Automated Segmentation and Morphological Analysis of Peritubular Capillaries for Improved Diagnostics in Kidney Transplantation

Abstract

Defining the level of inflammation in the peritubular capillaries (PTC), called peritubular capil laritis, is a crucial part of the rejection diagnosis through the Banff classification, a standardized system for grading kidney transplant rejection. However, assessing the degree and presence of inflammation can be difficult and prone to inter-observer variability between pathologists. In this thesis, an automated binary PTC segmentation model (nnU-Net V2) is presented, which can serve as a more objective baseline for assessing peritubular capillaritis. This study successfully improved segmentation accuracy (mean Precision and Dice of 0.775 and 0.751) compared to earlier approaches (mean Precision and Dice of 0.677 and 0.713), reducing errors significantly (p = 0.00001 and p = 0.00014). Furthermore, automatic per-PTC assessment of peritubular capillaritis is explored using a Patch Classification Model (PCM) and a Multi-Class Segmentation Model (MCSM). Both approaches achieved higher F1-scores for the more common classes (PCM: ptc0: 0.90, ptc1: 0.75) but showed lower performance on rare classes (PCM: ptc2: 0.29, ptc3: 0.44), indicating areas for future improvement. The analysis of morphometrics of PTCs, such as size, circularity, and density, demonstrated several significant correlations between morphometrics and Banff scores, as well as significant differences between rejection types, suggesting that PTC morphometrics might be helpful for diagnostic decision-making.