative.

Glomerulus images were randomly and equally sampled from murine (n = 40 per cohort, 20 each from control and disease) and human (n = 40 per DN stage) data, and podocyte nuclei were manually annotated in each image. Pipeline segmentations were compared pixel wise against manual ground truth. To quantify podocyte nuclear segmentation performance, sensitivity, specificity, precision, and accuracy were computed per glomerulus image (as defined previously).

Performance of podocyte detection was assessed with Hit-Miss analysis to determine the frequency at which podocyte nuclei were positively identified. Median performance was computed per cohort (median of n = 40 murine glomerulus images or n = 200 human glomerulus images) and then across all data sets.

Computational Podocyte Count and Density Estimation From WSIs

Segmented podocyte nuclei were automatically enumerated in each glomerulus image as a raw count. Image analysis techniques were then applied in sequence to emulate the single-section method (Supplementary Figure S3). First, the bounding box of each nuclear profile was derived, and the box length and width were averaged to find the apparent caliper diameter per profile (d). Then, d values were averaged to d, and established equations were used to estimate D, CF, the corrected podocyte count, and podocyte density (number per 106 μm³).

The performance of podocyte count and density estimates, which were output as continuous values, was assessed by calculation of error. To do this, computational estimates of podocyte count and density were compared against manual ground truth generated using MATLAB as described in Venkatareddy et al. (Estimation of D Using Image-Pro Software and the Quadratic Equation). For a detailed explanation of ground truth generation, see the Supplementary Methods.

Whole-Slide Podocyte and Glomerulus Feature Extraction

Built-in morphologic operations were applied to derive whole-slide coordinate locations and geometric features from podocyte nuclei and glomerulus profiles. Geometric features included image object area (μm2), bounding box area (μm2), convex area (μm2), eccentricity, equivalent diameter, extent, major and minor axis lengths (μm), orientations, perimeters (μm), and solidities. A brief description of each feature is provided in Supplementary Table S1. Feature statistics were computed per podocyte, per glomerulus podocyte population, and per WSI glomerulus population.

Biologically Inspired Podocyte Feature Engineering

Additional morphologic and spatial features were engineered from podocyte nuclear profiles and glomerulus units. Features validated in this work included total podocyte nuclear area (μm2) and podocyte nuclear coverage (Table 1). Validation of handcrafted features (e.g., podocyte nuclear distance to the glomerulus unit edge, Supplementary Table S2) was the scope of future work, and thus further discussion was omitted.

Table 1

Histologic image feature definitions for podometrics

Features	Definition
PC	Corrected podocyte count after application of the single-section method’s CF. Computed as number of podocyte nuclear profiles times the CF.
GA	Cross-sectional area of the glomerulus unit (μm2).
GPD	Podocyte volume density is computed as the ratio of the corrected podocyte count to the glomerulus volume and approximates the spatial density of podocytes (number per 106 μm3).
TPA	Total podocyte nuclear area is computed as the cumulative area of podocyte nuclear profiles for a given glomerular unit (μm2).
GPC	Glomerular podocyte nuclear coverage is computed as the ratio of total podocyte nuclear area to glomerulus unit cross-sectional area.

2D, two-dimensional; CF, correction factor; GA, glomerulus area; GPC, glomerular podocyte nuclear coverage; GPD, glomerular podocyte density; PC, corrected podocyte count; TPA, total podocyte nuclear area.

Podocyte morphometrics are invaluable tools for prognostication. Podometric methodologies compute podocyte nuclear count, size, and spatial density, relative to glomerulus area, to provide quantitative modeling of progressive glomerular disease. These features are incorporated into PodoCount as PC, GA, GPD, TPA, and GPC to quantify podocyte depletion through image features engineered from digitized renal histopathology. All reported podometric feature values are based on 2D quantification from glomerulus profiles in whole kidney sections.

Determination of the Number of Glomerulus Profiles Required to Reliably Estimate Podocyte Density

The number of glomerular profiles required to arrive at an estimate of podocyte density within 10% of the whole-slide mean value with 90% and 95% confidence was assessed per mouse model (Supplementary Figure S4). For each mouse (WSI), the true whole-slide podocyte density was calculated. Podocyte density estimates for n = 2 to 50 randomly sampled glomeruli were then computed, for 1000 trials apiece, and compared against the true value. The probability of arriving at an estimate within 10% of the true value (n successes of 1000 trials) was recorded per sampling, and the minimum n glomeruli required to achieve 90% and 95% confidence was determined. The study was repeated for estimates within 20% of the true value.

Statistical Analysis

Data were analyzed with Minitab Statistical Software version 19 (Minitab 17 Statistical Software 2010, Minitab, State College, PA) and MATLAB’s Statistics and Machine Learning Toolbox. Feature means and SDs were reported to allow for interpretation of biological and/or clinical significance. The normality of each feature distribution was assessed with the Anderson-Darling test statistic. Nonparametric tests were used for non-normal feature distributions. Differences between groups (e.g., WT vs. transgenic mice) were assessed with unpaired, 2-sample t tests. Nonparametric tests and Welch’s correction for unequal variances were performed when appropriate.

Correlation analyses between murine histologic image features and terminal urinary albumin-to-creatinine ratio (UACR) were completed using the Spearman rank-order correlation measure (Spearman’s ρ). Spearman correlation was selected over Pearson correlation because of the non-normality of image feature data.

Using the Statistics and Machine Learning Toolbox in MATLAB, linear mixed-effects models were implemented to assess whether image features differed between disease statuses while accounting for multiple glomeruli from the same mouse kidney section. In this study, WT and transgenic disease statuses consisted of the pooled glomerulus populations across WT and transgenic mice, respectively. In each linear mixed-effects model, the image feature was used as the outcome and the model included a fixed effect for binary disease status and a random intercept term for each mouse. Parameter estimation was conducted by maximizing the restricted log-likelihood of the model. Linear mixed-effects models were selected over unpaired, 2-sample t tests (i) to account for potential clustering of glomerulus subpopulations derived from the same mouse and (ii) because linear mixed-effects models can handle missingness at random and therefore different numbers of glomeruli per mouse.

For analysis of human data, differences among 3 or more groups (e.g., DN stages) were assessed with Kruskal–Wallis nonparametric tests followed by post hoc Dunn’s tests, as the data were not normally distributed and violated analysis of variance criteria., Logistic regression was used to study ESKD outcome. First, a model was fit using only eGFR because eGFR is a benchmark outcome indicator. To assess whether image features improved the prediction of ESKD beyond eGFR alone, subsequent models were fit using a single-image feature and adjusted for eGFR.

To correct for multiple hypothesis testing, we implemented the Benjamini-Hochberg procedure, with a false discovery rate of 0.05 for all analyses. A corrected P value (q) <0.05 was considered statistically significant. The Benjamini-Hochberg procedure was selected over the Bonferroni correction because of its robustness to high volume and nonindependence of tests.

Residual and absolute error were used to compare computational podocyte count and density estimates, respectively, against manual ground truth. Departures in computational estimates compared with ground truths were visualized with Bland-Altman plots. Residual error was calculated as the difference between ground truth and computational estimates (), and absolute error was calculated as the absolute value of the difference between ground truth and computational estimates (|). Correlation between computational and ground truth estimates was assessed using Pearson’s correlation analysis (Pearson’s R).

Deployment of Whole-Slide Podocyte Analysis With Cloud Computation

HistomicsUI, a distributed system with RESTful application programming interface, was developed by Kitware (Clifton Park, NY) and was used to deploy our algorithm as a plugin, thereby creating an online platform that would enable multiple users to detect and quantify podocytes via a web interface. The algorithm was packaged in the form of a Docker image using Docker software (Palo Alto, CA),43, 44, 45 a framework that enables users to build and run applications in containers. The generated container conforms to the Slicer CLI workflow interface, which allows HistomicsUI to display a user interface to adjust algorithm parameters.

Hardware

Computational processing was performed on a Linux distribution operating system (Ubuntu 16.04) with 2 Intel Xeon Silver 4114 processors, each with 10 cores, running at 2.20 GHz and equipped with 64 GB of random-access memory. Neural network training and predictions for glomerulus boundary detection were performed using a NVIDIA Quadro RTX 5000 GPU (16 GB of memory). HistomicsUI plugin is made available for end users in a research computer with Intel i5 6-core processor, running at 3.1 to 4.5 GHz and equipped with 16 GB of random-access memory.

Data Availability

To support reproducibility, we released fully annotated pipeline codes along with sample image data, glomerulus annotation files, and all pipeline output (i.e., podocyte nuclear annotations files, feature files). We also launched our cloud-based PodoCount plugin for the end user community and created an instructional video for first-time users (Supplementary Movie S1) accessible via http://hermes.med.buffalo.edu:8080. All codes and documentation, Docker image of the web cloud interface, and data are available via http://bit.ly/3rdGPEd and our Github repository, https://github.com/SarderLab/PodoCount.

Results

Qualitative and Quantitative Performance Analysis

Visual inspection of pipeline-derived podocyte nuclear, glomerular, and tissue boundaries confirmed successful region detection and segmentation (Supplementary Figure S5). Computational performance evaluation was first completed to assess the quality of image segmentation. Segmented image regions included tissue sections, boundaries of glomerular units, and podocyte nuclei. Across all randomly sampled images/ROIs, we observed high performance for all segmentation tasks (Table 2 and Supplementary Figure S6). The median sensitivity, specificity, precision, and accuracy in tissue boundary, glomerulus boundary, and podocyte nuclear segmentation tasks were 0.99/0.99/0.99/0.99 (tissue boundary), 0.97/0.99/0.92/0.99 (glomerulus boundary), and 0.85/0.99/0.93/0.99 (podocyte nuclear segmentation), respectively (see Table 2 for additional results, including averages and ranges of each performance metric). Furthermore, the Hit-Miss analysis determined that the frequency at which podocyte nuclei were positively identified was 0.98.

Table 2

Computational performance of podocyte detection and image segmentation

Segmented region(s)	Assessment of image segmentation and podocyte detection
	Sensitivity med. (avg./[range])	Specificity med. (avg./[range])	Precision med. (avg./[range])	Accuracy med. (avg./[range])	Hit-miss percent
Tissue boundary	0.996 (0.994/0.985–0.997)	0.998 (0.998/0.996–1.000)	0.990 (0.990/0.977–0.998)	0.997 (0.998/0.996–0.999)	—
Glom boundary	0.966 (0.965/0.954–0.977)	0.999 (0.999/0.999–1.000)	0.916 (0.917/0.875–0.943)	0.999 (0.999/0.999–1.000)	—
Podocyte nuclei	0.846 (0.834/0.425–0.993)	0.997 (0.994/0.980–1.000)	0.931 (0.933/0.763–1.000)	0.997 (0.992/0.943–1.000)	0.980

Avg., average; DN, diabetic nephropathy; Glom, glomerulus; Med., median.

Sensitivity, specificity, precision, and accuracy were computed for image segmentation tasks, and percentage accuracy in podocyte nuclear detection was assessed by Hit-Miss (i.e., frequency at which podocyte nuclei were positively identified). Performance analysis for segmentation of tissues and glomerulus unit boundaries was completed for n = 12 randomly selected whole-slide images, equally distributed across data sets and disease states. Assessment of podocyte detection and image segmentation tasks was completed using n = 240 glomerulus images randomly and equally sampled from murine (n = 40 per cohort, 20 each from control and disease) and human (n = 40 per DN stage) data. High performance was observed for segmentation of tissue and glomerulus unit boundaries. Although podocyte nuclear segmentation was less sensitive and precise, podocytes were positively detected 98% of the time. Lesser performance in podocyte nuclear segmentation was attributed to the challenge of manually delineating a consistent boundary about immunohistochemistry-labeled nuclei in brightfield images.

The accuracy of podocyte count and density in PodoCount was then assessed. For each cohort, automated, continuous counts and density estimates were compared against manual ground truth. Podocyte count error was bounded by 1 podocyte per glomerulus (Table 3 and Figure 2a). Automated and manual counts were strongly and significantly correlated across all cohorts. Departure of automated counts from ground truth was greatest for the DN cohort, where histologic manifestation of disease was most pronounced (Figure 1d). Absolute errors in PodoCount density estimates were near zero (Figure 2b), and across all cohorts, automated and ground truth density estimates were strongly and significantly correlated (Table 4).

Table 3

Comparison of podocyte counts by PodoCount and the single-section method

Cohort	Comparison of PodoCount automated counts vs. the single-section method
	Median error in estimation		Pearson correlation analysis
	Residual	Absolute	R2	95% CI for R2	P value
db/db	−0.01	0.26	0.95	(0.91–0.98)	<0.001
KKAy	−0.36	0.47	0.82	(0.68–0.90)	<0.001
FSGS	−0.17	0.25	0.89	(0.80–0.94)	<0.001
HIVAN	0.06	0.25	0.97	(0.94–0.98)	<0.001
Aging	−0.28	0.37	0.96	(0.92–0.98)	<0.001
Progeroid	0.10	0.16	0.94	(0.87–0.97)	<0.001
DN	0.10	0.17	0.57	(0.35–0.57)	<0.001

DN, diabetic nephropathy; FSGS, focal segmental glomerular sclerosis; HIVAN, HIV-associated nephropathy.

PodoCount estimates of corrected podocyte counts were compared against manual ground truth generated using the single-section method. Error in automated counts was bounded by 1 podocyte. Tendency toward over- or underestimation was cohort dependent. Automated counts were strongly and significantly correlated with ground truth counts across all cohorts. Correlation results between automated and ground truth podocyte density estimates were evaluated with parametric Pearson analysis (R2 value and 95% CIs reported, in addition to P values).

Figure 2

Comparison of podometric estimates by PodoCount and the single-section method. PodoCount estimates of corrected podocyte count and podocyte density were compared against those from manual ground truth measurements using the single-section method. (a) Error in automated counts was bounded by 1 podocyte. Tendency toward over- or under- estimation was cohort dependent. (b) The modified Bland-Altman plot highlights the departure in PodoCount podocyte density estimates from ground truth. No., number; FSGS (SAND), focal segmental glomerular sclerosis; HIVAN, HIV-associated nephropathy.

Table 4

Comparison of podocyte density by PodoCount and the single-section method

Cohort	Comparison of PodoCount automated counts vs. the single-section method
	Median error in estimation		Pearson correlation analysis
	Residual	Absolute	R2	95% CI for R2	P value
db/db	−0.01	0.75	0.96	(0.93–0.98)	<0.001
KKAy	−1.88	2.18	0.92	(0.85–0.96)	<0.001
FSGS	−1.59	2.29	0.97	(0.93–0.98)	<0.001
HIVAN	0.51	1.80	0.98	(0.96–0.99)	<0.001
Aging	−0.93	1.38	0.97	(0.95–0.99)	<0.001
Progeroid	0.59	1.59	0.94	(0.88–0.97)	<0.001
DN	0.11	0.15	0.39	(0.24–0.48)	<0.001

DN, diabetic nephropathy; FSGS, focal segmental glomerular sclerosis; HIVAN, HIV-associated nephropathy.

PodoCount estimates of podocyte density were compared against manual ground truth generated using the single-section method. Absolute error in automated estimates was on average 1.55 (number per 106 μm³). Tendency toward over- or underestimation was cohort dependent. Automated estimates were strongly and significantly correlated with ground truth estimates across all cohorts. Correlation results between automated and ground truth podocyte density estimates were evaluated with parametric Pearson analysis (R2 value and 95% CIs reported, in addition to P values).

Podocyte and Glomerulus Feature Significance Across Murine Models

Quantified image features were compared based on their ability to differentiate diseased tissue from normal tissue. Statistical analysis focused on the following image features: glomerulus area, podocyte density, total podocyte nuclear area, and podocyte nuclear coverage. For each model, statistical tests compared the image features of (i) diseased mice against WT mice and (ii) the glomerulus populations of diseased mice against those of WT mice. For glomerulus-level comparisons, see Supplementary Figure S7 and Supplementary Tables S3 to S8, part B. Strength and association of image features and UACR were assessed for those models where complete UACR data were available.

db/db model. Although no podocyte feature proved significant when comparing db/db and WT mice (Supplementary Table S3), the distributions were different across all features with greater variance observed in disease (Figure 3a). In db/db mice, glomerular area and total podocyte nuclear area were greater, whereas podocyte density was less. Moderate strength of correlation (|ρ| > 0.4) was observed pairwise between these image features and UACR; with increasing UACR, glomerular and total podocyte nuclear area increased whereas podocyte density decreased (Supplementary Table S9). Nuclear coverages were similar in diabetes and had little correlation with UACR.

Figure 3

Podocyte and glomerular morphometrics of control and disease mice across murine models. Distribution of podocyte feature values across disease states with each black dot corresponding to a single mouse in the (a) db/db model of type II diabetes mellitus, (b) KKAy model of type II diabetes mellitus, (c) FSGS model, (d) HIVAN model, (e) Aging model, and (f) Progeroid (Ercc) model. All podometric values are based on 2D quantification from glomerulus profiles in whole kidney sections. ∗q < 0.05. 2D, 2-dimensional; CTRL, control; FSGS, focal segmental glomerular sclerosis; Glom =, glomerulus; HIVAN, HIV-associated nephropathy; Pod, podocyte; WT, wild type.

KKAy model. In the KKAy model of diabetes, podocyte density was significantly less, whereas glomerular and total podocyte nuclear area were significantly greater (Supplementary Table S4, part A and Figure 3b). Strong, significant correlations (|ρ| > 0.75) were also observed between these image features and UACR (Supplementary Table S10). UACR increased with increasing glomerular and total podocyte nuclear area, as well as decreasing podocyte density.

FSGS model. SAND-treated mice showed significantly lesser podocyte densities and nuclear coverages with accompanying increases in the glomerular area. Total podocyte nuclear area was similar between WT and FSGS-affected mice (Supplementary Table S5, part A). The podocyte density distributions of control and FSGS mice were markedly different, with control mice featuring a broad range of average podocyte densities and FSGS mice featuring a narrow range of reduced podocyte densities (Figure 3c). Moderate-to-strong correlations described the relationships between UACR and image features (Supplementary Table S11). With increasing UACR, significant reduction in podocyte density was observed.

HIV-associated nephropathy (HIVAN) model. Image features did not significantly differentiate Tg26 mice from WT mice (Supplementary Table S6, part A). Before correction for multiple testing, the podocyte counts of Tg26 and WT mice were significantly different as demonstrated by the marked difference in their respective distributions (Figure 3d). All other feature values and distributions were comparable. In a transgenic model of HIVAN, variable penetrance among mice and within a single animal contributes to subtler histomorphologic phenotypes.

Aging model. In contrast with young mice, old mice featured significantly greater glomerular area, as well as significantly reduced podocyte density and podocyte nuclear coverage (Figure 3e, Supplementary Table S7, part A). Total podocyte nuclear area did not significantly differ with increased age.

Progeria (Ercc1) model. Significantly lower glomerular areas and podocyte densities were observed in Ercc1 mice (Supplementary Table 8, part A and Figure 3f). The podocytes’ nuclear coverages of WT and Ercc1 mice were comparable. Before correction for multiple testing, significantly lower total podocyte nuclear areas were noted in Ercc1 mice.

For each model, a single estimate of the apparent mean caliper diameter (d), true mean caliper diameter (D), and CF was determined (Supplementary Table S12). The CFs per model were found to be 0.19 in db/db, 0.20 in KKAy, 0.22 in FSGS, 0.21 in HIVAN, 0.20 in Aging, and 0.22 in Progeroid. These CF values align well with those previously reported.

Glomerular Sampling for Accurate Podocyte Density Estimates

The number of glomerular profiles required to arrive at an estimate of podocyte density within 10% of the true (whole-slide) value with 90% and 95% confidence was studied for each model based on murine phenotype. Across all cohorts, accurate estimation in the diseased state required more glomerular profiles (Supplementary Table S13). The number of profiles required to achieve an estimate within 10% of the true value with 90% confidence for each cohort’s control and disease groups, respectively, was found to be 32 and 38 in db/db, 27 and 29 in KKAy, 30 and 37 in FSGS, 27 and 28 in HIVAN, 28 and 33 in Aging, and 35 and 44 in Progeroid, respectively. As expected, the number of required profiles increased for estimates with 95% confidence. Achievement of a 90% confident estimate with 10 or fewer profiles required relaxation of the constraint to estimation within 20% of the true value. These data demonstrate the value of a robust pipeline that assesses all the available glomeruli to eliminate model-specific variability and potential sampling bias.

Podometrics in Clinical Human DN Biopsies

Biopsy-level features were compared among DN subjects based on their Tervaert classification (Table 5) and outcome (Table 6 and Figure 4a and b). Corrected podocyte count was the lead indicator of DN stage at the patient level, with a marked reduction in glomerular podocyte number (from 1.4 to 0.52) defining the transition from DN stage IIb to III (part B in Table 5). Total podocyte nuclear area and podocyte nuclear coverage were also significantly different across DN stages and were characteristic of the transition from stage IIb to III. Meanwhile, significantly greater glomerular area differentiated stage IIb from IV. These observations are consistent with established histopathologic classification criteria, and suggest a relationship with progressive podocyte injury, glomerular hypertrophy, and transition between DN stages. Single estimates of d, D, and CF were also determined for the human DN cohort (Supplementary Table S12). The CF value representative of the entire data set was 0.22. This human CF value is comparable to what has been previously reported.

Table 5

Podocyte morphometrics significantly differentiated diabetic nephropathy stages IIb and III

Image feature	Diabetic nephropathy cohort(A) Summary of feature values across patients’ DN stages according to the Tervaert classification schemeMean ± SD
	Stage I	Stage IIa	Stage IIb	Stage III	Stage IV
n	3	6	12	4	20
PC	1.09 ± 0.24	1.15 ± 0.72	1.40 ± 0.56	0.52 ± 0.31	0.85 ± 0.51
GA	29660.26 ± 4644.71	10085.43 ± 16639.20	31116.08 ± 8381.64	22629.21 ± 4758.69	23256.96 ± 9121.63
GPD	16.82 ± 3.85	19.96 ± 9.79	19.89 ± 6.31	12.55 ± 8.10	17.85 ± 7.91
TPA	95.76 ± 30.14	99.69 ± 70.29	155.42 ± 98.43	28.34 ± 27.70	84.94 ± 86.35
GPC	3.09 ± 0.65	2.82 ± 1.44	4.14 ± 2.09	0.98 ± 0.80	2.73 ± 1.66

DN, diabetic nephropathy; GA, glomerular area; GPC, glomerular podocyte nuclear coverage; GPD, glomerular podocyte volume density; n, number of unique patient thin-needle biopsies per DN stage; PC, corrected podocyte count; TPA, total podocyte nuclear area.

Pipeline-computed features were ranked based on their ability to differentiate between DN stages defined by the Tervaert classification scheme. (A) The table summarizes the mean and SD of each feature across Tervaert stages. (B) Corrected podocyte count was the most significant indicator of disease in DN, followed by total podocyte nuclear area and podocyte nuclear coverage. Pairwise tests revealed that differences in podocyte counts and nuclear morphometrics are consistently observed between DN stages IIb and III. Podometric values are based on 2D quantification from glomerulus profiles in whole kidney sections. Feature units: GA, μm2; GPD, number per 106 μm3; TPA, μm2. Statistical conclusions are based on Kruskal–Wallis and post hoc Dunn’s tests comparing population medians at a significance level of 0.05. H0: At least one population mean is different.

q < 0.05.

Table 6

Biopsy nuclear podocyte pathology is predictive of progression to end-stage kidney disease

Image feature	Diabetic nephropathy cohortTwo sample t test comparing nonprogressors (n = 31) against progressors to ESKD (n = 14)
	Feature summary (mean ± SD)		Difference of means (ESKD – no ESKD)		q value
	No ESKD	ESKD	Difference	95% CI
PC	3.44 ± 1.93	1.97 ± 1.05	−1.47	(−2.37 to −0.57)	0.002a
GA	28432.41 ± 9450.26	21426.01 ± 5384.96	−7006.40	(−11502.52 to −2510.29)	0.003a
GPD	19.22 ± 7.52	15.73 ± 7.38	−3.49	(−8.40 to 1.43)	0.156
TPA	124.26 ± 95.79	50.74 ± 29.88	−73.52	(−111.85 to −35.18)	<0.001a
GPC	3.01 ± 2.10	2.21 ± 1.31	−0.80	(−2.00 to 1.31)	0.002a

ESKD, end-stage kidney disease; GA, glomerular area; GPC, glomerular podocyte nuclear coverage; GPD, glomerular podocyte density; PC, corrected podocyte count; TPA, total podocyte nuclear area.

The table summarizes the mean and SD of each feature across patient outcomes. Significant reduction in corrected podocyte count, glomerular area, total podocyte nuclear area, and glomerular podocyte nuclear coverage was characteristic of progressor biopsies. Podometric values are based on 2D quantification from glomerulus profiles in whole kidney sections. Feature units: GA, μm2; GPD, number per 106 μm3; TPA, μm2. Statistical conclusions are based on 2-sample t tests comparing murine population means at a significance level of 0.05.

q < 0.05.

Figure 4

Podocyte and glomerular morphometrics in diabetic nephropathy kidney biopsy specimens predict outcome. Distribution of podocyte feature values comparing those with diabetes with progression to ESKD to those without with each black dot corresponding to a single patient (a) or glomerulus (b). All podometric values are based on 2D quantification from glomerulus profiles in whole kidney sections. ∗q < 0.05. 2D, 2-dimensional; ESKD, end-stage kidney disease; Pod, podocyte; Glom, glomerulus.

Podometrics Predict ESKD in DN

A logistic regression model fit using the DN cohort’s eGFRs at time of biopsy significantly predicted ESKD with an odds ratio of 0.75 (Table 7). Each additional unit increase in eGFR was associated with a 25% decrease in the odds of patient progression to ESKD, a finding supported by clinical standards that uphold eGFR as a key indicator of outcome. A series of logistic regression models each fit using an engineered image feature with adjustment for eGFR demonstrated that glomerular area and total podocyte nuclear area were also significant predictors of ESKD. For each 1000 μm2 increase in mean biopsy glomerular area, the odds of ESKD was increased by 43%. Similarly, each 10 μm2 increase in biopsy total podocyte nuclear area was associated with a 25% decrease in the odds of ESKD. Odds ratio-associated P values also suggested that, together, total podocyte nuclear area and eGFR were more significantly related to ESKD incidence than eGFR alone. When adjusted for eGFR, podocyte count, podocyte density, and glomerular podocyte nuclear coverage were not found to be significant predictors of ESKD (odds ratios 0.91, 0.80, and 0.89, respectively). These studies highlight the potential for encoded features in histomorphology to increase the predictive power of clinical metrics.

Table 7

Nuclear indicators of podocyte pathology may improve patient prognostication from time of biopsy

Image feature (adjusted for eGFR)	Diabetic nephropathy cohortLogistic regression for prediction of patient progression to ESKD (n = 14 ESKD, 31 no ESKD)
	eGFR			Feature
	OR	95% CI	P value	OR	95% CI	q value
eGFR	0.75	0.58–0.96	0.020a	—	—	—
eGFR + PC	0.54	0.23–1.29	0.167	0.91	(0.83–0.99)	0.285
eGFR + GA	0.74	0.57–0.96	0.021a	1.43	(0.34–6.06)	0.023b
eGFR + GPD	0.69	1.01–1.78	0.047a	0.80	(0.54–1.20)	0.389
eGFR + TPA	0.74	0.58–0.96	0.022a	0.75	(0.60–0.95)	0.018b
eGFR + GPC	0.74	0.58–0.95	0.022a	0.89	(0.31–2.61)	0.836

eGFR, estimated glomerular filtration rate at time of biopsy; ESKD, end-stage kidney disease; GA, glomerulus area; GPC, glomerular podocyte nuclear coverage; GPD, glomerular podocyte density; OR, odds ratio; TPA, total podocyte nuclear area.

Logistic regression models were evaluated for feature-based prediction of patient outcome in diabetic nephropathy. Each pipeline-computed image feature was evaluated in combination with eGFR as predictors of ESKD incidence (response variable). Response frequency was 14 of 45 patients. The values and 95% CIs for odds ratios were reported, in addition to their associated P values. Select image features were rescaled to provide interpretable unit changes in OR, including GA by 1000, GPD by 10, and TPA by 10. Feature units: GA, μm2; GPD, number per 106 μm3; TPA, μm2.

q < 0.05.

PodoCount in the Cloud

PodoCount was deployed as a cloud-based plugin on the Sarder Laboratory’s Digital Slide Archive (Supplementary Figure S8). This integration was facilitated by HistomicsTK, a web-based tool that allows for installation of user-defined algorithms as plugins in a virtual user interface, HistomicsUI, and is supported by the OpenSlide library for handling proprietary digital pathology WSI formats. PodoCount end users need only upload a WSI and glomerulus annotation file and select the option for PodoCount analysis. Predicted podocyte nuclear annotations are displayed in the cloud and available for download (.xml. for standard desktop pathology viewers) along with image features files (.csv). An instructional video for end users was shared along with open-source documentation and codes to maximize plugin accessibility (Supplementary Movie S1, http://bit.ly/3rdGPEd).

Discussion

In this work, we introduced PodoCount, a novel tool for automated, whole-slide assessment of podocyte depletion and nuclear morphometry. Compared with existing automated methods,, PodoCount takes a deterministic approach to podocyte quantification, using classical image analysis in place of deep learning. Benefits of this approach include computational simplicity, strong performance without big-data requirements, and thus no need for expensive GPUs or training time. PodoCount’s computational image analysis approach was informed by the benchmark method for estimation of podocyte density from a single histologic section developed by Venkatareddy et al. We validated this tool using WSIs from 6 distinct mouse models (n = 135) and human DN (n = 45) biopsy specimens from male and female mice and participants. These data were curated from multiple institutions and feature lab-to-lab technical variability, aspects that are known to hinder generalizable computational frameworks,52, 53, 54, 55, 56 including highly variable sectioning and staining methods, as well as image and tissue quality (Supplementary Figure S2). PodoCount navigated this challenging data set well, achieving precise segmentation and accurate detection of podocyte nuclei (Table 2), as well as highly accurate podocyte density estimates (Tables 3 and 4 and Figure 2). When paired with strategic, literature-informed, feature engineering, this computational performance facilitated robust quantification of podocyte depletion across varied renal pathologies using a podocyte nuclear marker and nuclear-based quantification from 2-dimensional cross-sections.

Our nuclear-based approach was informed by prior studies which demonstrated that podocyte nuclear quantification is sufficient to estimate glomerular podocyte density and predict podocyte pathology., Emphasis on nuclear quantification guided our choice of podocyte label. p57kip2 was a podocyte-exclusive nuclear label within the glomerular microenvironment enabling reproducible detection amidst proximate cell populations. Two-dimensional analysis was opted for over 3-dimensional podocyte visualization to maximize feasibility and efficiency while maintaining accuracy in podocyte estimation. Prior studies validated that podocyte estimation from single sections is comparable to 3-dimensional, multisection methods, irrespective of disease state.,,14, 15, 16 Murine data CFs and podocyte density estimates computed using PodoCount aligned well with the literature,,,57, 58, 59 as did human CF values. Podocyte density estimates for human DN data were slightly lower than those values reported for prior studies in early diabetes and clinical nephropathy.

According to the podocyte depletion hypothesis, podocyte depletion may manifest (i) absolutely, as a reduction in glomerular podocyte count, or (ii) relatively, when pathologic increase in glomerular area reduces podocyte spatial density. In this work, we found that computational podocyte features reproducibly predict disease-specific histopathology. Although the db/db model emulates early changes in human DN, the KKAy model, renowned for rapid glomerular basement membrane thickening, recapitulates the later, morphologically advanced stages of DN. Quantified feature trends in the db/db and KKAy models reflected these tendencies toward mild and advanced diabetic changes, respectively. Similarly, computed image features from the aging and progeroid models aligned well with our current understanding of glomerular senescence. Studies demonstrated that with age relative podocyte depletion from compensatory podocyte hypertrophy peaks at a 2.5-fold increase, giving way to absolute podocyte loss and progressive glomerulosclerosis. Observed feature trends in the aging and progeroid models aligned well with the literature, with marked increase in total podocyte nuclear area, in the presence of reduced podocyte volume density, reflecting peak podocyte hypertrophic capacity, followed by overt podocyte loss.

The postadaptive FSGS phenotype is characterized by glomerular hypertrophy and podocyte loss, and is best described as a combination of absolute and relative podocyte depletion. Computed, significant decrease in podocyte density, as well as increase in glomerular area, for SAND-treated mice in the FSGS model was consistent with FSGS pathology. Similarly, quantified podocyte loss in Tg26 mice was consistent with HIVAN pathology.,

From the human DN cohort, we learned that podocyte nuclear image features have the potential to be valuable diagnostic and prognostic tools. Podocyte nuclear metrics differentiated patient biopsy specimens (n = 45) according to Tervaert class, highlighting the transition from stage IIb to III as a key turning point in DN pathology (Table 5) and outcome (Table 6 and Figure 4). Intriguingly, logistic regression analysis demonstrated that podocyte morphometrics derived from patient biopsy specimens have the potential to improve prediction of ESKD beyond eGFR alone (Table 7). Total podocyte nuclear area was more significantly associated with ESKD than eGFR alone, and together, these 2 measures significantly predicted progression to ESKD. Glomerular sampling studies for reliable estimation of podocyte density were not completed for human data wherein the number of glomerulus profiles per patient biopsy was limited (Supplementary Table S14). We recognize that the number of human samples is a limitation of the study and emphasize that our findings warrant future studies with greater statistical power.

As the irreversible end point of chronic kidney disease, ESKD is characterized by complete loss of kidney function and patient dependence on dialysis or transplant for survival. The presence of significant feature-response relationships underscores the potential for biopsy podocyte features to increase the precision of clinical metrics in chronic kidney disease prognostication, and thus improve patient outcomes. With potential to augment both experimental and clinical workflows, PodoCount was launched as an open-source cloud-based tool to maximize accessibility and promote standardization of podocyte morphometrics.

Disclosure

LJN is the co-founder of NRTK Biosciences, a startup to develop novel senotherapeutics.