ECP 2023 Abstracts

S243 Virchows Archiv (2023) 483 (Suppl 1):S1–S391 13 E-PS-08-046 Automatic detection and counting of pathological specimens with a rejection option A. Vieira*, D. Montezuma, T. Albuquerque, L. Ribeiro, F. Rebolo, D. Oliveira, J. Monteiro, S. Gonçalves, I.M. Pinto, J. Cardoso, A. Oliveira *Instituto Superior Técnico, INESC-ID, Portugal Background & objectives: Pathology specimen counting aims to ver- ify that the number of fragments on the slides remains unchanged after grossing. Since this is a manual, time-consuming procedure, our pur- pose was to develop an automated system to replace this manual step. Methods: We applied a state-of-the-art object detection model to detect fragments and sets, YOLOv5, trained and evaluated on 2554 and 700 WSIs, respectively, from a labelled dataset of different pathology sam- ples. Subsequently, we implemented several rules to improve counting performance, and added a rejection option when confidence was low. Results: The rule to reject the automatic counting is based on dividing the number of fragments by the number of sets. If this number is not an integer (indicating an inconsistency in the number of fragments per set), a warning is given, and that sample is not classified by the model. Without rejection, the model achieves an overall accuracy of 87.9%, which increases to 92.8% if we reject 10.9% of the samples (which must be reviewed manually). Conclusion: The obtained results are relevant because they highlight the importance of the use of a rejection option, which improves accu- racy on the automated reviewed cases while still enabling the reduc- tion of the manual workload. In future work, we will further improve the model’s accuracy (by a second counting round of the rejected and misclassified fragments) and apply a threshold between 0 and 1, ena- bling us to fine-tune the final score, according to the desired accuracy. Funding: This work was supported by national funds provided by Fundação para a Ciência e Tecnologia (FCT), under projects PRELUNA, PTDC/CCI-INF/4703/2021) and UIDB/50021/2020. E-PS-08-047 Can AI teach us new features of clinical relevance in colorectal cancers? K. von Loga*, C. Saillard, R. Dubois, O. Tchita, L. Guillou, A. Filiot, A. Fouillet, M. Sefta *Owkin, France Background & objectives: MSIntuit is the only CE marked AI-based MSI detection tool. The model scores known features of CRCs and highlights new characteristics that may impact prognostication and treatment decisions. The objective is to evaluate these by a systematic pathological inter- pretability assessment. Methods: The 4-step MSIntuit pipeline consists of tissue detection, tiling & normalisation, feature extraction and feature aggregation. 600 consecutive resected cases underwent the pipeline and tiles were pooled according to their assigned risk scores, clusters and heatmaps and then systematically reviewed by a pathologist for tile phenotyping, cluster analysis and spatial distribution. Results: A total of 600 whole slide images (MSI: n=123), correspond- ing to 11 million 112um by 112um tiles with risk scores ranging from 0 to 1 were available for the analysis. Risk scores above 0.5 are indicative of an MSI phenotype. >15 clusters were identified and pathologically classified into 10 relevant groups and correlated with risk scores. 400 tiles most predictive of MSI (n=200) and MSS (n=200) were phe- notyped and spatially located. The majority of tiles predictive of both contained tumour cells, with MSI: 70%, MSS: 60%. In the heatmap analysis, high risk tiles were also found in the surrounding normal tissue and further categorised. Conclusion: AI adds relevant quantitative information to known fea- tures of MSI and MSS that can be used to objectively score and com- pare CRC cases. AI highlights new areas of interests and features that could complement standard pathology reporting of CRC cases. AI systematically points out and scores surrounding normal tissue which has not been taken into account in standard pathology report- ing yet. E-PS-08-048 How to get decision trees from a knowledge graph in (nephro-) pathology C. Weis*, M. Legnar, J.H. Simoneit, S. Porubsky, Z. Popovic *Institute of Pathology Heidelberg, Germany Background & objectives: Having a diagnostic algorithm for all potential differential diagnoses is beyond the capability of a single pathologist. To remedy this, we test generating decision trees automati- cally from domain-specific knowledge (from nephropathology as a use case) stored in a knowledge graph. Methods: Nephropathology knowledge is stored from multiple sources (textbooks, diagnostic texts) in a knowledge graph, initially based on the SnomedCT ontology. Each disease or rather diagnostic finding is represented by a node. Eventually, different graph algorithms are tested to retrieve the path between two nodes, which should represent the diagnostic steps. Results: Generating the nodes and edges of our knowledge graph, we learned that a) standard entity recognition tools work. However, due to many different names of one entity, a correction step is needed; and that b) the relation extraction between two entities in pathological texts fails in many cases due to the typical semantic style of such reports. Therefore, we had to develop a custom relation extraction model based on the medspaCy toolkit. In addition, node classification approaches are used to learn relation- ships between diagnosis nodes and certain concept nodes. Decision trees are then extracted from these models. Currently, we test amongst other MINDWALK-tree to generate decision trees. Conclusion: Converting a text classification task to a node classifica- tion task benefits from additional relation information stored within a knowledge graph. However, knowledge graph generation can not be done with standard tools due to the special language of pathologists. E-PS-08-049 Conversion of a vendor-specific to a vendor-neutral quantitative biomarker digital image analysis system K. Yao*, M. Leong, F. Chung, S. Sanati, M. Cervania, V. Kazarov, M. Venturina, M. Gayhart, M. Guindi, B. Balzer, F. Dadmanesh, R. Matter, B. Pham, T. Lee, D. Frishberg *Cedars-Sinai Medical Center, USA Background & objectives: Scanner vendor neutral digital image analysis for cancer biomarker can offer significant advantages over scanner vendor specific platforms. Our aim is to document our insti- tutional experience with the conversion and to illustrate the recurring challenges with adapting digital image analysis. Methods: We converted our Aperio System using the Aperio AT Turbo scanner to the vendor-neutral Visiopharm Oncotopix Discovery using the NanoZoomer S360 scanner for oestrogen receptor (ER), progester- one receptor (PR), human epidermal growth factor receptor 2 (Her2), and Ki67 image analysis. A collection of previously analysed breast cancer slides was assembled for validation followed by a post go-live analysis.