ECP 2023 Abstracts

S239 Virchows Archiv (2023) 483 (Suppl 1):S1–S391 13 the presence of blast cells, or recognising morpholgic patterns of AML in peripheral blood. When trained on monocentric datasets, generali- zation of algorithms to data from other settings can be challenging. Availability of training data from multiple sites is key to overcoming this hurdle using adapted training strategies. Conclusion: Deep learning-based diagnostic support algorithms for morphologic classification of haematologic samples have progressed in both diagnostic accuracy and computational efficiency in recent years. However, these systems must be evaluated for generalizability when used outside their training data domain. Key requirements for devel- oping generalizable algorithms are availability of diverse, multi-site morphologic data at train time and using training strategies optized towards generalizability and robustness. By comparing several differ- ent methods, we show how this aim can be achieved for hematologic classification tasks. E-PS-08-028 LymphoSight: an artificial intelligence QuPath companion applica- tion for the automated detection of tertiary lymphoid structures K. McCombe*, S. Craig, R. Gault, J. James *Queen’s University Belfast, United Kingdom Background & objectives: Tertiary Lymphoid Structures (TLS) are ectopic immune phenomena that occur in chronic inflammatory situ- ations including cancer. Research suggests there are prognostic and predictive of treatment. Here we develop of a deep-learning driven, “point-and-click” application for the automated detection of TLS. Methods: 1805 patient images across five cancer types were digitised, imported into QuPath and annotated for TLS. Images of TLS were extracted and split into training, test and independent validation sets at a 70-15-15 ratio at a patient level for model training. The model was evaluated on the independent validation set at whole slide level based on intersection-over-union and Spearman’s correlation. Results: 310 of the 1805 patients assessed possessed a mature TLS. Of these, 47 were set aside as the validation set and were matched with TLS-negative counterparts by cancer type. A strong correlation between predicted and expected TLS was achieved (Spearman’s r=0.853, p<0.0001). In the TLS-positive patients, an aver- age intersection-over-union (IOU) score of 0.716 was achieved, indicat- ing good model segmentation ability. Most false positives occurred due to gut-associated lymphoid tissue being detected as TLS in colorectal patients. Once these were manu- ally accounted for, IOU increased to 0.762 and Spearman’s correlation increased (Spearman’s r=0.943, p<0.0001). A “point-and-click” application was developed in Python to apply the model to images in QuPath projects. Conclusion: While studies have shown that the presence of TLS can be beneficial in multiple cancer types, the task of quantifying TLS in patient samples is often time-consuming, and potentially subjective. We take advantage of developments in digital pathology and artificial intelligence to develop a model to automate this process. In addition, we have developed a point-and-click application, LymphoSight, to apply the model directly to QuPath projects in a code-free manner, which may make such a model more clinically applicable. E-PS-08-029 Digital pathology shortens crucial steps of pathologist’s decision making Y. Molchanov*, A. Kobo-Greenhut, I. Barshack *Institute of pathology, Sheba medical centre, Israel Background & objectives: Digital pathology (DP) allows measure- ments of TAT at different stages. We tested the TAT in the pathologist’s work until decisions are made, One process of pathology can reduce TAT. Is the time between receiv- ing H&E slide to ordering immunohistochemical staining. Methods: Random lung biopsies were taken from a rapid diagnosis unit. H&E slide delivery to immunohistochemical stain ordering time was competed between 2018 and 2021 (before and after DP implementation). Results: The time from when the H&E staining was delivered until immunohistochemistry staining was ordered was shorter in May-July 2021 (M=0.98 days, SD=0.89) than in May-July 2018 (M=2.49 days, SD=2.04). This improvement was statistically significant, with t (54) = 3.59, p < 0.001 (one-tail). Conclusion: The laps time between slide delivery to release is important. After implementing DP and changes made to the work process, this time has shortened by more than 1.5 days.The time from H&E slide delivery to ordering of immunohistochemical staining been reduced by more than two times (from an average of about 2.5 days to an average of 0.98 days). E-PS-08-030 Future-proofing histological techniques in the Glasgow Tissue Research Facility H. Morgan*, J. Hay, J. Edwards *University of Glasgow, United Kingdom Background & objectives: Glasgow Tissue Research Facility (GTRF) bridges the gap between NHS Greater Glasgow and Clyde Bioreposi- tory, University and industry for tissue-based research, providing first- class automated tissue micro array (TMA) construction, digital pathol- ogy, image analysis and histology services. Methods: Creating a pathology-pipeline to meet increasing demands for high-quality research, the GTRF combines established histologi- cal protocols with brightfield and immunofluorescent slide scanning digital pathology techniques, and pioneering AI-based image analysis using Visiopharm®. TMAs enable analysis of multi-patient samples from highly characterised cohorts within single paraffin blocks, allow- ing standardisation of techniques including immunohistochemistry, RNAScope and efficient analysis of expensive ‘omics techniques. Results: Collaborations have generated TMAs examining tumour and stroma-rich regions of colorectal cancer (CRC) linked with clinical and genomic data from patents within the TransSCOT trial, and from pol- yps of patients within the INCISE project - designed to develop a com- prehensive risk stratification tool for CRC. Automation of TMA con- struction allows for precise block coring which may be used in DNA/ RNA extraction or placed in specialised maps, for research ranging from COVID-19 studies to targeting specific tissue regions of pancre- atic cancer which are subsequently used for downstream technologies such as GeoMx® and Visium®, thus allowing for highly specialised digital spatial profiling of gene and protein expression. Conclusion: Under the governance of the NHS Greater Glasgow and Clyde Biorepository, the GTRF works with researchers to enable access archived diagnostic tissue for a broad spectrum of research areas, pro- viding maximal preservation and use of limited and irreplaceable archi- val tissue samples. While the combined technologies and expertise within the GTRF and the collaborations we are involved in enables researchers to future-proof their research through adaption of ever- evolving techniques and technologies. Funding: Cancer Research UK (Scotland Centre) E-PS-08-031 Investigating tissue-source site-specific batch effects in H&E images for machine learning applications P. Murchan*, P. Ó Broin, A. Baird, A. Keogh, M. Barr, O. Sheils, S. Finn *The SFI Centre for Research Training in Genomics Data Science, Ireland Background & objectives: ML models risk learning confounding site-specific features in histology datasets. Our aim was to assess the performance of a ML model at predicting tissue-source site (TSS) from