ECP 2023 Abstracts

S66 Virchows Archiv (2023) 483 (Suppl 1):S1–S391 13 HER2-positive BCs. New therapies raised diagnostic issues with the emergence of HER2-low and HER2-ultralow categories. Our objective was to design an adapted Quantitative Image Analysis (QIA) algorithm. Methods: The evaluation used 133 slides providing from 2 external proficiency testing schemes (2021, n=52; 2022, n=81) conducted by the french interlaboratory comparison organization (AFAQAP) and comprising 8 different BCs with HER2 scores ranging from 0 to 3+. All statuses were assessed by 2 independent methods: visually by 3 expert pathologists, and by QIA using the IMSTAR PathoScan Tumour-Marker HER2 algorithm. Results: Expert pathologists identified 48 HER2 IHC slides receiv- ing an ”optimal technique” appreciation for each 4 BCs (2021, n=19 slides; 2022, n=29 slides). HER2-QIA algorithm performed an accu- rate evaluation of HER2 scores with an overall concordance of 94% with the experts for all optimal slides. The concordance between QIA and the experts was of 100% for HER2- positive BCs, 94% for IHC 1+ BCs, and 85% for IHC 0 BCs. Of note, concordance for the conventional IHC 0 BCs decreased from 100% in 2021 to 76% in 2022 with 7/29 BCs scored as 1+ by QIA, raising the question of the proper visual identification of HER2-physiological expression BCs, and HER2-ultralow BCs. Conclusion: Our QIA solution was efficient to identify HER2 expres- sion on technically optimal slides despite multiple laboratories IHC techniques being applied, highlighting the need for IHC standardization to obtain robust digital evaluations. HER2-QIA algorithm is accurate for the proper identification of HER2-positive BCs, while providing an objective, quantitative method discriminating between HER2-physio- logical expression, HER2-ultralow, and HER2-low BCs. In the lowest range of HER2 expression, it allows to generate HER2 new thresholds related to drug efficacy revealed by future clinical studies. PS-02-006 A colorectal AI model for triage and a second opinion in a digital workflow C. Girleanu*, R. Chetty, J. Weldon, M. Morrissey, R. Sykes, M. Colleluori, J. Fitzgerald, C. Durnin-White, S. Hutton, N. Mulligan *Mater Misericordiae University Hospital Pathology Dept, Ireland Background & objectives: Delayed diagnosis of advanced colorectal polyps result in increased morbidity and mortality. Rising workloads and pathologist shortages mean screening cases have reduced priority. A.I. driven case prioritisation within colorectal screening programmes may reduce turnaround time for clinically significant cases. Methods: A proprietary, weakly supervised learning classifier for slide-level, colorectal biopsy classification was created. The model was trained on over 15,000 whole slide images (WSI) of H&E colon biopsies obtained from The Mater Misericordiae University Hospital, Dublin (MMUH) at 20x magnification. This training cohort consisted of 32 unique biopsy diagnoses which were also provided by MMUH. Results: The model successfully triaged an unseen test cohort of 241 colon biopsies. This test cohort consisted of 156 urgent biopsies and 85 non-urgent biopsies. Of the 156 urgent biopsies, 20 contained colo- rectal cancer. The model classified each biopsy as urgent or non-urgent with 96% sensitivity and 92% specificity. Pathologist feedback on the 4% false negatives indicated that the majority were mislabelled ground truth and/or borderline cases. 100% of the biopsies containing colorec- tal cancer were classified as urgent. Conclusion: There is applicability of this A.I. for case triage and as a second opinion for decisions within a digital pathology platform that connects laboratories worldwide to an international network, where subspeciality pathologists can apply their expertise to clinical cases. This project received funding from the Enterprise Ireland Disruptive Technologies Innovation Fund (DTIF Project 2018_164) PS-02-007 Developing an efficient digital pathology cloud platform for collabo- rative image annotation and management C.K. Jung*, M. Yook, S.H. Lee, S.K. Yoon, I.Y. Choi *Department of Hospital Pathology, College of Medicine, The Catholic University of Korea, Cancer Research Institute, College of Medicine, The Catholic University of Korea, Republic of Korea Background & objectives: To integrate digital pathology into rou- tine diagnostic practice, efficient storage and processing of whole slide images (WSIs) data is crucial. Our objective was to develop a cloud platform with collaborative image annotation, multiple-user support, and managing large imaging datasets. Methods: Our dataset included mrxs, svs, ndpi, and tiff file formats scanned at 40x magnification. We customized Cytomine, incorporating annotation tools, clinical information, and pathology reports manage- ment. We tested a system that uses both cloud and on-premise servers to see how well it works for accessibility, cost-effectiveness, and how quickly images are displayed. Results: We compressed original WSI formats to tiff files using pyvips python, maintaining the same number of pixels as in the native WSIs. Compression with Q=50 reduced the file size by an average of 69% without significantly affecting visual quality, enabling data annotation. Using a hybrid cloud server system, native WSIs were stored on an on-premise server, while compressed tiff files were uploaded on object storage of a public cloud server. When used for annotation or viewed on demand, they were moved to the Network-Attached Storage (NAS). Annotations can be made directly on tiff WSIs in the NAS, accessible on any internet-connected device. XML ingestion/export options were added for local annotation. Conclusion: Our hybrid cloud system is cost-effective, sustainable, and offers fast and memory-efficient processing of large WSIs, with sup- port for annotation, visualization, and machine learning. Its flexibility, scalability, and ease of use make it a valuable tool for collaborative annotation, with options for multiple users and annotation types and for managing and analysing large imaging datasets. Its ability to support collaborative annotation across different platforms and to ingest and export annotations in XML provides additional user benefits. This researchwas supported by a grant (HI21C0940) from the KoreanHealth Technology R&DProject, Ministry of Health andWelfare, Republic of Korea. PS-02-008 Minimum cell counts for single gene testing – implications for tissue curation in molecular testing F. Kaar*, S. Conneely, A. O’Keefe, G. Callagy, S. Hynes *Anatomic Pathology, University Hospital, Galway, Ireland Background & objectives: Thresholds for molecular assessment of tumours vary depending on the technology used. The working thresh- old for single gene testing in most laboratories is 10%. However, the minimum number of tumour cells required for detection of a variant is not known. Methods: We analysed both colorectal and malignant melanoma cases with using a PCR based clinical assay for single gene assessment. Sen- sitivity was assessed using varying dilutions of the tumour cells. Abso- lute tumour cell and stromal counts were done using whole slide image assessment and QuPATH open-source software. Results: A colorectal cancer slide with known G12X and A146X vari- ants had 748,858 cells detected of which 441,962 cells were tumour cells. Based on dilutions of the target DNA a final minimal threshold of 189-2511 cells was estimated to be required for detection of these variants. A similar analysis of a BRAF V600E mutation in malignant melanoma was not able to find a dilution cut off point in order to illus- trate the threshold.