Artificial Intelligence-Driven Computational Analysis of Kidney Transplant Tissue

Download thesis

M. Hermsen

  • Promotor: J. van der Laak, L. Hilbrands and B. Smeets
  • Graduation year: 2024
  • Radboud University

Abstract

This thesis is organized as follows:

In Chapter 2 we present a CNN for the multi-class structure segmentation of WSI from PAS-stained kidney biopsies. Next to internal validation, the segmentation performance was assessed on kidney biopsies from an external medical center to demonstrate robustness for staining variations. Additionally, we show the applicability of the CNN on nephrectomy samples. Lastly, CNN-based measures were compared with visually scored histological (Banff) components in kidney transplant biopsies.

In Chapter 3 we describe a methodology to assess the microenvironment in sparse kidney transplant biopsy samples. Deep learning, multiplex immunohistochemistry and mathematical image processing techniques were incorporated to quantify lymphocytes, macrophages, and capillaries in kidney transplant biopsies of patients with delayed graft function. The quantitative results were used to assess correlations with subsequent development of interstitial fibrosis and tubular atrophy.

In Chapter 4 we assess the potential of CNNs to quantify chronic and inflammatory lesions in PAS and CD3 stained kidney transplant biopsies. A multi-class structure segmentation CNN and a lymphocyte detection CNN were applied, and non-linear image registration allowed for the quantification of inflammatory burden in specific structures. These automatically quantified features were compared to chronic and inflammatory Banff lesions scored by a panel of five pathologists.

In Chapter 5 we present a new CNN for the segmentation and classification of glomeruli. This network was applied on a large kidney transplant patient cohort to quantify the percentage globally sclerotic glomeruli at implantation and 5 years after transplantation. The CNN-based results were compared to quantifications based on morphometric analysis and we assessed whether the same correlations with clinical parameters are found when the percentage GSG is determined by either deep learning or manual morphometry.

In Chapter 6 we discuss the main findings and contributions of this thesis and provide a future outlook for further research possibilities.