ECP 2022 Abstract Book
Virchows Archiv (2022) 481 (Suppl 1):S1–S364 13 data entry under 92 separate headings with no protocol to deter- mine categories of incidents. Therefore, the following consensus categories were created: specimen collection, requisition, packag- ing, transportation, reception, accessioning, specimen preparation and missing histology alerts. Conclusion: To streamline PAD documentation, manual logs were eliminated, all reporting was moved to LIS under the above eight categories and accessioning staff was trained on data entry under the new categories. Mandatory deficiency check has been imple- mented in LIS for the accessioning bench, before the specimen is further processed. Deficiency entry protocol for the grossing bench and histology lab are also currently being optimized for LIS integration in the next phase of the project. PS-03-004 NLP in diagnostic texts from nephropathology M. Legnar, C. Weis* *University Medical Centre, Germany Background & objectives: Nephropathology is a sub-discipline with complex diagnostic patterns and terms. In addition, reporting in a structured manner is a special feature. Against this background, we investigate whether predicting the final diagnosis based on the written description is possible. Methods: For his work, 1,185 unlabelled nephropathological reports were included. (i) First, the diagnosis sections were clustered unsupervised to <20 diagnosis groups. Therefore, bag of word-based and embedding- based text-vectorization methods were used. (ii) Second, different natural language processing (NLP) methods for classification were trained to predict based on the descriptive report section the diagnosis group. Results: Regarding text clustering (i), the silhouette-score and the classification performance of a support vector machine were used to measure clustering accuracy. For both, embedding-based approaches (best is a Bidirectional Encoder Representations from Transformer (BERT)) performed slightly better compared to bag of word-based (best is latent Dirichlet allocation) approaches. Ana- lysing the clusters for keywords shows that some clusters can be mapped to diagnostic groups. For example, there is a cluster for IgA-nephropathy or one for diseases with glomerular necrosis. Again, the BERT-based approach worked best regarding diagnosis prediction based on the histological description (ii). Notably, there are classification quality differences between the diagnostic clusters. Only some groups are almost perfectly predicted. Conclusion: For nephropathological reports, the morphological description alone enables retrieving the correct diagnosis for some entities. For other entities, this associative approach does not work well. This is in accordance with a previous study on glomerular change patterns, where some diagnoses are associated with one pattern, and for others, there is a complex pattern combination. Mapping every diagnosis cluster to a diagnosis group (or a chapter in a textbook) is still under investigation while writing this abstract. PS-03-006 A multi-feature AI solution for diagnosis support in gastric biopsies: a multi-site clinical study J. Sandbank*, J. Calvo, A. Nudelman, T. Garcia, E. Lanteri, J. Reyre, I. Laouar, B. Terris, M. Montagne, C. Rancati, M. Makla- kovski, A. Albrecht Shach, A. Arad, G. Sebag, R. Mikulinsky, T. Amit, I. Gross, M. Grinwald, C. Linhart, M. Vecsler *Institute of Pathology, Maccabi Healthcare Services, Israel Background & objectives: This study aimed to clinically validate the performance of a multi-feature AI-based solution on the detec- tion of gastric carcinoma, high-grade dysplasia and high-grade lymphoma, and Helicobacter pylori against rigorous ground truth (GT) established by multiple blinded pathologists in gastric biopsies. Methods: The Galen™ Gastric algorithm was examined in a pro- spective stand-alone performance study using retrospectively col- lected histopathology slides from two sites. We compared GT diag- nosis of adult gastric biopsies with the algorithmic results on H&E. GT was reached by concordance between two pathologists (original report and a new blinded diagnosis by pathologist reviewing slides/ WSIs). Discrepancies were adjudicated by an expert pathologist. Results: The AI algorithm demonstrated very high accuracy for the detection of gastric adenocarcinoma, high-grade dysplasia and high- grade lymphoma, with AUC of 0.986. Analysing 544 cases (82 positive), demonstrated sensitivity of 96.34%, specificity of 88.74%, and NPV of 99.27%. Additionally, the algorithm achieved an AUC of 0.966 for the detection of H. pylori in analysis of 525 cases (112 positives), with sensitivity of 91.07%, specificity of 90.56%, and NPV of 97.40% %. We will further report on additional pathologies, e.g., low-grade lymphoma, low-grade dysplasia, and Adenoma. Conclusion: This study reports the successful clinical validation of the Galen™ Gastric multi-feature AI solution in the accurate detection of a broad range of pathological features, including gas- tric adenocarcinoma, H. pylori, neuroendocrine neoplasms and more, offering an important tool for computer-aided diagnosis in routine pathology practice, supporting pathologists in their diag- nostic work. PS-03-008 Computer aided iron quantification on liver biopsy whole-slide images D. Panzeri*, R. Scodellaro, G. Chirico, C. Lancellotti, L. Di Tom- maso, L. Sironi *Department of Physics "G.Occhialini", University of Milano- Bicocca, Italy Background & objectives: Iron overload disorders diagnosis currently relies on blood tests and genotyping. However, liver biopsies remain an invaluable tool for prognostic purpose. We developed a pipeline to quantify iron deposits in whole-slide images (WSI) of liver biopsy marked with Perls stain. Methods: The study is based on 10 WSI of liver biopsies with different amount of iron. WSI were analysed on different scales to quantitatively and objectively reproduce the standard clinical procedure followed by pathologists. The cells stained by Perls stain (PS) were segmented and quantified by exploiting two methods: Optical Density (OD) colour thresholding and a custom spectral phasor approach. Results: Our pipeline is able to quantitatively retrieve: the per- centage liver biopsy covered by PS, the mean intensity of PS stain, the density and dimension of PS granules, PS heterogene- ity. Different background properties of WSI were buffered using a colour correction procedure. Moreover, heatmap overlays are generated to address statistically significant tissue patches (i.e., more densely stained, higher/lower intensity areas, PS deposits dimensions). Data generated by the pipeline had < 5% of error from pathologist’s evaluation. Conclusion: Our pipeline warrants a rapid and objective method for the evaluation of hepatic iron. Moreover, it can also be tuned and applied to different staining protocols (for example Picro- Sirius Red, used for the quantification of liver fibrosis). Finally, the estimated parameters are coupled to heatmap overlays in order to obtain intuitive graphical representations of parameter differences in the WSI. S76
Made with FlippingBook
RkJQdWJsaXNoZXIy Mzg2Mjgy