ECP 2022 Abstract Book

Virchows Archiv (2022) 481 (Suppl 1):S1–S364 13 Sites of metastasis included liver, lung, and skin. Metastasis rates were as follows: 116 patients metastasised 0-24 months after diagnosis, 57 patients 24-48 months and 51 exhibited late metastasis ranging from 60-216 months after initial diagnosis. Patients died on average 1.5 years following metastatic diagnosis. Conclusion: From the analysis in our institution the incidence of metastatic disease corresponds with the European published figures. Our cohort includes a number of patients with a meta- static UM diagnosis prior to the age of 35. We found that 47.2% of patients treated for UM metastasised on average 43 months follow- ing diagnosis and the average time to death following metastasis was 1.5 years. Genetic analysis is currently underway to examine the contribution of specific genes role in early and late metastasis. CP-02 | Computational Pathology Symposium: Abstract presenta- tions and Best Abstract Award CP-02-001 Swarm learning for decentralized deep learning in gastric can- cer histopathology O.L. Saldanha*, H.I. Grabsch, H.S. Muti, R. Langer, B. Dislich, M. Kohlruß, G. Keller, J.N. Kather *UniKlinik Aachen, Germany Background & objectives: A limitation for computational pathol- ogy is the difficulty of data exchange. Swarm learning (SL) is a protocol for decentralized training of deep learning models. We evaluate SL for the prediction of microsatellite instability (MSI) from gastric cancer histopathology images. Methods: We collected tissue samples from four cohorts of patients with gastric cancer from four countries (Switzerland, Germany, the UK and the USA). Each dataset was stored in a physically sepa- rate computer. We trained a deep learning-based classifier to detect microsatellite instability using SL from digitized haematoxylin and eosin-stained resection slides without annotating tumour containing regions. Results: We evaluated the patient-level performance for the pre- diction of MSI status in the TCGA cohort (N=334 patients). We found that local models achieved AUROCs of 0.7016 (+/- 0.0087), 0.5600 (+/- 0.0238) and 0.6638 (+/- 0.0170) when trained on local datasets. Merging the three training cohorts on a central server (merged model) improved the prediction of AUROC to 0.7508 (+/- 0.0074). This was compared to the per- formance of SL-trained models, and we assessed the performance of a weighted Swarm Learning model (w-chkpt) for MSI muta- tion prediction. In this task, w-chkpt achieved an AUROC of 0.7469 (+/- 0.0214), which was not significantly different from the merged model (p=0.7806). Conclusion: Computational pathology problems in gastric cancer requires large datasets. Preferably, such data should be derived from different centres so as to avoid bias. However, the collection of such datasets faces practical, ethical and legal obstacles. These obstacles can be overcome using SL. In the future, this could be an alternative for sharing patient-related data across sites. CP-02-002 BLEACH&STAIN, a novel multiplex fluorescence immuno- histochemistry framework that facilitates a fast high through- put analysis of >15 biomarkers in more than 3000 human carcinomas E. Bady, K. Möller, N.F. Debatin, T. Mandelkow, C. Hube-Magg, N.C. Blessin* *University Medical Center Hamburg, Germany Background & objectives: Multiplex fluorescence immunohis- tochemistry (mfIHC) approaches were yet either limited to 6 markers or limited to a small (1.5cmx1.5cm) tissue size that hampers translational studies on large tissue microarray (TMA) cohorts. Methods: To assess more markers in a large patient cohort, we have developed a BLEACH&STAIN mfIHC approach that enables the analysis of ≥15 biomarkers in 3098 tumour sam- ples from 44 different carcinoma entities within one week and without costly instrumentalization. An artificial intelligence- based framework –incorporating three different deep learning systems– for automated marker quantification was used to inter- operate the BLEACH&STAIN data. Results: This approach was used to study the relationship between PD-L1 expression on multiple different cell types and the relation- ship with various leucocyte subtypes (PD-L1,PD-1,CTLA-4,pan CK,CD68,CD163,CD11c,iNOS,CD3,CD8,CD4,FOXP3,CD20, Ki67,CD31). Comparing the automated and deep learning-based BLEACH&STAIN PD-L1 analysis framework with conventional brightfield PD-L1 data revealed a high concordance in tumour cells (p<0.0001) as well as immune cells (p<0.0001) and an accuracy of our approach ranging from 90% to 95.2%. Unsupervised clustering showed that a major proportion of the three PD-L1 phenotypes (i.e., PD-L1+ tumour and immune cells [G1], PD-L1+ immune cells [G2], PD-L1 negative [G3]) were either inflamed (G1.1, G2.1, G3.1) or non-inflamed (G1.2, G2.2, G3.2) and showed distinct spatial orches- tration patterns. Conclusion: BLEACH&STAIN mfIHC in combination with a deep learning-based framework for automated PD-L1 assessment on tumour and immune cells enabled a rapid and comprehensive assessment of 15 biomarkers across more than 3000 tumour entities that is quick and easy to establish in all laboratories. In breast cancer, the PD-L1 relative expression on tumour cells showed a significantly higher predictive performance for overall survival compared to the commonly used PD-L1 tumour proportion score. CP-02-003 Pathologists’ first perspectives on barriers and facilitators of computational pathology implementation in histopathology J. Swillens*, I. Nagtegaal, S. Engels, A. Lugli, R. Hermens, J. van der Laak *Radboudumc, The Netherlands Background & objectives: Computational pathology algorithms detect, segment or classify cancer in whole slide images in histo- pathology. Currently, challenges have to be overcome before they can be used. We aim to explore international perspectives on the role of computational pathology in clinical practice. Methods: We will focus on opinions and first experiences regarding barriers and facilitators, which will inform establishment of validation studies, implementation trajectories and communication activities to generate widespread stakeholder acceptance. We conducted an international explorative eSurvey study and semi-structured interviews with pathologists and pathology residents. We used an implementation framework to classify potential influencing factors. Results: Results of the eSurvey showed remarkable variation in opinions regarding attitude, understandability and validation of computational pathology. Results of the interviews showed that barriers focused on the quality of available evidence, while most facilitators concerned strengths of using computational pathol- ogy. A lack of consensus was present for multiple barriers and facilitators, such as the determination of sufficient validation using computational pathology, the preferred function of computational pathology within the digital workflow and the appropriate timing of computational pathology introduction in pathology education. S56