ECP 2023 Abstracts

S17 Virchows Archiv (2023) 483 (Suppl 1):S1–S391 13 Conclusion: High-grade tumour budding, low immunoscore, non-brisk or absent lymphocytic infiltrate, and mutated p53 immunoprofile are negative prognostic markers in penile squamous cell carcinoma. Grade, histological subtype, and HPV status (as determined by p16 immunohis- tochemistry) showed, surprisingly, little or no prognostic significance. This work was supported by the Czech Health Research Council (Grant number NU21J-03-00019). OFP-04-010 Detection of muscularis propria invasion in urothelial carcinoma using artificial intelligence I. Fahoum*, R. Naamneh, R. Hagege, K. Pragier, D. Hershkovitz *Institute of Pathology, Tel-Aviv Sourasky Medical Center, Israel Background & objectives: Muscularis propria invasion in urothelial carcinoma indicates surgical treatment and chemotherapy. Detecting muscularis propria invasion in transurethral resection specimens can be time consuming for pathologists. We aim to develop an algorithm to identify muscularis propria invasion in urothelial carcinoma. Methods: Training the algorithm involved manual segmentation of muscularis propria and tumour using 925 images from 50 cases of urothelial carcinoma. Analytical validation was performed using 97 additional manually segmented images from 10 new cases, and the algorithm was compared to annotations of pathologist for tumour and muscularis propria. Detection rate, false alarm rate and intersection over union were calculated. Results: In the validation cohort the detection rates of the algorithm for muscularis propria and tumour were 85% and 90%, respectively. On the other hand, the false alarm rates were 21% for tumour and 30% for muscularis propria. Causes of false alarms included thick muscular wall of blood vessels in the lamina propria and urothelium with reactive changes. When we defined event as muscle and tumour tissue were in nearest proximity, the algorithm showed lower false alarm rates. Conclusion: Artificial intelligence can be an efficient tool and assist patholo- gists identifying urothelial carcinoma invasion into the muscularis propria. Our algorithmic approach yielded good results based on a relatively small training cohort. Inclusion of the proximity factor between the tumour and the muscularis propria significantly reduced the false alarm rate. OFP-04-011 Assessing the robustness of a glomerular segmentation tool: char- acterization and automatic detection of annotation errors H. Weishaupt*, N. Mola, J. Besusparis, S. Sund, S. Leh *Department of Pathology, Haukeland University Hospital, Bergen, Norway Background & objectives: In nephropathology, segmentation of glomeruli is one of the tasks most extensively studied using artificial intelligence. While some software applications are already becoming publicly available, it remains largely unclear how robust such tools are when applied to new cases. Methods: To address this open question, the current study investigated the performance of a publicly available glomerular segmentation tool when applied to over 1000 whole slide images, yielding more than 125000 glomerular predictions. A combination of qualitative and quan- titative measures was employed to determine the types and prevalence of annotation errors and how to automatically detect them. Results: Through visual inspection, the project revealed that, in addition to false-positive and false-negative predictions, there were three major categories of annotation errors reflecting different segmentation mistakes. Furthermore, utilizing a wide range of morphometric features, i.e., applied quantitative metrics characterizing the shape of annotations, these errors could be detected automatically and were estimated to account for about 5% of all predictions. Conclusion: Having investigated a large number of glomerular annotations predicted by a publicly available segmentation tool, the project demonstrates what kind of mistakes can occur, how often they occur, and how they can be automatically detected. We believe that the results of the study are highly relevant for future use of segmentation tools by raising awareness of different segmentation mistakes and how they can be identified, thus enabling their subse- quent correction. OFP-04-012 Digital pathology and gene expression analysis as potential tools of standardisation for renal transplant pathology D. Dobi*, A. Edwards, Z.G. Laszik *Semmelweis University, Hungary Background & objectives: Renal transplant diagnostics is based on semi-quantitative assessment of the biopsy sample. This method is prone to inter-, and intraobserver variability. Here, we used the tools of digital and molecular pathology to explore ways to improve the current diagnostic system. Methods: The enrolled kidney biopsies covered the entire spectrum of the T-cell alloimmune response, from normal morphology to bor- derline changes to T-cell-mediated rejection (n = 96). Histological sections from formalin-fixed, paraffin-embedded (FFPE) tissue blocks were stained with multiplexed immunohistochemistry, then were digi- tized and analysed using machine learning. Gene expression stud- ies were performed on sections from FFPE tissue on the Nanostring platform. Results: The F-score, recall and precision for the glomeruli were 0.91, 0.89 and 0.94 respectively. Similarly excellent parameters could be achieved for the tubules (0.93, 0.95 and 0.91). Object recognition capacity for inflammatory cells (0.79, 0.84 and 0.74) and peritubular capillaries (0.66, 0.82 and 0.60) were fair. The vari- ables generated with the help of the algorithm showed a moderate to strong correlation with the gene sets characteristic of T cell-mediated rejection. Conclusion: Digital pathology methods and gene expression testing are useful additional tools for semiquantitative kidney biopsy evaluation. OFP-04-013 Hepatitis E virus infection mimics T-cell mediated rejection in transplanted kidneys J. Schmitz*, J. Kracht, K. Evert, J.J. Wenzel, M. Schemmerer, M. Panning, L. Pape, M. Pohl, J.H. Bräsen *Hannover Medical School, Institute of Pathology, Nephropathology, Germany Background & objectives: Immunocompromised patients are at risk of chronic hepatitis E (HEV) infection, which can be treated with riba- virin. Recurrent T-cell and borderline rejections in a paediatric patient with high HEV copy numbers led us to study HEV infection within renal transplants. Methods: To investigate frequency of HEV infection in renal allografts, we identified a total of 15 samples from patients with contemporaneous diagnosis of HEV-infection (diagnosed by PCR of blood/stool) at Hannover Medical School. Ten samples had suf- ficient residual paraffin tissue for immunofluorescence (IF) and fluorescence-in-situ-hybridization (RNA-FISH). The biopsy of the paediatric index patient was additionally sufficient for tissue-PCR and electron microscopy. Results: We detected HEV virus RNA in paraffin tissue of the pae- diatric index patient by tissue-PCR. Subsequently, we localized HEV infection in tubular epithelial cells by immunofluorescence and RNA- FISH (HEV Oligo-FISH) within residual paraffin tissue as well as in electron microscopy using a routinely embedded epon block. One of