Kidney Week

Abstract: TH-PO1075

Implementation of an Artificial Neural Network for Automated Podometrics in Human Kidney Specimens

Session Information

• Glomerular Diseases: Podocyte Biology - I
  November 07, 2019 | Location: Exhibit Hall, Walter E. Washington Convention Center
  Abstract Time: 10:00 AM - 12:00 PM

Category: Glomerular Diseases

• 1204 Podocyte Biology

Authors

• Klaus, Martin, III. Department of Medicine, University Medical Center Hamburg-Eppendorf, Hamburg, Germany

Martin Klaus,

• Zimmermann, Marina, Center for Molecular Neurobiology, University Clinic Hamburg-Eppendorf, Hamburg, Germany, Hamburg, Germany

Marina Zimmermann,

• Gernhold, Lukas, III. Department of Medicine, University Medical Center Hamburg-Eppendorf, Hamburg, Germany

Lukas Gernhold,

• Kuppe, Christoph, University Hospital of RWTH Aachen, Aachen, Germany

Christoph Kuppe,

• Wong, Milagros N., III. Department of Medicine, University Medical Center Hamburg-Eppendorf, Hamburg, Germany

Milagros N. Wong,

• Wanner, Nicola, III. Department of Medicine, University Medical Center Hamburg-Eppendorf, Hamburg, Germany

Nicola Wanner,

• Person, Fermin, University Hospital Basel, Basel, Switzerland

Fermin Person,

• Wiech, Thorsten, Department of Pathology, University Hospital Hamburg Eppendorf, Hamburg, Germany

Thorsten Wiech,

• Kramann, Rafael, RWTH Aachen University, Lemiers, Netherlands

Rafael Kramann,

• Bonn, Stefan, Institute for Medical Systems Biology, University Medical Center Hamburg-Eppendorf, Hamburg, Germany

Stefan Bonn,

• Huber, Tobias B., III. Department of Medicine, University Medical Center Hamburg-Eppendorf, Hamburg, Germany

Tobias B. Huber,

• Puelles, Victor G., III. Department of Medicine, University Medical Center Hamburg-Eppendorf, Hamburg, Germany

Victor G. Puelles,

Background

Advanced morphometrics to study podocytes (podometrics) may provide robust readouts for diagnosis, prognosis and management of patients with glomerular diseases. However, their clinical implementation is limited by their time-consuming nature. Thus, artificial neural networks emerge as interesting tools to bring podometrics closer to the bedside.

Methods

Ground truth data was determined in 318 images (144 training, and 174 testing), acquired using immunofluorescence and confocal microscopy. An artificial neural network (U-Net) was implemented, optimised via a systematic grid search and compared to an automatic ImageJ-based segmentation tool. Dice scores (pixel-based), F1 scores (object-based), and spearman correlations were calculated to validate each method against the ground truth. Model-based stereology podometrics were determined using segmented data.

Results

In nephrectomy samples, U-Net provided higher Dice and F1 scores than those obtained with ImageJ (P<0.0001), with stronger correlation indices for U-Net (R=0.94-0.95, P<0.0001) compared to ImageJ (R=0.61-0.66, P<0.01). In ANCA-associated glomerulonephritis, Dice and F1 scores were also higher in U-Net (P<0.0001) compared to ImageJ with stronger correlation indices in U-Net (R=0.91-0.94, P<0.0001) compared to ImageJ (R=0.59-0.79, P<0.01).

Conclusion

Our optimised artificial neural network (U-Net) provides readouts that are comparable to manual segmentation and superior to conventional segmentation tools, even in the context of glomerular disease. These findings bring us one step closer to the use of automatic podometrics as a clinical instrument.

We have implemented an artificial neural network to automatically extract the data required for the estimation of podometrics in clinical samples.