ECP 2023 Abstracts

S41 Virchows Archiv (2023) 483 (Suppl 1):S1–S391 13 Long-term storage of information on already diagnosed cases doesn’t need maintaining the highest resolution for most of them, which will only occasionally be consulted again in future. With our recompression system, we reduce original size to 10%, which allows long-term storage at affordable prices and with quality if not optimal sufficient for re-evaluation of old cases. OFP-10-012 HPV detection in oropharyngeal squamous cell carcinoma: com- parison of morphology and artificial intelligence S. Craig*, V. Gaborieau, R. Gault, K. McCombe, A. Moyes, Y. Sun, T. Wang, A. Schache, T. Jones, J. Risk, P. Gunning, P. Brennan, J. James, B. Abedi-Ardekani *Queens University Belfast, United Kingdom Background & objectives: HPV infection has a key role in the pathogenesis and prognosis of oropharyngeal squamous cell car- cinoma (OPSCC). This study compares pathological assessment of HPV based on morphological features in H&E and p16-stained slides, against a novel artificial intelligence (AI) model. Methods: Based on an EfficentNet-b4 architecture, H&E images from a UK cohort were used to develop a deep-learning model for predicting HPV status. This model was independently tested in an unseen cohort of OPSCC in European, and American patients (n=418). HPV statuses were reported as positive when non-kerati- nising morphology with p16 strong expression in >70% of tumour cells was met. Results: Using p16 immunohistochemistry as a surrogate bio- marker for HPV status, p16-related OPSCC was identified in 45% (190/418) of European and American patients assessed. Non- keratinising morphology was observed in 37% of cases (156/418). Morphological features accurately predicted p16-related OPSCC in 77% (95% CI: 72-81) of cases with a sensitivity of 65% (95% CI: 81-90) and specificity of 86% (95% CI: 81-91). In contrast, by using a deep learning AI model to assess tumour morphology we were able to accurately predict 85% (95% CI: 82-89) of p16-related European and American OPSCC cases with a sensitivity of 85% (95% CI: 79-90) and specificity of 86% (95% CI: 81-90). Conclusion: This study highlights the subtly of phenotypic differ- ences in tumour morphology driven by high-risk HPV infection in malignant disease and the difficulty in recognising these features reli- ably by eye. Our AI model demonstrated significantly better accuracy in identifying p16-related disease compared to manual assessment of non-keratinising morphology. This model has the potential to be tested for usability in HPV-related squamous cell carcinomas from other sites and organs and offers a potentially clinically relevant tool for determining HPV status. OFP-10-013 Artificial intelligence’s impact on prostatic needle biopsies’ diagnostics E. Torresani*, L. Cima, M. Gentilini, E. Bragantini, F.G. Carbone, L. Maccio, M.G. Disanto, C. Sartori, S. Grassi, L. Morelli, T. Cai, M. Brunelli, C. Doglioni, M. Barbareschi *Department of Laboratory Medicine, Unit of Surgical Pathology, Ospedale Santa Chiara di Trento, APSS, Trento, Italy Background & objectives: Artificial intelligence algorithms applied to digital slides is promising to have huge impact on work’s quality and workflow in pathology. Aim of our study is to compare the performance of Paige Prostate AI to human performance in a routine setting. Methods: We selected 106 consecutive patients who underwent prostatic needle biopsies (1431 cores and consequently slides overall). Glass slides were digitalized with 3DHistech P1000 scanner and ana- lysed with Paige Prostate AI. We compared AI analysis with the diagnostic reports already provided by seven expert uropathologists and by one junior pathologist, consid- ering cancer status and Gleason Score grading as parameters of interest. Results: We compared AI analysis to three datasets: expert uropatholo- gists’ diagnosis, junior pathologist’s diagnosis on glass slides, junior pathologist’s diagnosis on digital slides. Agreement on cancer detection, considering each core individually, ranged from AC1=0.929 (CI 0.912-0.946) (expert pathologists vs AI) to AC1=0.937 (CI 0.912-0.946) (junior pathologist on digital slides vs AI). Agreement on Gleason Score grading, considering each core indi- vidually, ranged from AC2=0.881 (CI 0.851-0.911) (junior patholo- gist on digital slides vs AI) to AC2=0.923 (CI 0.898-0.947) (expert pathologists vs AI). AI’s sensitivity was 0.975 (IC 0.958-0.992) and specificity was 0.973 (IC 0.963-0.982). Conclusion: Paige Prostate AI has almost perfect agreement with pathologists’ performance, with excellent sensitivity and specific- ity values, supporting its eligibility for a routine use in pathology departments with a dedicated uropathology service. Its implementation in a laboratory workflow could have a great impact in standardizing prostate needle biopsies’ diagnostics. Moreover, it could ensure the dropping of inter- and intraobserver variability and finally could be used as a quality control tool. Funding: Internal APSS research funding. OFP-10-014 OvarIA: a deep learning approach for BRCA somatic mutations detection in high-grade ovarian cancer based on an innovative tumour segmentation method from whole-slide images R. Bourgade*, N. Rabilloud, T. Perennec, T. Pecot, C. Garrec, C. Delnatte, S. Bézieau, A. Lespagnol, M. De Tayrac, S. Henno, C. Sagan, C. Toquet, J. Mosnier, S. Kammerer-Jacquet, D. Loussouarn *Department of Pathology, University Hospital of Nantes, France Background & objectives: BRCA mutations represent an effective predictor of sensitivity of high-grade ovarian cancer to PARP inhibi- tors. However, their testing by NGS is costly and time-consuming. This work presents a novel approach for predicting BRCA mutations in ovarian cancer using deep learning. Methods: We included 775 patients with high-grade ovarian cancer and BRCA mutational status. A first step of tumour segmentation based on an innovative deep learning technique was performed and a total of 1,69M tumour tiles were predicted. We used 599K tiles to train a ResNet-50 with momentum contrast while 1,087M tiles were used to train the BRCA classifier with multiple-instance learning. Results: The tumour segmentation model trained on 8 whole-slide images obtained a Dice Score of 0.915 (± 0.05) and an IoU of 0.847 (± 0.079) on a testing set of 50 whole-slide images. The BRCA classifier achieved the state-of-the-art AUC of 0.739 (± 0.024) in 5-fold cross-validation, 0.681 (± 0.014) over the testing set, and 0.631 (± 0.03) over an external cohort from The Cancer Genome Atlas. We performed an additional multi-scales approach whose results have suggested that the relevant information for pre- dicting BRCA mutations seems to reside more in the tumour spatial conformation than in the cell morphology. Conclusion: Our results suggest that somatic mutations of BRCA have a phenotypic impact in high-grade ovarian cancer and this information lies more in the tumour context than in the cell mor- phology. Even if this study needs to be validated on a larger and multicentre cohort with homologous recombination deficiency, it paves the way to clinical application with the future implementation of pre-screening tools for a more personalized medicine.