ECP 2022 Abstract Book

Virchows Archiv (2022) 481 (Suppl 1):S1–S364 13 Conclusion: IP is a benign sinonasal tumour with malignant poten- tial. In our series, 1/22 had squamous cell carcinoma transforma- tion and 13/22 had more than one recurrences. The aetiology is yet unknown and probably multifactorial with a main role of HPV infection. Our findings suggest that the HPV status may be associ- ated with a higher risk of recurrences and dysplastic transforma- tion, but further investigation is needed. PS-03 | Poster Session IT in Pathology / Computational Pathol- ogy Symposium PS-03-001 MarrowQuant 2.0: clinical application of a user-friendly digi- tal hematopathology tool for human bone marrow trephine biopsies R. Sarkis*, O. Burri, C. Royer-Chardon, S. Blum, F. Schyrr, M. Cos- tanza, S. Sherix, C. Bárcena, B. Bisig, V. Nardi, R. Sarro, O. Spertini, S. Blum, M. Weigert, A. Seitz, B. Deplancke, L. de Leval, O. Naveiras *Laboratory of Regenerative Hematopoiesis, Institute of Bioen- gineering, Ecole Polytechnique Fédérale de Lausanne (EPFL) & Department of Biomedical Sciences, University of Lausanne (UNIL), Switzerland Background & objectives: Bone marrow(BM) assessment is a multiparametric evaluation of which cellularity constitutes one parameter. In diagnostic practice, the assessment is based on a semi-quantitative estimation which is time-consuming. In this study, we validated MarrowQuant2.0, within QuPath software, a digital hemopathology tool. Methods: MarrowQuant2.0 quantifies four compartments within the BM (hematopoietic cells, adipocytes (using Stardist), interstitiumμvasculature, and bones) and measures the cellularity of human BM trephine biopsies. We calculated the cellularity in a series of retrospective biopsies (training set n=36; experimental set n=157H&E). Using intraclass coefficient of correlation(ICC), specificity and sensitivity tests, we measured the agreement between MarrowQuant 2.0’s quantification and clinical reference. Results: Our algorithm was capable of accurate, rapid, and robust segmentation (average accuracy 0.86, n=36). There was an excellent agreement between MarrowQuant 2.0 and the clinical reference (ICC=0.978(95% CI0.955-0.989), R2=0.93). MarrowQuant 2.0 performed in a comparable way as to the clinical reference when used on a set of BM trephine biopsies from clinical routine diagnosis(n=42). We found reciprocity between the hematopoietic and adipocytic compartments in the context of an extreme case of BM remodelling, except for cases with stromal expansion. MarrowQuant2.0 can also leverage an adipocyte- based StarDist model, a deep-learning-based segmentation algorithm, implemented as an extension in QuPath, offering an accurate segmentation of individual adipocytes and a size-based classification. Conclusion: Our tool may represent a useful adjunct for experi- mental and clinical hematopathology. We will use MarrowQuant 2.0. to link output with clinical parameters using dimension reduc- tion and clustering methods to visualize and explore a potential prognostic value in myeloid malignancies. Funding: PHRT-Personalized and Health-Related Technologies PS-03-002 Convolutional neural network-based algorithm for the detec- tion and quantification of the components of the histologic grading of breast ductal adenocarcinoma - the first results G. Olteanu*, D. Kumar, M. Köteles, I. Mihai *Spitalul Clinic de Boli Infec ț ioase ș i Pneumoftiziologie Dr.Victor Babe ș Timi ș oara, Romania Background & objectives: The current grading system for ductal adenocarcinoma NOS of the breast (DAC-NOS), is prone to low reproducibility, subjectivity, and is overall time-consuming. Here, we trained a convolutional neural network-based algorithm (CNN- bA) to detect and quantify the components for grading DAC-NOS. Methods: 100 whole slide images (WSI) of diagnostic slides and 10 training slides (TS) with DAC-NOS were selected from the TCGA-BRCA dataset and subsequently uploaded to a WSI manage- ment server (Aiforia Technologies Oy, Helsinki, Finland). Briefly, the CNN-bA (Aiforia version 4.8, Aiforia Create, Aiforia Tech- nologies Oy) was trained on the 10 TS to detect and quantify the components for grading DAC-NOS. Results: After a successful model was established, the model was used to detect and quantify the histological components of DAC- NOS: tumour tissue (TT), tubule and gland formation (T + GF), solid tumour (ST) aspect, nuclear pleomorphism (NP) and mitotic count (MC). Next the results were exported and interpreted. Here, we report the first results of this pilot program. Successful training for TT, T, and GF or ST aspects was achieved after 3 iterations of the model. Unsuccessful training for NP and MC was reported and this represents the limitations of the results, and the baseline for future improvements that need to be addressed for the subsequent improved artificial intelligence (AI) model. Conclusion: AI in digital pathology represents the successful evolution of diagnostic pathology. While at times difficult, this transition is successfully implemented in in vitro diagnostics (IVD) platforms. Complex diagnostic algorithms used by pathologists like grading DAC-NOS are laborious to translate into an AI model. We report preliminary data (successful and unsuccessful data points) for a CNN-bA AI model that detects and quantifies components of the histologic grading of DAC-NOS with the aim of constructing the framework for a future IVD model. Funding: Aiforia Aiforwarding Program PS-03-003 Refining pre-analytic deficiency reporting and capture in the anatomical pathology laboratory: a quality improvement initiative T. Truong*, P. Roopchand, T. Jordan, Z. Ghorab, E. Slodkowska, M. Downes *Sunnybrook Health Science Centre, Canada Background & objectives: Pre-analytic, notably pre-laboratory, deficiencies are estimated to account for ~70% of laboratory errors. Herein, we reviewed pre-analytic deficiency (PAD) data at our institution to develop strategies for improved PAD documentation, allowing for identification of PADs and improvement of quality performance. Methods: We retrospectively reviewed 12 months of data at an academic, tertiary referral centre, which was captured from three sources: Laboratory information system (LIS), corporate risk man- agement reporting system and manual paper logs. We also inter- viewed accessioning staff to determine barriers to data recording. Results: 237 PAD were recorded in one year. Of these, 72% (n=171) were pre-laboratory and 28% (n=66) were in labora- tory. Specimen procurement accounted for 79% (n=134) of all pre-laboratory PAD followed by deficient requisitions, 12% (n= 21). Failure to adhere to specimen handling protocols accounted for 37.5% (n=25) of in-laboratory PAD with accessioning errors contributing to 23% (n= 15). Barriers to incident documentation were: time required, multiple systems being used and free-hand S75