Introduction
The Banff classification system is used for diagnosis and classification of kidney transplant biopsies. An important feature of antibody-mediated rejection (ABMR) is peritubular capillaritis (ptc), defined as the presence of inflammation in peritubular capillaries (PTCs). However, assessing the extent of peritubular capillaritis suffers from interobserver variability and is timeconsuming. Automated assessment would offer great benefits.
Material and methods
Kidney transplant biopsies (n=67) were stained with periodic-acid Schiff (PAS) and scanned with a P1000 (3DHISTECH, Hungary) whole slide image (WSI) scanner (0.24 um/pixel). To obtain reliable annotations, the PAS-stained WSI were re-stained using anti-CD34-antibody. Guided by the restaining, a pathologist manually annotated over 20,000 PTCs on the PAS-stained WSI. For PTCs versus non-PTCs segmentation, we trained a U-Net model (ImageNet pre-trained ResNet50 backbone) with 160,000 patches (512 x 512 pixels, 0.24 um/pixel) per epoch.
Results and discussion
A Dice score of 67.9% and 98.3%, and a Jaccard Index of 52.3% and 96.6%, for the PTCs versus non-PTCs regions, respectively was achieved. While there was a satisfactory performance in most cases, we observed less accuracy in cases with prominent interstitial changes, such as atrophic tubules and interstitial matrix deposition, which make PTCs less recognizable.
Conclusion
We developed a segmentation model for PTCs in PAS-stained kidney transplant biopsies, which in contrast to healthy tissue also include areas of inflammation and chronic damage. This is a first step towards a more accurate, reproducible scoring of peritubular capillaritis. Our next goal is to develop an algorithm for inflammatory cell detection, as this is necessary for automated ptc scoring.