ECP 2023 Abstracts

S53 Virchows Archiv (2023) 483 (Suppl 1):S1–S391 13 Methods: A single pathologist selected 31 difficult-to-annotate regions of interest (ROIs) from gastric biopsies. The ROIs were independently reviewed by three pathologists in-person gathering, and individual cells were classified into 5 categories (positive/negative tumour cell, posi- tive/negative immune cell, and unclear). Fibroblasts/endothelial cells were excluded. Discrepancies were resolved by consensus among the three pathologists, and output was used to inform the roadmap. Results: Consensus building and discussions produced the follow- ing principles: a) nuclear size, shape, and chromatin are helpful for the recognition of epithelial cells and require assessment of tumour architecture and potential reference to the H&E stain; b) in cases with shared membrane positivity only one cell is considered positive; c) at the tumour-immune boundary, distinct membranous staining is required for tumour cell positivity; d) scoring of regions with extensive granular membranous staining should be avoided or staining repeated; e) multi- nucleated cells are considered as single cells; in case of uncertainty the least number of overlapping cells should be counted; f) cytoplasmic positivity in immune cells needs to be convincing at 20x. Conclusion: The roadmap outlined here provides the basis for standard- ized cell annotations to train AI algorithms and produce accurate scoring. This will provide a consistent framework for clinical usage. While PD-L1 is used as a case study, the principles may be generally applied to assist in a variety of single-cell annotations necessary to advance AI-based analy- sis of single or multiplex immunohistochemical stains. Funding: Bristol Meyers Squibb CP-01-007 Deep-learning model multiplexing CD68 virtual stain with digi- tal whole slide images of scanned PD-L1 22C3 pharmDx immu- nostained non-small cell lung cancer tissue slides S. Aviel-Ronen*, O. Ben-David, E. Arbel, F. Aidt, K. Kersch, T. Hagedorn-Olsen, D. Rabkin, I. Remer, A. Ben-Dor, L. Jacobsen, A. Tsalenko *Adelson School of Medicine, Ariel University, Pathology Department, Sheba Medical Center, Tel Hashomer, Israel Background&objectives: Tumour Proportion Score (TPS) of PD-L1 22C3 pharmDx (GE006) immunostaining in non-small cell lung cancer (NSCLC) excludes all immune cells including macrophages which is challenging. Our objective is to provide virtual stain multiplexing CD68 and PD-L1 to support accurate scoring. Methods: Forty-nine NSCLC tissues were sequentially stained and scanned with CD68 PG-M1 (GA613)-Envision FLEX HRP Magenta chromogen (GV925) on top of PD-L1 IHC 22C3 pharmDx and the whole slide images (WSI) were aligned. A deep convolutional neural-network model was trained and validated using matched pairs of PD-L1 (input) & PD-L1+CD68 (ground-truth) patches to create CD68 virtual stain based on PD-L1 stain. Results: CD68 staining highlights surprisingly more PD-L1 positive macrophages infiltrating the tumour than on initial estimation. Our model produced virtual CD68 staining resembling the actual CD68 stain, as quali- tatively assessed by pathologist. Most macrophages in the validation set were stained by the virtual CD68 stain. To quantitatively evaluate the contri- bution of the model to pathologist detection of macrophages we compared pathologist cell annotations as macrophages/not macrophage performed on PD-L1 stain only vs. multiplexed PD-L1/virtual CD68 stain, using annotations performed on PD-L1/CD68 sequentially stained slides as the ground truth. In ~4000 cells annotated, sensitivity/positive predictive value increased from 0.32/0.44 to 0.79/0.67 respectively. Paired McNemar’s test yielded a p-value of less than 10E-30. Conclusion: We demonstrated a promising CD68 virtual stain-mul- tiplexing model in NSCLC digital WSI, able to identify and virtually stain macrophages on an internal validation set. The model can be extended to virtually multiplex additional stains, including for other types of immune cells. Such virtual stain multiplexing could serve in the future as an assistive layer, allowing pathologists to score PD-L1 slides more accurately and reliably, and present an opportunity for studying the spatial relations between tumour and immune cells. Funding: Agilent Technologies CP-02 | Oral free presentations and Best Abstract Award CP-02-002 Transformer-based automated biomarker prediction from colorectal cancer histology S.J. Wagner*, D. Reisenbüchler, J.M. Niehues, J.A. Schnabel, M. Boxberg, T. Peng, J.N. Kather *Helmholtz Munich, Germany Background & objectives: Deep learning (DL) can extract predictive and prognostic biomarkers from routine pathology slides in colorectal cancer. However, current approaches rely on convolutional neural net- works (CNNs). More recently, transformer networks are outperforming CNNs in many computer vision tasks. Methods: We developed a new fully transformer-based pipeline for end- to-end biomarker prediction from pathology slides. We combine a pre- trained transformer encoder and a transformer network for patch aggre- gation, capable of yielding single and multi-target prediction at patient level. In contrast to previous methods, the transformer-based aggregation relates all tiles to each other instead of considering each tile individually. Results: A fully transformer-based approach massively improves perfor- mance, generalizability, data efficiency, and interpretability as compared with current state-of-the-art algorithms. By using the transformer-based aggregation module and stain colour augmentation during the preproc- essing, we can improve the prediction of MSI in colorectal cancer on the public dataset TCGA from 0.79 to 0.9 AUROC score compared to existing attention-based models. External validation on the public dataset CPTAC improves from 0.66 to 0.85 AUROC compared to state-of-the-art meth- ods showing significantly better generalization capabilities. The approach learns faster from fewer data samples than existing methods. Furthermore, multiple attention heads attribute high scores to different medical features increasing the interpretability of the model. Conclusion: A fully transformer-based end-to-end pipeline yields clini- cal-grade performance for biomarker prediction. Using a feature extractor pretrained on millions of pathology patches, stain color augmentation, and a new transformer-based aggregation module for MSI prediction on colorectal cancer histology significantly outperforms current state-of-the- art methods. Notably, our approach improves generalization to unseen cohorts, important for translation to clinical application. Funding: SJW and DR are supported by the Helmholtz Association under the joint research school “Munich School for Data Science - MUDS” and SJW is supported by the Add-on Fellowship of the Joachim Herz Foundation. JNK is supported by the German Federal Ministry of Health (DEEP LIVER, ZMVI1-2520DAT111), the Max-Eder-Pro- gramme of the German Cancer Aid (grant #70113864) and the German Academic Exchange Service (SECAI, 57616814). JNK and JNK is sup- ported in part by the National Institute for Health and Care Research (NIHR) Leeds Biomedical Research Centre. CP-02-003 Automated Ki67 hot-spot detection and analysis leads to higher Ki67 proliferation indices M. Zwager*, T. Koopman, H. Buikema, T. Ramsing, J. Thagaard, B. van der Vegt *Department of Pathology, University of Groningen, University Medical Center Groningen, The Netherlands Background & objectives: Visual identification and manual scoring of Ki67 hot spots is difficult and prone to inter- and intra-observer