ECP 2022 Abstract Book

Virchows Archiv (2022) 481 (Suppl 1):S1–S364 13 A. Romano Martínez Martínez, D. Jiménez-Sánchez, J.I. Ech- eveste, S. Martín Algarra, M.D. Lozano Escario, C.E. de Andrea* *Clínica Universidad de Navarra, Spain Background & objectives: Melanoma is one of the most frequent types of skin cancer, as well as one of the most aggressive, com- monly known for having very poor survival outcomes. Methods: Although lymphocyte presence in a tumour has been described as important, there is no standardization of the criteria pertaining to its analysis; it is highly subjective and dependent on the individual looking at the sample. Therefore, an annotation-free artificial intelligence technique, called NaroNet, was developed to identify melanoma cells and tumour infiltrating lymphocytes objectively. Results: It was then applied to samples in order to explore the differences between biopsies at their baseline and at progression. As NaroNet’s parameters determine the scope of tissue pattern learning, they were optimized to identify both cell types. Patch analysis approach (70x70 pixels) was used on 22-paired samples to detect cells of different morphologies and determine their distribution pattern. Each patch could hold one large melanoma cell and up to nine lymphocytes, which allowed the algorithm to discover interactions within their environment. NaroNet predicted baseline from progression with a 95.45% accuracy using phenotype abundances. T-distributed stochastic neighbor embedding of each phenotype showed high confidence when identifying cell types. Conclusion: Biopsies at progression had little to no brisk infiltration of the tumour, which was significantly different (p=0.03) from baseline samples. NaroNet can be trained to detect cell types without manual annotations, and used to identify important patterns of cell interactions in melanoma that could differentiate samples at baseline and progression. With further development, the algorithm could be optimized to become a useful tool for pathologists. OFP-11-011 Artificial intelligence-guided spatial transcriptomics in high grade serous carcinoma: toward image-analysis based precision oncology A. Laury*, S. Zheng, O. Youssef, J. Tang, O. Carpen *University of Helsinki, Research Program in Systems Oncology, Finland Background & objectives: H&E images of high-grade serous ovarian carcinoma (HGSC) may contain prognostic information detectable only by artificial intelligence (AI). We hypothesise that AI can pinpoint regions with prognostic value, and the biology of these regions can be revealed with spatial transcriptomics. Methods: The cohort included 55 stage III-IV patients with distinct platinum-free intervals (PFI) (<6 vs >18 months). A deep learning neural network tool identified tumour regions most indicative of out- come. These high-confidence (HC) and background regions were probed with 10x Visium for FFPE spatial transcriptomics technology. Output was visualized with 10x Loupe browser and analysed using the R package Seurat. Results: The neural network was first trained to identify tumour tissue, then to classify the tumour into short or long PFI group. Using a HC mask, regions indicative of outcome were identified. These HC regions were then used to train the final neural network. Testing a combined inference pipeline to classify an independent tumour set showed high sensitivity (73%) and specificity (91%). UMAP visualization of the spatial transcriptomics demonstrated that while data from the same patient are close to each other, HC and background regions are mostly distinct within the cluster for each patient. Transcriptomics profiles from HC regions predicted PFI group status significantly better than background regions. Conclusion: Artificial intelligence-based image-analysis (AI-IA) of HGSC tissue can identify morphologic patterns invisible for human eye and guide selection of biologically meaningful regions for spatial transcriptomics. When combined, these novel technologies identified several signalling pathways and transcripts separating HGSC tumours with short vs. long PFI. In conclusion, AI-IA together with spatial transcriptomics offers a promis- ing toolkit to identify biological features associated with cancer behaviour, making the AI-based diagnosis more interpretable and clinically relevant. Funding: This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 965193 for DECIDER. OFP-11-012 Development of trial quality assurance program for digital pathology of the Korean Society of Pathologists Y. Chong*, J.M. Bae, D. Kang, K.I. Kim, H.S. Han *The Catholic University of Korea, Republic of Korea Background & objectives: Digital pathology (DP) can fundamentally change the way of working in pathology. Since the Korean Society of Pathologists (KSP) published the consensus recommendation paper for DP application recently, the need for quality assurance program (QAP) for DP has been raised. Methods: To provide standard baseline reference for internal and external QAP for DP, the Committee of Quality Assurance of KSP developed a checklist for DP QAP and started a trial QAP in 2021. After several revisions, the checklist was finalized. Five leading institutes participated the trial QAP in the first year and we gathered feedback from these institutes afterwards. Results: The newly developed checklists of QAP for DP contains a total of 39 items (212 score) to check, 8 items for quality control of DP systems, 3 items for DP personnel, 9 items for hardware and software requirement for DP systems, 15 items for validation, opera- tion, and management of DP systems, and 4 items for data security and personal information protection. Full text in both Korean and English is attached as appendices. Most participant institutes in the trial QAP replied that continuous education on unfamiliar terminol- ogy of new technology and more practical experience is demanding. Conclusion: QAP for DP is essential for the safe implementation of DP in pathologic practice. Each laboratory should prepare institu- tional QAP according to this checklist and consecutive revision of the checklist with the feedback from trial QAP for DP needs to follow. Funding: This study was supported by the Korean Society of Pathologists Study Group/Committee Supporting Research Grant (2020). OFP-11-013 Detecting premalignant lesions in the Fallopian tube, using a deep-learning model. A pilot study J. Bogaerts*, J. Linmans, M. van Bommel, M. Steenbeek, J. Bulten, J. de Hullu, M. Simons, J. van der Laak *Radboud University Medical Centre, The Netherlands Background & objectives: Serous Tubal Intraepithelial Carcinoma (STIC) is a precursor lesion to High Grade Serous Carcinoma (HGSC). Interobserver variability in STIC diagnosis is high. We aim to develop an Artificial Intelligence model that can detect STIC in digitalized whole slide images. Methods: We collected, digitalized and annotated 91 cases of STIC/ STIL and 75 control cases. Diagnosis was confirmed using p53 and Ki-67 immunohistochemical stains, when available. The cases were S47