ECP 2023 Abstracts

S69 Virchows Archiv (2023) 483 (Suppl 1):S1–S391 13 Results: The trained segmentation network, the feature extrac- tor achieved high performances with an accuracy, recall, precision, F1-score and mIOU of 0.987, 0.9742, 0.9742, 0.9742, and 0.95 respec- tively on the external validation sets. We further verified the disentan- gled representation of the encoder of the segmentation network using UMAP. The risk classifier achieved an AUROC of 0.7487 (± 0.0021) in predicting LNM status. The top features which affect LNM status were mostly geometric features from the total tumour and the undif- ferentiated tumour, including the area, and diameter of the total tumour and the perimeter of undifferentiated one. Conclusion: This is the first multi-institution study to develop machine learning algorithm for predicting LNM status in patients with EGC using H&E-stained histopathology images. Our findings have the potential to better help in selecting patients who need surgery among EGC patients showing high-risk histologic features. PS-02-016 Artificial intelligence-assisted daily quality control system for his- tologic diagnosis of gastrointestinal endoscopic biopsies: 1-year experience S. Yoo*, Y. Park, J. Jang, S. Yun, Y. Hwang, Y.S. Ko *Pathology Center, SeegeneMedical Foundation (SMF), Republic of Korea Background & objectives: In March 2022, SMF has launched an artificial intelligence (AI)-based quality control system (SeeDP) that double-checks all gastrointestinal endoscopic biopsy (EB) slides for possible incorrect diagnosis. We aim to review its operational records and clinical impact over the past year. Methods: Operational records were retrieved for the total number of EB specimens submitted, slides scanned and assessed by AI models, and cases with discordant assessments between the AI and pathologists’ diagnoses. Cases of which the diagnosis was revised after SeeDP’s suggestion were collected and compared to revised cases identified by conventional routes such as random review and clinician’s enquiry. Results: From 2022-03-01 to 2023-02-28, 67.7% (572,254/844,906) of EB slides were scanned. SeeDP failed to analyse 0.8% (4,531/562,203) of gastrointestinal EB slides due to various technical errors. AI’s judge- ment differed from pathologist’s diagnoses in 7.7% (42,760/557.672) of the cases assessed by SeeDP. Review of discordant cases revealed that true misdiagnosis accounted only for 5.1% (21/410) of the disa- greement, with most discordance attributable to inherent limitation of the current AI models. Compared to conventional error recogni- tion routes, SeeDP detected more misdiagnosed cases (14 versus 8) in significantly shorter interval of time (average of 3.6 days versus 38.8 days; P<0.001), including one signet ring cell carcinoma case initially diagnosed as gastritis. Conclusion: This is the first report of implementation and utilization of AI-based daily QC system for histologic diagnosis, and demonstration of its usefulness in routine clinical practice. Promising results were yielded over the past one year, but technical errors and unexpected events compromised 100% coverage of the QC system. The accuracy of AI models may not be improved under the current patch-based framework. Further effort is needed to systematically manage SeeDP coverage and to construct a more histologically relevant AI framework. PS-02-017 Interobserver variability in semantic segmentation for urothelial carcinoma S. Zurac*, B. Ceachi, L. Nichita, M. Cioplea, C. Popp, A. Cioroianu, L. Sticlaru, M. Busca, A. Vilaia, J. Dcruz, P. Mustatea, C. Mogodici *Colentina University Hospital, Carol Davila University of Medicine, Romania Background & objectives: Artificial intelligence (AI)-based algo- rithms for automatic detection in urothelial carcinoma (UC) are not available yet. Semantic segmentation for creating a dataset for UC requires heavily annotation on pixel level. We analysed the inter- observer variability and its consequences for annotation process. Methods: We selected 307 areas of interest (AOI) of minimum 1028x1028 pixels originating in 3 whole slide images (high-grade inva- sive UC; high-grade non-invasive UC; low-grade non-invasive UC); we used Cytomine application (Cytomine Corporation); 8 pathologists with various expertise in UC diagnosis & seniority annotated 13 differ- ent classes on each AOI (tumour-related, stroma-related, non-diagnos- tic, no_tissue, electrocoagulation). We evaluate annotations similarity with Sørensen–Dice coefficient (SDC). Results: SDC varied largely: low-grade tumour 0.91-0.93, smooth muscle 0.88-0.90, high-grade tumour 0.86-0.88, no_tissue 0.81-0.84, stroma 0.80-0.81, electrocoagulation 0.74-0.79, vessels 0.71-0.74, emboli 0.70-0.91, non-diagnostic 0.66-0.744, interstitial haemorrhage 0.55-0.78, invasion 0.47-0.73, inflammation 0.45-0.6. For invasion, the most similar pair of annotators (MSPA) had 40.74% annotations with similarity <0.5. Image analysis reveal an issue of interpretation: large tumoral areas were labelled as invasive in context. Different cut- off levels for inflammation gave the lowest SDC score. Higher cut-off levels for electrocoagulation influence less-experienced pathologists labelling of “non-diagnostic”. Low-grade UC annotations differed by pixel-level inconsistencies (manual delineation). Stroma differences (32.37% similarity <0.5 for MSPA) arise from delicate strands of col- lagen (pixels-wide) in low-grade non-invasive UC. Conclusion: Interobserver variability is significant when manual anno- tations are performed. Two main causes were identified: interpretations problems (with similarity scores less than 0.75) and technical problems due to manual delineation of each area (with similarity scores approx. 0.9). Interpretation issues (with considerable larger discrepancies between pathologists) must be mitigated in consensus debates in order to establish similar method of approach (i.e., “invasion” category) and similar cut-off levels (“inflammation”, “electrocoagulation” etc). This work was partially supported by a grant of the Ministry of Research, Innovation and Digitization, CNCS - UEFISCDI, project number PN-III-P4-PCE-2021-0546, PCE 109/2022, within PNCDI III PS-02-018 Quantitative aspects of neutrophils, Paneth cells, and adipocytes in digital H&E slides from uninvolved proximal ileal resection margins can predict post-operative recurrence in Crohn’s disease N. Zurek, D. Shiramizu, P. Gu, A. Mujukian, Y.J. Lee, K. Nurzynska, E. Chang, A.E. Walts, P. Fleshner, D.P. McGovern, A. Gertych* *Cedars-Sinai Medical Center, USA Background & objectives: Evidence suggests that histology at the proximal ileal resection margin (PIRM) has prognostic significance in post-operative recurrence (POR) of Crohn’s disease (CD). In this prospective study we used AI to analyse histologic features in PIRMs and to predict POR. Methods: Our AI pipeline identified neutrophils, Paneth cell gran- ules, stroma/smooth muscle, and submucosal fat in H&E slides from 139 PIRMs with no evidence of active CD per pathologist review. Extracted relevant features, and fractal dimension and adipocyte flattening in ROIs were analysed using Kaplan-Meier (K-M) plots and/or the rank sum test. Post-resection follow-up was >6 months in each case. Results: Utilizing average neutrophil density in the stroma/muscle, the K-M estimator significantly stratified patients into low- and high-risk groups for POR (HR=1.7, CI (1.11-2.61), p=0.0138). The stratifica- tion improved when this feature was combined with median Paneth cell granule count (HR=1.83, CI (1.19-2.81), p=0.0042), and when these two features were further combined with the percentage of submucosal fat pixels in the digital slides (HR=2.06, CI (1.34-3.17), p=0.00067). Analysis of submucosal fat ROIs from recurrent (n=17)