ECP 2022 Abstract Book

Virchows Archiv (2022) 481 (Suppl 1):S1–S364 13 without nuclear atypia, lymphocytes and neutrophils, with thin and thick colloid. Numerous filiform structures were identified, with branching at acute angles raising the possibility of a fungal infec- tion, and rare hair follicles. The diagnosis was benign (colloid nod- ule) through The Bethesda System for Reporting Thyroid Cytopa- thology. After an exhaustive search by the reference books, it was not possible to determine the species name of these structures. The Google Lens application was used and allowed to select structures with similar morphology, confirming contamination by Verbascum sp. plant (which is indigenous in the patients residential area). Conclusion: The second cytology did not show the presence of these structures leading to the conclusion that it was a contamination. At the moment, no empirical evidence for using Google Lens in a medical setting is known to the authors. Nevertheless, Google Lens seems to be good for a quick evaluation, as smartphone cameras have become a widespread cost-effective method for pathology image acquisition and this software is freely available in our devices. E-PS-14-014 A robust, centerwise-adaptive hybrid machine learning approach for HER2 scoring R. Fick, C. Bertrand, A. Moshayedi, V. Di Proietto, S. Ben Hadj* *Tribun-Health, France Background & objectives: To enable precise human Epidermal growth factor Receptor 2 (HER2) quantification in digital pathology, we propose a hybrid deep- and machine learning approach, which exhaustively localizes invasive cancer in a slide and is adaptive to centerwise HER2 scoring criteria. Methods: Our dataset contains 350 Whole Slides Images (WSI) comprising 3 scanners. To quantify HER2 expression, we first localize all invasive cancer in a WSI using a U-NET. Then, we use hand-crafted features based on DAB colour devolution to determine HER2 expression in cancer. The slide label is determined by comparing aggregated expression with that of the laboratory’s HER2 calibration slides. Results: HER2 quantification is based on the staining completeness and intensity of all cancer cell membranes in the WSI, where the classes 0, 1+, 2+, and 3+ are distinctive in that specific percentages of cells are stained more intensely than the centre’s calibration slides. In practice these properties can only be eyeballed, leading to large inter-observer variability in HER2 slide classification. We compare eyeballing HER2 scoring with our algorithm’s evaluation based on clinical guidelines, calibrated on the centre’s calibration slides. Our results indicate high algorithm/eyeballing agreement for images with homogeneous HER2 expression within a slide, but less so for heterogeneous expressions. Conclusion: The results indicate that our AI-based HER2 quanti- fication can come to a different conclusion than expert eyeballing for hard cases. This may be the result of our exhaustive calculation of the total tumour surface, which is used as the denominator to determine the percentage of cells that exhibit HER2 expression. Visual inspection, in contrast, may underestimate HER2 non- expressive invasive cancer areas. In such cases, our approach may improve clinical treatment decisions based on verifiable algorithm outcomes. E-PS-14-015 Blended learning in histology using a virtual microscope: from adaptation to adoption at the University of Geneva P. Soulié, J. Perrin-Simonnot, L. Borgs*, R. Hoyoux, G. Vincke *Cytomine, Belgium Background & objectives: Priscilla Souilé and Jackie Perrin-Simon- not, two collaborators at the Faculty of medicine of the University of Geneva, teach histology using a blended learning model alternating online virtual microscopy and on-site activities to enrich and complete the learnings outcomes. Methods: The virtual microscopy practical work sessions are given using the open-source CYTOMINE software, integrated with the open-source MOODLE learning management system. The teaching assistants, supervised by the professors, are coaching the students using the virtual microscope and a chat space to exchange ques- tions and answers. Fully online during the COVID pandemic, this pedagogical strategy now also includes face-to-face remediation. Results: To maintain quality of histology teaching during the COVID pandemic period would have been impossible with a clas- sic model based only on practical sessions in face-to-face using microscopes. Beyond having taken up this challenge, and made it possible, this blended model based on web-based virtual micros- copy allowed teachers and students to collaborate with more auton- omy, transparency, and traceability, to ultimately having developed more persistent knowledges and skills. Alternating with face-to- face sessions after the pandemic allowed to keep the development of microscope related skills and develop more personalized reme- diation. This complementarity meets the objectives of both teachers and students. Conclusion: The histology practical work sessions during the pan- demic were quickly adapted to full e-learning based sessions thanks to virtual microscopy. Students, supervisors, and professors founded this virtual environment easy and intuitive, allowing them to quickly switch to new ways of teaching and learning, essential in these troubled times. In addition, it trains students to use tools that foreshadow what will be those they will use in their future career, in an easy and intuitive way. E-PS-14-016 The application of artificial intelligence in the diagnosis of prostate cancer A. Dovbysh, T. Savchenko, R. Moskalenko*, A. Romaniuk *Sumy State University, Ukraine Background & objectives: Prostate cancer (PC) is one of the leading causes of death in men from cancer. Artificial intelligence can reduce subjectivity and improve the effectiveness of diagnosing this disease using fewer resources than the standard diagnostic scheme. Methods: We have created a mathematical algorithm based on his- tological image recognition by machine learning and image recogni- tion methods. The input data for training and testing the mathematical algorithm were the results of PC histology obtained by biopsy. The structured vector consisted of nine characteristics of PC cells and one variable that captures benign or malignant tumours. Results: We have built a categorical functional model of informa- tion-extreme machine learning in the form of an oriented graph of sets used in the operation of the diagnostic system in the exam mode. The introductory recognition class is proposed, in relation to which the system of control tolerances for diagnostic features is determined, to choose the category according to the most signifi- cant variance of its input training matrix of brightness. With the increasing power of the alphabet of recognition classes, it is advis- able to move from a linear structure of input data to a hierarchical one, which will be the subject of further research. Conclusion: Comparative analysis of the results obtained by the authors allows us to consider the proposed method of information- extreme machine learning as a promising alternative to neuron-like structures in the analysis of large amounts of data. Funding: This research has been performedwith the financial support of grants from the Ministry of Education and Science of Ukraine No. 0122U000773 S302