Deep learning-based histopathologic segmentation of peritubular capillaries in kidney transplant biopsies

Computers in Biology and Medicine, vol. 193, pp. 110395

Abstract

BACKGROUND: Assessing the extent of inflammation in peritubular capillaries (PTCs) is important for diagnosing antibody-mediated rejection in kidney transplant biopsies. However, this assessment is time-consuming and suffers from interobserver variability, making it a promising candidate for automation. This study introduces a deep learning-based approach for PTCs detection and segmentation in PAS-stained kidney transplant biopsies, providing a first step towards automated scoring of peritubular capillaritis (ptc). METHODS: PAS-stained whole slide images of kidney transplant biopsies (n = 69) were manually annotated, guided by CD34 immunohistochemical re-staining, resulting in approximately 28.000 PTCs for training and evaluation. RESULTS: We trained a U-Net model using a patch-sampling approach in a two-class setting (PTCs versus other tissue), achieving a PTCs segmentation performance of 71.3 %, 56.1 % and 73.0 % for the Dice score, Jaccard index and normalized surface Dice, respectively. For the individual PTCs detection, we achieved a precision of 67.7 %, and a recall of 71.8 %. CONCLUSIONS: Despite satisfactory performance of the DL-model developed in this study, we observed that prominent interstitial inflammation and fibrosis, where PTCs become less recognizable, and the presence of structures mimicking PTCs (e.g. atrophic tubules), posed the largest challenges for the model. Our model is novel as it was trained on a set of transplant biopsies with morphological abnormalities, requiring only PAS-stained kidney biopsies and confirmation of ground truth with the use of CD34 immunohistochemistry. Next steps include incorporation of inflammatory cell segmentation, which may enhance accuracy, reproducibility and quantitative scoring of peritubular capillaritis to guide pathologists in routine diagnostics.

Authors 9

1. Dominique van Midden corresponding

   Radboud University Nijmegen · Radboud University Medical Center

   Affiliation as printed

   Department of Pathology, Radboud University Medical Center, Nijmegen, the Netherlands. Electronic address: Dominique.vanMidden@radboudumc.nl

2. Linda Studer

   Radboud University Medical Center

   Affiliation as printed

   Department of Pathology, Radboud University Medical Center, Nijmegen, the Netherlands

3. Meyke Hermsen

   Radboud University Medical Center

   Affiliation as printed

   Department of Pathology, Radboud University Medical Center, Nijmegen, the Netherlands

4. Eric J. Steenbergen

   Radboud University Medical Center

   Affiliation as printed

   Department of Pathology, Radboud University Medical Center, Nijmegen, the Netherlands

5. Jesper Kers Aachen

   Leiden University Medical Center · Amsterdam University Medical Centers · University of Amsterdam

   Affiliation as printed

   Department of Pathology, Amsterdam University Medical Center, University of Amsterdam, Amsterdam, the Netherlands; Department of Pathology, Leiden University Medical Center, Leiden, the Netherlands

6. Nicolas Kozakowski

   Medical University of Vienna

   Affiliation as printed

   Department of Pathology, Medical University of Vienna, Vienna, Austria

7. Željko Kikić

   Medical University of Vienna

   Affiliation as printed

   Department of Urology, Medical University of Vienna, Vienna, Austria

8. Luuk B. Hilbrands

   Radboud University Medical Center

   Affiliation as printed

   Department of Nephrology, Radboud University Medical Center, Nijmegen, the Netherlands

9. Jeroen van der Laak

   Linköping University · Radboud University Nijmegen · Radboud University Medical Center

   Affiliation as printed

   Department of Pathology, Radboud University Medical Center, Nijmegen, the Netherlands; Center for Medical Image Science and Visualization, Linköping University, Linköping, Sweden