ECP 2023 Abstracts

S40 Virchows Archiv (2023) 483 (Suppl 1):S1–S391 13 degrees of fibrosis and inflammation (overall IoU 0.76 vs 0.66 for >50% inflammation/fibrosis). Vessels were the hardest to segment in severely fibrotic/inflamed biopsies where SB-M2F increased the IoU to 0.58 compared to 0.35 for R18-U-Net (IoU 0.22 for R50 U-Net backbone). SB-M2F resulted in visibly crisper and more uniform segmentations. Conclusion: We show that the widely used U-Net architecture with various ResNet encoders is outperformed by the recently released Trans- former-based architecture Mask2Former for multiclass segmentation of kidney histology. Importantly, we show improved performance in test biopsies with increasing extent of fibrosis and inflammation, lesions that are known to deteriorate segmentation performance. We have written code to extract object-based annotations from the SB-M2F output for further segmentation improvement in a Human-AI-Loop (HAIL) setting and easy object labeling in the open-source software “Slidescape”. Funding: JK is financially supported by grants from the Dutch Kidney Foundation (17OKG23) and the Research Priority Area Human(e) Ai of the University of Amsterdam. OFP-10-009 Computationally-derived stromal phenotypes, along with tumour- infiltrating lymphocytes, are associated with progression-free sur- vival in high-grade serous ovarian carcinoma digital pathology slides C. Walker*, L. van Wagensveld, J. Sanders, R.F. Kruitwagen, M.A. van der Aa, G.S. Sonke, K.K. Van de Vijver, S. Rottenberg, H.M. Horlings, A. Janowczyk * Institute of Animal Pathology, University of Bern, Switzerland Background&objectives: Tumour development is critically dependent on the supporting stroma and tumour-microenvironment. Here we investigate the prognostic significance of stroma phenotypes and the spatial arrange- ment of tumour-infiltrating lymphocytes (TIL) in high-grade serous ovarian cancer (HGSOC) using a fully-automated digital pathology pipeline. Methods: n=360 patients with advanced-stage HGSOC, treated with primary debulking surgery(PDS), or neo-adjuvant chemotherapy(NACT) and interval debulking(IDS). Stroma was delineated using a DeepLabv3 segmentation model, with subsequent extraction of texture features. TILs were detected using nuclei detection followed by an optimized ResNet18 for cell classification and analysed using graph-based features. Features were selected and validated using a 50:50 training/validation split. Results: Visual inspection showed high accuracy for stroma deline- ation and TIL detection. Using feature forward selection, 10 features were selected based on progression free survival (PFS) for stroma and TIL respectively. Partial log hazard estimation based on the stromal texture features and TIL graph and density features were both sig- nificantly associated with PFS (HR=1.07, CI=1.03-1.12; HR=1.09, CI=1.03-1.16) on the validation set. Furthermore, learning of a com- bined risk score on the training set based on the stromal and TIL based partial log hazard scores leads to a significant risk score (HR=1.16, CI=1.09-1.17, p<0.005), which remained significant in multivariable analysis (HR=1.12, CI=1.04 -1.23, p<0.005). Conclusion: Stroma composition may be an independent prognostic bio- marker in HGSOC for PFS. Furthermore, combining stroma texture features with TIL graph features is associated with PFS when adjusted for treatment, patient age and FIGO stage. Interestingly, visual assessment of these phe- notypes is likely challenging and subject to large inter-observer variability. Consequently, computational assessment of stroma phenotypes and TIL graph features provides the high-reproducibility needed for imaging-based biomarkers that may further help in stratifying risk of recurrence in HGSOC patients. OFP-10-010 Automated diagnostic coding (SNOMED-CT) from narrative pathology reports using natural language processing G. Cazzaniga*, V. L’Imperio, F. Pagni *Università di Milano-Bicocca - Department of Pathology, IRCCS San Gerardo, Monza, Italy Background & objectives: Pathology reports contain a wealth of information, but their unstructured and free-text format presents chal- lenges for analysis and knowledge extraction. In this study, we aim to automate the diagnostic coding (SNOMED-CT) from narrative reports with Natural Language Processing (NLP). Methods: We extracted the diagnosis text and corresponding SNOMED-CT over the past four years from the Laboratory Informa- tion System (LIS) of the IRCCS San Gerardo Pathology Department, Monza, Italy, excluding uncoded cases and retaining only the D (diag- nosis) or M (morphology) codes. The diagnoses associated with the 70 most frequent codes were selected to train a 3-layer LSTM (Long- Short-Term-Memory) network model. Results: The final dataset consisted of 36,855 well-balanced labelled diagnoses, with most represented categories being generic codes like “Chronic Inflammation” and “Negative for Tumour Cells”. The LSTM model was trained on 6 epochs and 64 batch size, showing an accuracy of 0.83 for the training set and 0.78 for the test set, with precision and recall of 0.78, and an F1-score of 0.77. Notably, best results were achieved in malignancies, while confounding factors were the pres- ence of adjectives as in the case of tubular vs tubulo-villous adeno- mas. Explainability graphs were used to identify influential monograms and bigrams for each category and fine-tune the model by identifying outliers. Conclusion: The present study demonstrates the feasibility of retro- spective classification and coding of a large dataset of narrative reports using NLP, which has potential applications in identifying unlabelled cases, automating the diagnostic process and monitoring disease trends in real-time. This approach, along with the prospective introduction of synoptic reports, can represent the basis to build an invaluable resource for the patients, especially through the desirable integration with com- mercially available LIS. OFP-10-011 Resizing and recompression of pathology whole slide images for affordable long term storage L. Alfaro*, M.J. Roca *Hospital Vithas 9 de Octubre, Spain Background & objectives: Digital pathology has spread to healthcare practice with innumerable advantages. However, its implementation is limited by high financial investment required in storage systems. In order to optimise its use, we designed an image recompression system to reduce long-term storage costs. Methods: Pathology cases are scanned for diagnosis at high resolution according to the specifications of manufacturers, with three scanners from different companies in .ndpi, .svs and .mrxs formats. Three months after the diagnosis, digital files are moved from the diagnosis server to a storage repository, after a recompression and resizing process. Results: We use three different software: Hammatsu NDP Convert, Aperio-Leica Digital Slide Studio and Objective Converter. JPEG2000 compression systems used at rates 10-20 (originally 80-90) and files are reduced to 25--33% of their original size. Depending on the degree of compression and resizing, resulting files weight varies between 5 and 15% with average of 10%. To assess the quality of these files, authors submitted the cases to patholo- gists with experience in digital diagnosis for re-evaluation. In all cases the overall assessment of the diagnosis was possible, and files were found to be valid for review in practice and for decision making. Only details requiring high resolution at maximum magnification presented limitations. Conclusion: The organisation of storage in digital pathology with redundant systems and backups that guarantee the integrity of infor- mation requires large and expensive storage environments.