Deep learning–based segmentation and quantification of podocyte foot process morphology suggests differential patterns of foot process effacement across kidney pathologies

Abstract

Morphological alterations at the kidney filtration barrier increase intrinsic capillary wall permeability resulting in albuminuria. However, automated, quantitative assessment of these morphological changes has not been possible with electron or light microscopy. Here we present a deep learning-based approach for segmentation and quantitative analysis of foot processes in images acquired with confocal and super-resolution fluorescence microscopy. Our method, Automatic Morphological Analysis of Podocytes (AMAP), accurately segments podocyte foot processes and quantifies their morphology. AMAP applied to a set of kidney diseases in patient biopsies and a mouse model of focal segmental glomerulosclerosis allowed for accurate and comprehensive quantification of various morphometric features. With the use of AMAP, detailed morphology of podocyte foot process effacement was found to differ between categories of kidney pathologies, showed detailed variability between diverse patients with the same clinical diagnosis, and correlated with levels of proteinuria. AMAP could potentially complement other readouts such as various omics, standard histologic/electron microscopy and blood/urine assays for future personalized diagnosis and treatment of kidney disease. Thus, our novel finding could have implications to afford an understanding of early phases of kidney disease progression and may provide supplemental information in precision diagnostics. Morphological alterations at the kidney filtration barrier increase intrinsic capillary wall permeability resulting in albuminuria. However, automated, quantitative assessment of these morphological changes has not been possible with electron or light microscopy. Here we present a deep learning-based approach for segmentation and quantitative analysis of foot processes in images acquired with confocal and super-resolution fluorescence microscopy. Our method, Automatic Morphological Analysis of Podocytes (AMAP), accurately segments podocyte foot processes and quantifies their morphology. AMAP applied to a set of kidney diseases in patient biopsies and a mouse model of focal segmental glomerulosclerosis allowed for accurate and comprehensive quantification of various morphometric features. With the use of AMAP, detailed morphology of podocyte foot process effacement was found to differ between categories of kidney pathologies, showed detailed variability between diverse patients with the same clinical diagnosis, and correlated with levels of proteinuria. AMAP could potentially complement other readouts such as various omics, standard histologic/electron microscopy and blood/urine assays for future personalized diagnosis and treatment of kidney disease. Thus, our novel finding could have implications to afford an understanding of early phases of kidney disease progression and may provide supplemental information in precision diagnostics. Translational StatementWe here describe that using deep learning–based segmentation and new imaging protocols, foot process morphometry can be automatically extracted from images of human biopsies. Our results indicate that individual patterns of foot process effacement differ between patients and correlate with levels of proteinuria. Although there are technical challenges in its clinical implementation, automatic morphologic analysis of podocytes could potentially complement other readouts, such as omics approaches, standard histopathology, electron microscopy, and blood and urine assays, for future personalized diagnosis and treatment of kidney disease. We here describe that using deep learning–based segmentation and new imaging protocols, foot process morphometry can be automatically extracted from images of human biopsies. Our results indicate that individual patterns of foot process effacement differ between patients and correlate with levels of proteinuria. Although there are technical challenges in its clinical implementation, automatic morphologic analysis of podocytes could potentially complement other readouts, such as omics approaches, standard histopathology, electron microscopy, and blood and urine assays, for future personalized diagnosis and treatment of kidney disease. The glomerular filtration barrier consists of a fenestrated endothelium, the glomerular basement membrane, and postmitotic epithelial cells, called podocytes.1Benzing T. Salant D. Insights into glomerular filtration and albuminuria.N Engl J Med. 2021; 384: 1437-1446Crossref PubMed Scopus (67) Google Scholar Damage to any of these layers results in increased filter permeability and consequent albuminuria.1Benzing T. Salant D. Insights into glomerular filtration and albuminuria.N Engl J Med. 2021; 384: 1437-1446Crossref PubMed Scopus (67) Google Scholar Pathologic alterations to podocyte foot processes (FPs) and slit diaphragm (SD) morphology (i.e., effacement) are seen in most glomerular diseases with shortening, widening, and ultimately loss of FPs together with a progressive reduction of SD abundance.2Suleiman H.Y. Roth R Jain S et al.Injury-induced actin cytoskeleton reorganization in podocytes revealed by super-resolution microscopy.JCI Insight. 2017; 2e94137Crossref PubMed Google Scholar,3Unnersjö-Jess D. Scott L. Blom H. Brismar H. Super-resolution stimulated emission depletion imaging of slit diaphragm proteins in optically cleared kidney tissue.Kidney Int. 2016; 89: 243-247Abstract Full Text Full Text PDF PubMed Scopus (62) Google Scholar Recently, super-resolution microscopy techniques have allowed for visualizing and quantifying such morphologic alterations,3Unnersjö-Jess D. Scott L. Blom H. Brismar H. Super-resolution stimulated emission depletion imaging of slit diaphragm proteins in optically cleared kidney tissue.Kidney Int. 2016; 89: 243-247Abstract Full Text Full Text PDF PubMed Scopus (62) Google Scholar, 4Unnersjö-Jess D. Butt L. Höhne M. et al.A fast and simple clearing and swelling protocol for 3D in-situ imaging of the kidney across scales.Kidney Int. 2021; 99: 1010-1020Abstract Full Text Full Text PDF PubMed Scopus (11) Google Scholar, 5Pullman J.M. New views of the glomerulus: advanced microscopy for advanced diagnosis.Front Med. 2019; 6: 37Crossref Scopus (9) Google Scholar, 6Siegerist F. Ribback S. Dombrowski F. et al.Structured illumination microscopy and automatized image processing as a rapid diagnostic tool for podocyte effacement.Sci Rep. 2017; 7: 11473Crossref PubMed Scopus (44) Google Scholar and our group recently used this approach to provide a new model in which size selectivity of the filtration barrier is dependent on glomerular basement membrane compression.7Butt L. Unnersjö-Jess D. Höhne M. et al.A molecular mechanism explaining albuminuria in kidney disease.Nat Metab. 2020; 2: 461-474Crossref PubMed Scopus (66) Google Scholar In that study, FP morphology correlated robustly with levels of albuminuria. Until now, morphologic analyses required time-consuming manual input, which is prone to investigator-dependent biases, impeding broad use in research and diagnostics. New machine-learning methods to automate bioimage analysis at a human-level accuracy are developing fast, with cancer histopathology as a prominent application. In nephrology, deep learning methods have been proposed for segmenting entire renal structures or cells.8Bouteldja N. Klinkhammer B. Bulow R. et al.Deep learning–based segmentation and quantification in experimental kidney histopathology.J Am Soc Nephrol. 2021; 32: 52-68Crossref PubMed Scopus (67) Google Scholar, 9Jayapandian C.P. Chen Y. Janowczyk A. et al.Development and evaluation of deep learning–based segmentation of histologic structures in the kidney cortex with multiple histologic stains.Kidney Int. 2021; 99: 86-101Abstract Full Text Full Text PDF PubMed Scopus (74) Google Scholar, 10Zimmermann M. Klaus M. Wong M.N. et al.Deep learning-based molecular morphometrics for kidney biopsies.JCI Insight. 2021; 6e144779Crossref PubMed Scopus (18) Google Scholar However, none of these methods allows automatic recognition of podocyte substructures and FP effacement analysis. We here present automatic morphologic analysis of podocytes (AMAP)—an automated method for FP and SD segmentation from light microscopy images of podocytes. It is trained and tested on a broad range of disease and imaging conditions, correctly detecting FPs across our data sets of mice and reproducing morphometric parameters originally extracted using an ImageJ macro—referred to as macro throughout the article. AMAP generalizes also to human samples and to confocal microscopy images, an imaging system commonly available in most clinical pathology laboratories, opening up possibilities for fast and robust studies of the complex nanoscale pathologies in kidney disease. See online Supplementary Methods for details of patients, animals, sample preparation protocols, and macro analysis. We modified U-Net segmentation11Ronneberger O, Fischer P, Brox T. U-Net: convolutional networks for biomedical image segmentation. In: Navab N, Hornegger J, Wells W, Frangi A, eds. Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015. Lecture Notes in Computer Science. 2015;9351. Springer, Cham. https://doi.org/10.1007/978-3-319-24574-4_28Google Scholar by adding 2 convolutional layers after the last layer. The first convolution produced 3 output channels and was followed by cross-entropy loss. The second convolution produced 16 output channels and was followed by discriminative loss, as described previously.12De Brabandere B. Neven D. Van Gool L. Semantic instance segmentation with a discriminative loss function. arXiv.https://doi.org/10.48550/arXiv.1708.02551Google Scholar This loss forced 16-dimensional pixel representations to group together, in terms of euclidian distance, if pixels belonged to the same instance and to separate from one another if pixels belong to distinct instances. We chose 16 as the smallest dimensionality that resulted in correct clusters. As images were too large to fit in the hardware memory, we used randomly sampled 384-×-384 pixel crops of the original images during training. Images were additionally randomly flipped along x- and y-axis and rotated by 0°, 90°, 180°, or 270° during training. We used RMSProp optimizer and starting learning rate of 0.05 in training. We additionally used a scheduler that decreased learning rate by a factor of 0.8 every 100 epochs if no decrease in value of the loss function was observed. We used a batch size of 8 and performed 3 training iterations of 1000 epochs each. In each training session, the starting learning rate was set at 0.05. This approach resulted in convergence in a better local optimum with each iteration. We performed an additional training iteration on the train set expanded by 55 confocal microscopy images, allowing adaptation of the segmentation method to images generated with the fast protocol. During inference, we used windows of size 384-×-384 pixel crops from the original images with an overlap of 256 pixels between neighboring windows. The overlap allowed us to better resolve segmentation of FPs that fall on the boundaries of a given window. When merging windows, we took the maximum class label in the overlapping regions. In instance segmentation, overlapping instances were merged together. To separate individual FP instances, pixel representations were clustered. As the correct number of clusters was unknown, we performed several clusterings for each image window and chose the one that was the best according to a set of criteria. We used BICO,13Fichtenberger H. Gillé M. Schmidt M. et al.BICOBIRCH meets coresets for k-means clustering.in: Bodlaender HL Italiano GF – Lecture Notes in Computer Springer, Scopus Google Scholar a that data allowing for into multiple of clusters. We pixel representations with a the of a pixel into a given of FP separate were individual this additional information the the clusterings we performed on segmentation we one according to such as the number of distinct clusters that were number of clusters that were of a size and standard of of in 3 was The clusterings we tested for each image window were on the of the of FP in this window. We performed of the FP to the number of FPs in the train of clusters between and of the number of FPs on this were tested as described The resulting instance segmentation was by FPs with or which we as resulting from segmentation We also FPs on the image boundaries that were not in an image and that in of the morphometric We used batch k-means with a batch size 1000 as of the large number of clusterings we performed during the we this process across processing To the between and we performed an instance In this for each we for a FP instance that showed overlap with the given of overlap between the FPs was as instances that not overlap with any FP were as and FPs with no overlap with instances were as training and were performed on processing with processing and 2 We our together with and sample data available a super-resolution stimulated emission depletion imaging D. Scott L. Blom H. Brismar H. Super-resolution stimulated emission depletion imaging of slit diaphragm proteins in optically cleared kidney tissue.Kidney Int. 2016; 89: 243-247Abstract Full Text Full Text PDF PubMed Scopus (62) Google Scholar we generated images mouse at various and of podocyte the SD is by we SD and FP We trained a segmentation convolutional on the training set of of these images Brabandere B. Neven D. Van Gool L. Semantic instance segmentation with a discriminative loss function. arXiv.https://doi.org/10.48550/arXiv.1708.02551Google Scholar the of the convolutional performed segmentation, image pixels into 3 and SD segmentation FP pixels into separate FP instances We trained the and its segmentation on the pixel AMAP correct to and of FP and SD The of the and of FP and SD was to manual of training images, and in the boundaries of structures and of and not in the same in FP To this we additionally performed a of in results 3 We FP instances with the instances overlapping with FPs were as and the of overlap with the label was In images, we of FPs FP instances at a value of as of and instances. to FPs were with AMAP, were not were to in the training set and not be To the of our we tested the segmentation approach on an data set and in a clinical and We a segmentation accuracy in images, which were of a FP resulting from AMAP segmentation and challenges in As a of we if AMAP morphometric parameters extracted using a analysis of a focal segmental glomerular mouse model and L. Unnersjö-Jess D. Höhne M. et al.A molecular mechanism explaining albuminuria in kidney disease.Nat Metab. 2020; 2: 461-474Crossref PubMed Scopus (66) Google Scholar The parameters were SD SD FP FP and FP is a of a and FPs We AMAP on images from the original and mice of FPs in the data we correctly with an overlap of on We additional FPs not in the data of these that correct FPs that were not using the macro for an of This the of in the FP and automatic segmentation is that 2 AMAP with the macro analysis in the mouse This not be as a segmentation analysis as the one described in In the macro was used to an that was to a the FPs by AMAP be we if AMAP the morphometry of disease progression in the mouse We found of morphometric parameters extracted by AMAP of and were the and that additional FPs by AMAP to be and with and extracted by AMAP allowed from mice and FP not for this FP morphometry extracted by AMAP with the macro Supplementary In morphometric parameters of FPs are robust to the variability in the number of FPs between the macro and AMAP With our of (i.e., the SD was to be We expanded AMAP to automatically on the SD allowing quantification of the SD these AMAP extracted for SD with the macro This is to the on by AMAP and Supplementary SD by AMAP was on by AMAP of the capillary the not in increased SD resulting in for SD these in and SD the correlated robustly between the macro and AMAP the decrease of SD in with mice was using AMAP also for the decrease in SD with SD and levels of albuminuria correlated as was using the L. Unnersjö-Jess D. Höhne M. et al.A molecular mechanism explaining albuminuria in kidney disease.Nat Metab. 2020; 2: 461-474Crossref PubMed Scopus (66) Google Scholar We size and found that for parameters on the SD segmentation (i.e., SD and In AMAP the morphometry of FP pathology in an model with decrease in and of for a trained macro to segmentation of this data We have recently a protocol for fast and confocal imaging of FPs to here as fast which the for sample preparation and nanoscale imaging to D. Butt L. Höhne M. et al.A fast and simple clearing and swelling protocol for 3D in-situ imaging of the kidney across scales.Kidney Int. 2021; 99: 1010-1020Abstract Full Text Full Text PDF PubMed Scopus (11) Google Scholar To AMAP to the of this we generated an additional set of and human and mouse images, and we these images to the training set and the This segmentation also in images of and We also the accuracy of AMAP segmentation as that the 3 or as in which at 2 of the 3 to the challenges in an accurate for a new protocol for D. A. Höhne M. et imaging of kidney 2021; PubMed Scopus Google Scholar we samples from patients with disease We AMAP on confocal images from these patients along with and 2 patients, with robust segmentation of the SD and individual FPs and Supplementary of images and AMAP segmentation from this are in Supplementary In the mouse in SD are by in FP this was in the human patients with To effacement we multiple morphometric parameters SD FP FP and FP using and dimensionality L. and for Scholar to of patients and of the of to is used See Supplementary for an of morphometric parameters in these We that SD and FP and that the is of This that quantitative parameters patterns of a decrease in SD is not with an increase in FP AMAP morphometry not the of FP also the of individual FPs on patients of patient with patient in the value of and of AMAP kidney nanoscale In and variability is which can be with a of on a with a large of We applied AMAP to patients with patients, for the patient with are of a data D. Butt L. Höhne M. et al.A fast and simple clearing and swelling protocol for 3D in-situ imaging of the kidney across scales.Kidney Int. 2021; 99: 1010-1020Abstract Full Text Full Text PDF PubMed Scopus (11) Google Scholar and no between SD and FP was and However, patients with showed distinct in the that disease of the number of patient samples in the we FP morphometry in of diverse of glomerular disease in We mice with L. Unnersjö-Jess D. Höhne M. et imaging of the filtration barrier a for in to glomerular Am Soc Nephrol. PubMed Scopus Google S. B. S. et of in experimental Am Soc Nephrol. PubMed Scopus Google Scholar We that FPs were on in the mice and This was also SD was in the mice and the FP was as Thus, on the SD FPs in the mice with the for FP In standard of parameters in the mice a of FP morphology with the results patterns of FP effacement on the not with other imaging and analysis We have that FP morphometry with albuminuria in our mouse L. Unnersjö-Jess D. Höhne M. et al.A molecular mechanism explaining albuminuria in kidney disease.Nat Metab. 2020; 2: 461-474Crossref PubMed Scopus (66) Google Scholar We morphometry with albuminuria in patients with We a of albuminuria levels the from to are in the In SD decreased as levels increased resulting in a between SD and was no between and most patients with of with if SD at the that an increased (i.e., be a first of podocyte loss of SD and proteinuria. our data indicate that SD treatment FP that of the patients with in this are on treatment or patients not the FP effacement and that is a of we between morphometry and treatment in patients with We found an morphometric of patients and patients standard The was prominent in patients on patients a with The other patients treatment showed a Although the sample size is this the of AMAP as a tool to treatment on FP morphology. In light microscopy imaging of podocyte has large these techniques complement electron microscopy also the of imaging quantitative image D. Scott L. Blom H. Brismar H. Super-resolution stimulated emission depletion imaging of slit diaphragm proteins in optically cleared kidney tissue.Kidney Int. 2016; 89: 243-247Abstract Full Text Full Text PDF PubMed Scopus (62) Google F. Ribback S. Dombrowski F. et al.Structured illumination microscopy and automatized image processing as a rapid diagnostic tool for podocyte effacement.Sci Rep. 2017; 7: 11473Crossref PubMed Scopus (44) Google L. Unnersjö-Jess D. Höhne M. et al.A molecular mechanism explaining albuminuria in kidney disease.Nat Metab. 2020; 2: 461-474Crossref PubMed Scopus (66) Google Scholar Although these quantitative analyses on manual input, which is time-consuming and prone to deep we here a automatic segmentation of FPs and the which results in a mouse model for L. Unnersjö-Jess D. Höhne M. et al.A molecular mechanism explaining albuminuria in kidney disease.Nat Metab. 2020; 2: 461-474Crossref PubMed Scopus (66) Google Scholar Thus, AMAP can be to FP morphology with the same accuracy and with our AMAP the of SD and levels of the proposed mechanism of albuminuria in podocyte L. Unnersjö-Jess D. Höhne M. et al.A molecular mechanism explaining albuminuria in kidney disease.Nat Metab. 2020; 2: 461-474Crossref PubMed Scopus (66) Google Scholar We that AMAP can be to the SD and FPs in human patient clinical from images acquired using our recently fast and simple D. Butt L. Höhne M. et al.A fast and simple clearing and swelling protocol for 3D in-situ imaging of the kidney across scales.Kidney Int. 2021; 99: 1010-1020Abstract Full Text Full Text PDF PubMed Scopus (11) Google D. A. Höhne M. et imaging of kidney 2021; PubMed Scopus Google Scholar morphometric analysis in human samples revealed distinct morphologic alterations in patients with on with such as and also distinct We additionally morphologic patterns between a model and an morphologic SD and are complex that the of FP effacement as a process have to be and that effacement patterns could differ on not studies in also to clinical diagnostics. we that FP morphometry is to treatment in patients with on data size (i.e., number and size of AMAP can be on a or any advanced with a processing of this the of of the which results in This is for images of or in images of The is in and to the accuracy of our automated method that potentially segmentation across in the training AMAP be a tool for changes to and not a tool for accurate segmentation of FPs in an for images and images of the SD of which the SD is automatically of the capillary resulting in for SD with the macro analysis. the human patient be that the size is and the results be as the technical to image and human FP pathology from biopsies. When to the possible clinical of AMAP, the sample preparation and imaging which of the methods to the It on the other these new techniques that the accurate and comprehensive analysis in this AMAP is at present to FP with and of we that of AMAP to for the glomerular basement membrane and the in the future be We have in this automated and be to also to automate microscopy image Although we have followed to automated imaging any possible and increase the of the analysis. In AMAP allows for automated quantification of nanoscale FP morphology. We that AMAP is to human samples with recently fast protocols, no super-resolution microscopy. The of AMAP and a fast and simple protocol in the future for and quantitative kidney using one sample and analysis N. Klinkhammer B. Bulow R. et al.Deep learning–based segmentation and quantification in experimental kidney histopathology.J Am Soc Nephrol. 2021; 32: 52-68Crossref PubMed Scopus (67) Google Scholar, 9Jayapandian C.P. Chen Y. Janowczyk A. et al.Development and evaluation of deep learning–based segmentation of histologic structures in the kidney cortex with multiple histologic stains.Kidney Int. 2021; 99: 86-101Abstract Full Text Full Text PDF PubMed Scopus (74) Google Scholar, 10Zimmermann M. Klaus M. Wong M.N. et al.Deep learning-based molecular morphometrics for kidney biopsies.JCI Insight. 2021; 6e144779Crossref PubMed Scopus (18) Google Scholar the no We the on in and the for their in the of microscopy This was by the and as as by of the and were by the and the in and by the for This was by the and the number the by the for and to and and number with Supplementary Supplementary Supplementary of the data sets used in this Supplementary Supplementary of variability in image Supplementary to data in evaluation of method on a data set and in a Supplementary of segmentation accuracy with on data sets of Supplementary foot process and slit diaphragm (SD) Supplementary of morphometric parameters as by macro and automatic morphologic analysis of podocytes Supplementary of morphometric parameters with number of foot processes Supplementary in slit diaphragm (SD) between macro and automatic morphologic analysis of podocytes Supplementary of morphometric parameters with of Supplementary of training data for the fast protocol images Supplementary of segmentation results on 2 or Supplementary segmentation of slit diaphragm (SD) and foot process in images acquired in biopsies from patients with disease Supplementary of morphometric parameters in and Supplementary of the and on the method Supplementary slit diaphragm (SD) a and individual glomerular Supplementary Automatic morphologic analysis of podocytes foot process morphometry in human Supplementary of morphometric parameters in and are generated on a with Supplementary which the same on a in this et that the of of that kidney epithelial could be used to for kidney that intrinsic is the of a is a factor that kidney after an factor the as an PDF

Topics & Concepts