ECP 2023 Abstracts

S55 Virchows Archiv (2023) 483 (Suppl 1):S1–S391 13 Results: Data included 988,272 tiles (size=256x256 pixels) containing 303,318 IBC, 348,945 CIS, and 336,000 NT tiles from 403WSI. Classifica- tion performances of the independent test set containing 134,760 tiles yielded an average accuracy of 89%, average precision score of 0.89, average recall score of 0.89 and average F1 score of 0.89. The area under the curve comput- ing each class against the rest was 0.98 for IBC, 0.99 for CIS, and 0.95 for NT. Conclusion: Our results suggest that this novel deep neural network model can accurately diagnose IBC, CIS and NT on breast core needle biopsies and excisional specimens. If validated, this computational framework could be incorporated into digital pathology workflows to facilitate the diagnostic workup of breast specimens and integrated in biomarkers scoring algo- rithms. In addition, the learning frameworks of this model may be trans- ferred to applications in other disease sites. CP-02-007 NPIC AI FORGE: a unique multi-scanner digital pathology testbed for artificial intelligence development and evaluation M. Humphries*, D. Kaye, G. Stankeviciute, C. Dunn, H. Pye, A. Wright, D. Brettle, D. Bansal, D. Treanor *Leeds Teaching Hospitals NHS Trust and National Pathology Imaging Cooperative, Leeds, United Kingdom Background & objectives: Artificial intelligence (AI) has poten- tial to revolutionize pathology diagnosis; accuracy is dependent on the training dataset. Sensitivity to input changes reduces real-world performance. Generalizable AI requires diverse data obtained from multiple sources, scanners, and staining platforms. Methods: To address this problem, NPIC have created a unique multi- scanner facility, the NPICAI FORGE (Facilitating Opportunities for Robust Generalisable data Emulation). This facility comprises 15 scanners from 8 vendors as follows: Results: The NPICAI FORGE has broad capabilities including high-capac- ity WSI brightfield scanning of 1x3 inch slides; 3x2 inch slides; multiple magnification capture; z-stacking across 3 instruments, and WSI fluores- cence scanning. The NPICAI FORGE can generate over 3000WSI images per day, equating to over 5Tb of data. The NPICAI FORGE is supported by a research image management system, and secure research data repository with 15 petabytes of backed up storage, and disaster recovery mitigations, ensuring the safety, security and longevity of all digital pathology data gen- erated. Our PACS system presents as an easy to access online portal for rapid QC, viewing, annotation, project creation and sharing, and download. Conclusion: The ability to compare multiple scanning platforms provides an opportunity to map differences between data sets, to accurately measure the impact this has onAI specificity. We are able to accuratelyQC each system to specific calibration settings, ensuring robust and reliable data for AI training. Furthermore, unlike other areas of healthcare, slides can be re-scannedmulti- ple times, which offers an opportunity develop high-quality cross-platformAI. To our knowledge, this is the only such digital pathology facility in the world. All authors are funded by NPIC (Project no. 104687), supported by a £50m investment from the Data to Early Diagnosis and Precision Medicine strand of the government’s Industrial Strategy Challenge Fund, managed and delivered by UK Research and Innovation. (UKRI). MD-01 | Selected Abstracts MD-01-001 External Quality Assessment (EQA) for PIK3CA testing in breast cancer: lessons learned and the need for continued quality improvement V. Williams*, M.H. Cheetham, A. Maqsood, N.L. Wolstenholme, S. Patton *EMQN CIC, United Kingdom Background & objectives: Breast cancer is the most common cancer in women and may also occur in men. We present 3 years (2020-2022) of data from an EQA scheme designed to assess the quality of clinical testing for PIK3CA testing in breast cancer. Methods: EMQN sent artificial FFPE reference materials and/or patient derived FFPE tumour samples to each EQAparticipant for testing of clinically actionable PIK3CA variants using their routine methodologies. Anonymised results were assessed and peer reviewed. Individual laboratory and overall summary scheme reports were fed back to laboratories to help improve their performance and to enable comparison (benchmarking) of results and reporting. Results: Over the three years, there has been an improvement in the quality of testing, the clinical reporting of results, and standardisation of variant nomenclature used for reporting sequence variants. Improve- ment in the clinical interpretation of the impact of the analytical result is also evident; further work is still required to raise standards. Conclusion: Although errors persist, we conclude that annual par- ticipation in this EQA scheme can improve the quality of clinical diagnostic testing for breast cancer and contribute to the realisation of accurate personalised medicine. MD-01-002 Health care providers’ knowledge, confidence and practice of molecular pathology testing for oncology, within an Irish university hospital setting S. Kearney* *University College Dublin, St Vincent’s University Hospital, Innopharma Education, Ireland Background & objectives: Healthcare providers’ (HCPs) pathology knowledge was measured, via both a subjective appraisal alongside an objective assessment. Self-confidence and practice ordering, interpret- ing and explaining molecular pathology tests were evaluated, alongside HCPs’ views on clinical versus research utility of molecular pathology testing. Methods: The survey was conducted via electronic means, utilis- ing Google Forms, which was distributed to a wide variety of HCPs. 30 HCPs from a large academic teaching hospital in Ireland replied, including nurses, doctors in training, qualified consultants from a vari- ety of specialities and general practitioners. Results: More than half of respondents (65.6%) thought they knew less than enough about molecular pathology to do their jobs. Regarding con- fidence in the steps of ordering, interpreting and explaining molecular pathology tests, only a minority were fully or somewhat confident. Most HCPs were positively predisposed to utilising clinical informatics tools, genetic counsellors, second opinions services and molecular tumour boards (MTBs) in their molecular pathology workflows. Finally, the majority (83.9%) were supportive or strongly supportive that molecular pathology testing was important to their clinical practice, rather than research. Conclusion: Despite a small subset of respondents having high self- appraisals of their molecular pathology knowledge, when objectively measured most Irish HCPs lack knowledge and confidence in molecular pathology. This is important, given the growing prominence of molecular pathology in mainstream clinical practice, as reported by respondents. MD-01-003 Retrospective blind assessment of SOPHiA DDM™Dx HRD solution for evaluation of olaparib maintenance treatment efficacy in ovarian cancer patients from the randomized, phase III PAOLA-1 trial A. Buisson*, M. Lamkhioued, P. Constantoulakis, K. Oikonomak, P. Har- ter, S. Pignata, A. Gonzalez-Martin, C. Schauer, K. Fujiwara, I. Vergote, N. Colombo, E. Pujade-Lauraine, I. Treilleux, P. Saintigny, I. Ray-Coquard *Centre Léon Bérard, Lyon, France Background & objectives: SOPHiA DDM™ Dx HRD Solution com- bines analysis of genomic instability and homologous recombination