Development of an automated estimation of foot process width using deep learning in kidney biopsies from patients with Fabry, minimal change, and diabetic kidney diseases

Abstract

Podocyte injury plays a key role in pathogenesis of many kidney diseases with increased podocyte foot process width (FPW), an important measure of podocyte injury. Unfortunately, there is no consensus on the best way to estimate FPW and unbiased stereology, the current gold standard, is time consuming and not widely available. To address this, we developed an automated FPW estimation technique using deep learning. A U-Net architecture variant model was trained to semantically segment the podocyte-glomerular basement membrane interface and filtration slits. Additionally, we employed a post-processing computer vision approach to accurately estimate FPW. A custom segmentation utility was also created to manually classify these structures on digital electron microscopy (EM) images and to prepare a training dataset. The model was applied to EM images of kidney biopsies from 56 patients with Fabry disease, 15 with type 2 diabetes, 10 with minimal change disease, and 17 normal individuals. The results were compared with unbiased stereology measurements performed by expert technicians unaware of the clinical information. FPW measured by deep learning and by the expert technicians were highly correlated and not statistically different in any of the studied groups. A Bland-Altman plot confirmed interchangeability of the methods. FPW measurement time per biopsy was substantially reduced by deep learning. Thus, we have developed a novel validated deep learning model for FPW measurement on EM images. The model is accessible through a cloud-based application making calculation of this important biomarker more widely accessible for research and clinical applications.

LAY SUMMARY

Podocyte injury is critical to pathogenesis of many kidney diseases. Increased podocyte foot process width (FPW) is an important measure of podocyte injury and dysfunctional glomerular filtration barrier. However, precise direct measurement of FPW is time consuming and not widely available. We developed a deep learning model to automate FPW measurements. The deep learning results in kidney biopsies from patients with Fabry disease, type 2 diabetes, minimal change disease and controls were interchangeable with gold standard measurements performed by expert technicians while the time needed for the measurements was reduced from several hours to less than one minute per biopsy. Our model is accessible through a cloud-based application. This approach can make FPW measurement more widely accessible for research and clinical applications.

Graphical Abstract

Podocytes play a key role in the maintenance of glomerular structure and filtration barrier.¹ The extent of podocyte injury typically correlates with the severity of urinary protein excretion and chronic kidney disease (CKD) progression in various settings.² Since podocytes have limited regenerative capacity, their injury and loss leads to cumulative podocyte depletion and glomerulosclerosis. ^3–5 Accurate quantification of podocyte injury may be of prognostic significance and could improve treatment decisions. Increased foot process width (FPW) is a common feature of podocyte injury and can occur before microalbuminuria or proteinuria. ^6–9 If properly quantified, average FPW correlates with severity of urine albumin and protein excretion rate. ^6,10 Currently, there is no consensus on how to estimate the average FPW. While the unbiased stereological approach is considered the gold standard, ^10–13 it is resource and time consuming, technically complex, and not widely available. Recent advances in artificial intelligence have made this field increasingly available for image analysis. Deep learning (DL) is already revolutionizing light microscopy digital pathology. ^14–17 However, application of these approaches to EM pathology is limited. ^18,19 We developed DL methods to automate FPW estimation on EM images. The method was developed in kidney biopsies from patients with Fabry disease and secondary focal segmental glomerulosclerosis (FSGS), conditions known for podocyte injury ^6,20,21 and from living kidney donors, and tested the method in diabetic kidney disease (DKD) and minimal change disease (MCD).

METHODS

A detailed version of Methods is provided in Supplementary Methods.

Subjects and Clinical Information

Kidney biopsies from patients with Fabry disease, type 2 diabetes, and MCD were performed either as part of clinical trials, or standard of care.^22–24 Research kidney biopsies from living kidney transplant donors were studied as controls. Written informed consent was provided by all participants with Fabry disease, diabetes, or kidney donors and/or their designees for the research use of these tissues. IRB waiver of consent was obtained to include clinical kidney biopsies from patients with MCD.

Kidney Biopsy Studies

Kidney biopsies were processed at eight electron microscopy laboratories. One μm sections of 2.5% glutaraldehyde fixed plastic embedded tissues were stained with toluidine blue for identification of glomeruli. Random glomerular sections were prepared for stereological studies. ⁶ Overlapping digital images (~8,000 x) of entire glomerular profiles were obtained using a JEOL 1010 EM for masked review by two observers (BN and MM) to select 2–6 (median = 3) non-sclerosed glomeruli per biopsy with no major artifacts. High magnification (~30,000 x) images were obtained using a systematic uniform random sampling (SURS) strategy²⁵ for estimation of average FPW by expert technicians (ET) masked for the clinical information and using stereology.^6,10,13 Additional structural parameters relevant to podocyte injury, Fabry nephropathy or DKD were quantified as described elsewhere. ^{6 26 22}

The Training Set and Ground Truth

The same images used for FPW stereological measurement were used for DL studies. The training dataset was created on 820 EM images from kidney biopsies from participants with Fabry disease and controls. Generation of the training dataset involved three major steps: annotation, normalizing, and augmentation. To annotate the dataset, we developed a custom multi-class semantic segmentation utility which we named “the UW Segmentation Utility” using wxPython.²⁷ This utility allowed the manual and assisted classification of the podocyte-glomerular basement membrane interface (PGBMI) and foot process slit pores (slits) (Figure S1). In total, 7342 slits and 8,152,303 nm of PGBMI were manually segmented in the training set. Each segmentation category was assigned a numerical label which was mapped to a 3D voxel, allowing for overlapping segmentation labels. Once manually segmented, the images and segmentations were exported to a hdf5 dataset,²⁸ and processed using Dask²⁹ for normalization. Down-sampling was performed on each input image using lanczos4³⁰ [8×8 pixel region] and the output voxel stack using nearest neighbor. Finally, the images were standardized by subtracting the mean and dividing by the standard deviation, resulting in a distribution centered at the mean.

Model Design

The processes involved in designing the model are summarized in Figures 1 and 2. A U-Net³¹ architecture was designed to detect both PGBMI and the slits in two separate layers (Figure 1). The network was designed with two pooling layers before the fork and one shared branch on the 7th upscale layer. Each pooling used a factor of 2 filters starting from 32 for the first layer (with a 3×3 kernel). ReLU^32,33 activation was used for all, except the last layers where sigmoid was used on the two output branches. This enabled the isolation of decoder branches for specific classes, as there were multiple distinct branches for both the slit and PGBMI classes. The Convolutional Neural Network (CNN) model branches labeled each pixel either as background or category [i.e. slit or PGBMI]. The slit branch had a skip connection from the PGBMI branch in the 7th upscale layer, in the decoder. Due to the memory requirements of the model, a gradient accumulation technique was employed. This involved accumulating gradients over a few batches before taking an optimization step. In our case, the target batch size was set to 16 and desktop trained used a 2 batch size [physical] (8 accumulation rate) for 150 epochs (~123,000 samples). Additionally, a learning rate (LR) scheduler reduced LR, initially at 18e-4, by a factor of ~3 when training dice plateaued. Lastly, the calculation of the foot process width (FPW) was performed in a post-process script. This script was designed to analyze the segmented output and estimate the FPW based on the obtained results from the deep learning model.

Figure 1. Model of the ForkNet Architecture.

Top: A schematic representation of the architecture showing an input EM Image to the left, conv2d blocks max pooling for downscaling, split branches, deconv2d and conv2d blocks for upscaling, and output layers to the right. Layer connections are identified as solid directional lines, and skip connections are identified as dashed directional lines. Note: the skip connection from the membrane branch to the slit branch. Most residual connections used convolutional blocks rather than identity connections. Bottom: A-E and F-J show two examples of input images (A and F), resulting in corresponding PGBMI (B and G) and slit masks (C and H) with merged images, showing the masks superimposed on the input images (D and I); and the post-processing output with individual foot process width (FPW) measurements (E and J). F’-J’ show magnified views of the red boxes in corresponding F-J images. The size of the dots in the slit mask and merged images reflects the model confidence in prediction of a slit. The scale bar in A, F, and F’ represents 1 μm.

Figure 2. ForkNet post-processing script general workflow.

Top left image: a low magnification view of a glomerulus. The small squares represent systematic uniform random sampling (SURS) images taken at higher magnification (30,000 x) shown in bottom left that are used for foot process width (FPW) measurements. The flowchart on the right shows the post-processing workflow. The input images are normalized and go through various filters and processing, the output of which will be semantic segmentation of filtration slits and podocyte – glomerular basement membrane interface (PGBMI). The length of the individual segments (PGBMI limited between two adjacent slits) is measured and exported. Different components of the architecture are shown in boxes coded with different shapes and shades that are connected through solid or dashed lines as explained in the “Key” box at the bottom left corner. In summary, from left to right, the input image goes through conv2d downscaling blocks, split branches, and deconv2d upscaling to result in PGBMI and slit output masks. ROI: region of interest. The scale bar in the bottom high mag image represents 1 μm.

Model Training

Several standard techniques were employed to mitigate overfitting and enhance validation accuracy. The first was image augmentation (random scaling, rotation, flipping, contrast changes, skew, translation, and artifacts) aimed to introduce variability and enhance the model’s ability to generalize. To evaluate overfitting, a random subset comprising 10% of the dataset was reserved for validation tests. Training involved a semi-online learning approach, where a small sample of manually labeled images was used to train the model to assist technicians with predictions to produce a larger sample. This iterative process continued until the training dataset was fully utilized. Then, the model was retrained fully to capture the highest validation dice similarity coefficient (DSC).

Compiling the Package and Cloud-Based Utility

Once the model was trained successfully on the complete dataset, all necessary components were compiled into a Docker³⁴ image to facilitate easy distribution and deployment. A cloud-based web utility along with a dataset uploader was developed to support running the model on machines that are not capable of running the model locally and provide researchers and users an easy way to generate reports. The web utility, currently in beta, is accessible at https://fpwdl.smerkous.com. Instructions for downloading and processing the datasets used in this paper may be found at https://fpwdl.smerkous.com/instructions. For more specific details regarding the model evaluation please refer to the public repository listed below.

Workflow and Post-Processing

The DL model workflow, including the post-processing steps is summarized in Figure 2. The predicted segmentation layers are processed in order of layer dependence, where the slit layer depends on the PGBMI layers. A post-processing script was developed using the python OpenCV, ³⁵ Mahotas, ³⁶ and Numpy³⁷ frameworks to make associations between the slits and independent PGBMI segments. All extraneous noise and small segments were removed to only leave valid PGBMI segments. The PGBMI masks were skeletonized, contoured, and segments with breaks that are nearly aligned were reconnected to make a continuous line. Using a hit-miss transform for branch features, any branching occurring in the PGBMI would be broken into individual segments to prevent measuring FPW across multiple branches. ³⁶ Before combining, the points were ordered by randomly picking a segment end, using a branch end hit-miss detection, then ordering the slits along each segment’s contour from that end. The distance on the PGBMI between each slit segment was measured by calculating the arc-length of that specified PGBMI segment contour. The uploaded images can be viewed within the website prior to processing, where the images can be individually examined and excluded from analysis using a toggle button if affected by major artifacts. In addition, the images with superimposed segmentations can be viewed to observe the accuracy of the model on each image individually.

Assessments of Model Performance and Validation

The model performance was analyzed using DSC. ³⁸ Each manual segmentation mask, per category, was compared to the prediction mask of the model. DSC was calculated by dividing the intersection by the total area of the masks, with an additional smoothing constant of 1 $DSC=\frac{2|X\cap Y|+1}{|X|+|Y|+1}$ for each category. ³⁸ DSC values range from 0 to 1, where 0 represents no pixel overlap and 1 represents perfect pixel overlap between the two masks. In addition, DL FPW measurements were compared with ET measurements.^6,10

Statistical Methods

Statistica 13.0 (TIBCO Software Inc.), R-4.3.0 (The R Foundation for Statistical Computing), and Graphpad Prism 9.5.1 (Graphpad Software, LLC) were used for statistical analyses and generation of graphs. Paired DL and ET FPW measurements were compared using the sign test (nonparametric) in each group. DL or ET measured FPW values across the groups were compared using the Kolmogorov-Smirnov test. Relationships between parametric variables were assessed using Pearson correlation. DL and ET FPW measurements interchangeability was assessed using Bland-Altman plots with limits of agreement of ±1.96 SD from the average difference between the two methods. Asymptotic test for the equality of coefficients of variation from multiple groups³⁹ was used to compare inter-glomerular dispersion of FPW values among the groups. To identify the minimum number of images needed to obtain a robust average of FPW, incremental image sampling with bootstrapping was used. p ≤ 0.05 was considered statistically significant.

RESULTS

Subject Characteristics

Subjects with Fabry disease consisted of 39 males and 16 females with no prior Fabry-specific treatment. Seven patients had GLA mutations associated with the later-onset Fabry disease phenotype and the remainder had classic GLA mutations. Their age was 33 [5–68] years, median [range] years. Mean glomerular filtration rate (GFR) was 114 ± 37 (range 45–289) ml/min/1.73 m². Urine protein creatinine ratio (UPCR) was 182 [0–3380] mg/g. Subjects with type 2 diabetes were 6 males and 9 females with age 48 [38–69] years, GFR 144 ± 62 ml/min/1.73 m², and urine albumin creatinine ratio (UACR) 41 [5–2991] mg/g. Subjects with MCD included 6 males and 4 females aged 7 [1–16] years. MCD diagnosis was established clinically with pre-treatment biopsy confirmation; however, renal function data for these patients were not available. The control subjects were 16 males and one female aged 41 [3–65] years.

Model Performance and Validation

The model training accuracy measured against manual segmentation was 61–67% (15% SD) for slits and 76–86% (7.5% SD) for PGBMI. Model precision performance on segmentation of 6960 filtration slits (i.e. positive predictive value) was 91%, with a recall (i.e. sensitivity) of 78%, and an F-score (the harmonic mean of precision and recall) of 0.84. DSC values for various training groups are shown in Table S1. Lower DSC values were associated with inclusion of more suboptimal quality images in the dataset. The time required to estimate average FPW in a biopsy was substantially reduced (~45 seconds to complete ~90 source images on an NVIDIA GTX 1080) compared to an ET (6–8 hours). DL errors were mostly related to smaller datasets, low contrast PGBMI regions, or occasionally mistaking inclusions or endothelial cells with PGBMI or slits. However, since post-processing alignment of slits on the PGBMI was necessary for FPW measurements, semantic segmentation errors involving misalignment of these masks did not affect the values. After compilation of the model and utility in one executable package, it was tested on 32–326 (median = 101) SURS images per biopsy, which had not been included for training, from the same 51 subjects.

Interchangeability of Deep Learning and Expert Technician Stereological Measurements of Foot Process Width

DL and ET FPW measurements were not statistically different, and were directly correlated in any of the groups studied (Table 1 and Figure 3A). Eighty-seven to 100% of the differences between respective FPW values measured by DL or ET fell within limits of agreement (LOA) defined as ± 1.96 SD of their average difference on Bland-Altman plots, indicative of overall interchangeability between the methods (Table 1). When all groups combined, there was a strong correlation between FPW measured by DL and ET (r=0.92; p<0.0001) with the regression line almost being exactly superimposed on the line of identity (Figure 3B) and 94% of the differences falling within the LOA (Figure 3B). Together, these results confirm overall interchangeability between the two methods.

Table 1.

Comparisons and correlations of deep learning (DL) and expert technician (ET) measured FPW values

Group	N	FPW (nm)	ET FPW (nm)	p-value	Correlation (r; p-value)	Bland-Altman (% within LOA)
Controls	17	663 ± 106	608 ± 92	ns	0.73; 0.001	94%
Female Fabry	16	705 ± 124	657 ± 116	ns	0.76; 0.001	94%
Male Fabry	39	793 ± 144	776 ± 213	ns	0.80; <0.0001	95%
DKD	15	1037± 136	1052 ± 550	ns	0.70; 0.004	87%
MCD	10	3613 ± 1406	3858 ± 1120	ns	0.65; 0.04	100%

Figure 3. Interchangeability of deep learning (DL) and expert technician (ET) measurements of foot process width (FPW).

(A) Comparison of DL (circles) and ET (triangles) FPW measurements in control subjects (grey), males with Fabry disease (red), females with Fabry disease (green), diabetic kidney disease (DKD; blue), and minimal change disease (MCD; pink). The corresponding violin plots (grey) are shown in the background of individual values. Dashed lines in the violin plots show means and dotted lines show SD. P-values of DL vs. ET measurement comparisons are shown above the violin plots (ns = non-significant; asterisk = both DL or ET values are statistically significantly different from controls); (B) Simple linear correlation between DL and ET FPW measurements in the entire cohort. The regression line (dashed red) and the line of identity (solid black) are almost superimposed (R=0.92; p<0.0001). The black dotted lines show the 95% confidence limit of the regression line; (C) Bland-Altman graph plotting the averages of FPW measurements by DL and ET vs. the differences between the values measured by the two methods in the entire cohort. The middle dotted line shows the bias of DL measurements (in nm) and the two other dotted lines show the upper and lower limits of agreement defined by ± 1.96 SD. A color code key is provided on the right; (D-H) Representative electron micrographs (~30,000x) from biopsies from the various groups shown in the above graphs.

We also compared relationships found between FPW measured by ET or DL in male and female patients with Fabry disease and in persons with DKD with other relevant clinical and structural variables (Tables S2 – S7). In male Fabry disease patients, both DL and ET measured FPW correlated directly with age, UPCR, podocyte volume and podocyte GL3 volume [V(Inc/PC)]; However, only DL FPW correlated inversely with podocyte numerical density and directly with podocyte GL3 inclusion volume fraction; and only ET measured FPW correlated with mesangial cell GL3 inclusion volume fraction. In female Fabry disease patients, only DL FPW directly correlated with UPCR, podocyte volume and V(Inc/PC) and inversely with % non-Fabry phenotype podocytes. In the DKD group, both DL and ET FPW correlated directly with mesangial matrix volume fraction. On the other hand, only DL FPW directly correlated with UACR, mesangial volume fraction, mesangial cell volume fraction and podocyte volume, and inversely with glomerular filtration surface density [Sv(PGBM/glom)].

Intra and Interglomerular Variability of Foot Process Width in Fabry Disease and Normal Controls

Objective assessment of segmental vs. diffuse foot process effacement is a common element of diagnostic Renal Pathology^11,40. Our DL approach provided detailed information about the distribution of FPW in the glomeruli. Individual FPW values in controls ranged from 70 – 3030 nm with the majority falling between 200–300 nm (Figure 4). We arbitrarily defined FPW ≥ 2000 nm (~3 fold greater than in the control group average FPW) as “prominent widening” and included all such instances into one histogram bin (Figures 4A–E). This approach provided quantitative correlates of “segmental” vs. “extensive” foot process widening (Figures 4F–G). Thus, only a tiny fraction (~0.4%) of FPWs in the control group were prominently widened, whereas this category included ~3% of FPWs in the Fabry female group, ~8% in each of the Fabry male and DKD groups, and ~46% in the MCD group. Of note, the FPW histogram showed a clear shift to the right in MCD. Interglomerular variability of average FPW was not statistically different among the groups and the range of average FPW in the biopsies when normalized for the biopsy average value was comparable among the groups (Figure 4H). Likewise, FPW coefficients of variation were not statistically different among the groups (Figure 4I).

Figure 4. (A-E) Comparison of individual foot process width distributions in Fabry disease, DKD, MCD, and control subjects.

Histogram of aggregate individual foot process widths (FPWs) in biopsies from (A) control subjects, (B) female patients with Fabry disease, (C) male patients with Fabry disease, (D) subjects with diabetic kidney disease (DKD), and (E) patients with minimal change disease (MCD). The gray histogram of individual FPW distribution of controls is added to the background of each of the other categories for comparison. All FPW ≥ 2000 nm are aggregated in the last bucket, with the number above the last bar representing the %FPW measurements ≥ 2000 nm. (F) Intact foot processes in a control biopsy. (G) Segmental foot process widening (red arrowheads) in a biopsy from a patient with Fabry disease. The blue lines in F and G show podocyte glomerular basement membrane interface and the yellow dots show the slit diaphragms segmented by DL. The scale bars represent 1 μm. (H) Interglomerular variability average FPW among the groups studies (gray: control; green: Fabry female; red: Fabry male; blue: DND; and purple: MCD). Each circle represents average FPW in a glomerulus. Each vertical line represents a biopsy. Small black dots represent medians. (I) comparison of coefficient of variation (CV) of interglomerular average FPW per group.

The Minimum Number of Electron Microscopy Images Required for Robust Estimates of Foot Process Width

Examination of individual FPW distributions confirmed that foot process widening was not a uniform phenomenon within a glomerulus (Figures 4A–E). SURS, used in this study, is an efficient method to obtain unbiased representative sampling. ²⁵ To identify the minimum number of EM images needed for a robust average FPW estimate, graphs of estimated FPW convergence were created for each condition. The empirical cumulative distribution function (CDF) of the event is shown in Figure 5A–E. The majority (85%) of biopsy FPW pixel averages converged at approximately 69 images for controls and DKD, 85 images for Fabry Males, 94 images for Fabry Females, and 117 images for MCD. However, for practical purposes, 50–60 images provided adequately stable average FPW values across the groups.

Figure 5. Convergence of average FPW by incremental sampling of images in biopsies from (A) controls, (B) females with Fabry disease, (C) males with Fabry disease, (D) DKD, and (E) MCD.

X-axis: number of input EM images used to obtain average FPW; left Y-axis: biopsy FPW pixel average represented in the colored graphs (shades of green to purple); right Y-axis: biopsy convergence cumulative distribution function (CDF) represented in the stepped line (grey). The vertical blue dashed line shows the number of images resulting in a CDF = 0.85.

DISCUSSION

This is the first study demonstrating the successful application of DL in quantitative assessment of kidney EM. The ForkNET model produced accurate estimates of the average FPW compared with those provided by ET using stereological methods. DL and ET measured FPW were overall interchangeable. While our DL cloud-based application provides visualization of model predictions, tracking the measurement process of previously analyzed images by ET is not feasible. Thus, a precise explanation for small differences observed between the two methods cannot be offered. DL FPW values showed relationships with relevant demographic, clinical or structural variables in male and female Fabry disease and DKD patients. Importantly, in female subjects with Fabry disease only the DL measurements correlated directly with UPCR, podocyte volume (a measure of podocyte hypertrophy) and V(Inc/PC), and inversely with % non-Fabry phenotype podocytes. The latter finding is in line with a prior study, where the presence of more non-Fabry podocytes was associated with lower FPW, suggestive of lower FPW in the non-Fabry podocytes or a podocyte protective effect related to the mosaicism status. ⁴¹ Likewise, in male Fabry disease patients, the findings of a direct relationship between age and FPW and an inverse relationship between FPW and podocyte number density was limited to DL measurements. Moreover, in DKD patients, DL FPW measurements and not the ET measurements were directly correlated with UACR, mesangial volume fraction, mesangial matrix volume fraction, mesangial cell volume fraction and podocyte volume and inversely with Sv(PGBM/glom), supportive of equal or greater biological relevance for DL vs. ET FPW measurements.

Widening of foot processes is a morphological manifestation of podocyte injury which correlates closely with proteinuria/albuminuria in a wide variety of renal disorders. ^13,42–45 Increased FPW has been documented in patients with Fabry disease with no increased albuminuria or proteinuria, suggesting that this is a sensitive marker of podocyte injury which occurs before glomerular leakage of proteins exceeds tubular reabsorption capacity. ^{6,7,20,41,46–49} The magnitude of GL3 reduction in podocytes in male patients with Fabry disease and amenable GLA mutations following treatment with migalastat strongly correlated with the reduction in FPW, consistent with a reduction in podocyte injury. ⁴⁷ The results of the present study support the clinical applicability of this DL approach in podocyte injury assessment in MCD and DKD. The observation of correlations between DL measured FPW and key structural parameters of DKD is particularly promising.

The extent of foot process widening provides diagnostic insights in clinical Renal Pathology. Presence of extensive, rather than segmental, foot process effacement in the absence of deposits is characteristic of primary podocytopathies, such as MCD or primary focal segmental glomerulosclerosis (FSGS). ^{11,50 51,52} however, distinction between segmental vs. extensive foot process effacement has been arbitrary and subjective. In this study DL introduced a quantitative approach for this phenomenon and showed that prominently wide foot processes were rare in controls, comprised <10% of FPW values in Fabry disease and DKD and ~50% of FPW values in MCD. In addition, inter-glomerular average FPW variability was not different among the groups studied, supporting of a uniform foot process widening among the glomeruli in biopsies from these conditions.

Arbitrarily taken images used for diagnostic Renal Pathology does not guarantee a representative sampling of foot processes. In this work, we provided a systematic data-driven approach to determine the minimum sample size (i.e. number of images) for obtaining a robust estimate of average FPW. Incremental sampling analysis revealed that ≈70 images were needed to achieve a very narrow convergence threshold in control and DKD biopsies, while this number was greater for MCD or Fabry nephropathy; however in practice, 50–60 images seems to provide sufficiently robust estimates of average FPW.

Compared to the standard encoder-decoder architecture of U-Net with an argmax for pixelwise classification at the end, the ForkNET model used in this study allows class branch isolation to possibly discard branches and associated filters for some classes. The separate binary masks allow for easily separated membrane and slit image processing and faster and more consistent interpretation of the data compared to ET.

While important progress has been made in application of DL in light microscopy analysis of kidney structure, ^53–57 very few studies have been previously published regarding application of DL in kidney EM. Hacking and Bijol applied DL algorithms on EM images for discrimination between disease and normal specimens, which resulted in an average area under the curve for precision and recall of 88.89% and a recall of 66.67% for diabetic nephropathy. ⁵⁸ Cao et al. used a random forest-based machine learning method involving two level (training and testing) integrations for automatic segmentation of the glomerular basement membrane. ¹⁸ Semi-automated approaches have also been proposed for segmentation and measurement of glomerular basement membrane thickness. ^59,60

Super-resolution microscopy techniques, such as three-dimensional structured illumination microscopy (3D-SIM), ⁶¹ stimulated emission depletion imaging (STED), ⁶² and expansion microscopy^63,64 have been used to study podocyte foot processes. ⁶⁵ Combination of immunofluorescence microscopy and ultrastructural studies is a significant advantage of super-resolution microscopy; however, most current super-resolution microscopy approaches typically need either special tissue handling or access to sophisticated equipment and software not used in routine diagnostic Renal Pathology. A recent study showed that expansion-enhanced super-resolution radial fluctuations (ExSRRF) can achieve a resolution of up to 25 nm using LED-based widefield microscopy which is enough for estimation of FPW;⁶⁶ however, this technique requires special tissue handling for expansion microscopy and includes protocols very different from those used in routine diagnostic Renal Pathology. In contrast, any EM lab can usre our deep learning model by uploading images taken according to the instructions provided in our website and downloading FPW results without a need for special tissue handling.

Our study faces several limitations. EM studies are traditionally performed on a limited number of glomeruli. While there is FPW interglomerular variability, this variability was not different among the conditions studied. Importantly, average FPW obtained from three glomeruli among ~2X10⁶ glomeruli in the kidneys correlated with albuminuria and/or proteinuria and other relevant variables. Similarly, other studies showed that stereology measured FPW in three glomeruli in diabetes, ⁶⁷ Alport syndrome, ⁶⁸ and membranoproliferative glomerulonephritis⁶⁹ correlated with albuminuria or proteinuria, supporting functional relevance of these sampling strategies. Compared to conventional assessment of foot process widening in diagnostic Renal Pathology, our DL approach requires storage of at least 50–60 digital EM images per biopsy (~50–60Mb) which is comparable with the size of a one mm² whole slide image (~48Mb) scanned at ×40 magnification with a resolution of ~0.25 μm per pixel and 24-bit color depth. ⁷⁰ The generalizability of the network is limited by the architecture and the number of training samples. Future work seeks to improve generalizability using more advanced network methods, domain generalization, augmentation, and training samples from conditions other those represented in this study, including in animal models.

In summary, we have developed and validated a novel DL model approach for automated and fast measurement of FPW in EM images. The cloud-based availability of this novel tool and it’s applicability to EM specimens taken for diagnostic purposes makes it suitable not only for research but also potentially for clinical grounds.

DATA SHARING STATEMENT

The repository containing the code for analysis, model definition, and datasets are available at the following git repository https://github.com/najafian-lab/fpw-dl-v1. The Docker container that has the required packages and code preinstalled to run on any dataset is available at https://hub.docker.com/r/smerkd/forknetv5. More details about usage can be found in the public repository documentation. Figures and tables using the dice and convergence metrics to evaluate the model can be found at the public repository with the specified link above.

Supplementary Material

1

Figure S1. A screenshot from UW Segmentation Utility.

Table S1. Examples of dice coefficient (DSC) values for various training groups for podocyte - glomerular basement membrane interface (PGBMI) and filtration slit masks used for evaluation of model performance

Table S2. Glomerular structural parameters of male patients with Fabry disease

Table S3. Comparison of correlations between foot process width measurements by deep learning vs. expert technician in male patients with Fabry disease

Table S4. Glomerular structural parameters of female patients with Fabry disease

Table S5. Comparison of correlations between foot process width measurements by deep learning vs. ET in female patients with Fabry disease

Table S6. Glomerular structural parameters of type 2 diabetic patients

Table S7. Comparison of correlations between foot process width measurements by deep learning vs. ET in type 2 diabetic patients