ECP 2022 Abstract Book

Virchows Archiv (2022) 481 (Suppl 1):S1–S364 13 representative slide of each. We then used a built-in algorithm on Aperio ImageScope software [12.4.3.5008] to estimate the average nuclear size and the nuclear chromasia of cervical epithelium with HSIL and used statistical software to compare the data. Results: Our sample comprised 18 cases of HPV16, 26 cases of HPV-HR and 13 cases of HPV16 and HPV-HR coinfection. The presence of HPV16 genotype, alone or together with HPV-HR, was associated with CIN III (chi-squared = 5,28; p = 0,02). When comparing the HPV16 group to coinfected 16+HR, we found that HPV16 alone had significantly larger nuclei than the coinfec- tion group (t-test = 2,3; p = 0,03). However, these groups showed no difference regarding the nuclear chromasia. Conclusion: Digital image analysis applied to pathology seems a promisor field since it can provide us with essential information about our specimen that human eye simply can´t measure. In our study, the stronger association was between HPV 16 and CIN III but the clinical value of that is uncertain and may need a larger study. The same applies to the fact that HPV16 specimens have larger nuclei than coinfected 16+HR. PS-03-030 Using a digital platform to organize and manage pathology resi- dents curricula in Portugal F. Ramalhosa*, F. Pereira, C. Alves-Vale, D. Argyropoulou, C. Dahlstedt-Ferreira, C. Padrão, D. João, A. Lozada, G. Gerardo, J. Vaz da Silva, A.M. Gonçalves Pereira, R. Coelho, J. Gancho Figueiredo, R. Veiga *Serviço de Anatomia Patológica, Centro Hospitalar e Universi- tário de Coimbra, Portugal Background & objectives: In Portugal, for yearly and national board exams (NBE),a formal curriculum must be written down. In order to organize the registry of routine cases signed out during residency, a free internet platform dedicated to curricula organization and management (XERPA-MD) has been working together with Portuguese residents. Methods: Members of the residents’ committee of the Portuguese Society of Pathologists (NiSPAP) and an independent pathologist (coordinator of Anatomic Pathology for XERPA-MD) met regularly over a period of 12 months. A database, with neoplastic and non- neoplastic entities, was built according to the most recent editions of the WHO Classification of Tumours books, as well as other well- known bibliographic references in the field of surgical pathology. Results: The database follows the Portuguese College of Patholo- gists curricular guidelines. After three rounds of work (data input, data correction, and data review), a total of 5500 entities were organized. Users are now able to access the database online and to register all their curricular activities in a structured and systematic way. All inserted records can be reviewed, organized, managed, and exported automatically, saving several hours of work. Conclusion: Pathology residency is a critical period in the lives of all pathologists worldwide, with lots of study, routine cases and diverse scientific activities to be performed. The formal curriculum must include, among several other extensive registries, a thorough and organized list of all routine cases signed out during residency. Quite understandably, this becomes a daunting, exhaustive and stressful task, to be done every year, until the NBE. Automatiza- tion of curricula creation and management greatly decreases the time spent by Portuguese Pathologists writing down and organ- izing data. PS-03-032 AI(H): deep learning model for standing and grading autoim- mune hepatitis from histology C. Ercan*, K. Kordy, A. Knuuttila, X. Zhou, D. Kumar, P. Mes- enbrink, S. Eppenberger-Castori, L.M. Terracciano, M.C. Pedrosa *Institute of Pathology and Medical Genetics, University Hospital Basel, University of Basel, Switzerland Background & objectives: Autoimmune Hepatitis(AIH) is one of the most challenging diagnoses in liver pathology. We aim to develop a deep learning model, Artificial Intelligence for Hepatitis[AI(H)] that evaluates liver biopsies, to provide granular, quantifiable, and rapid analysis of histological features of AIH. Methods: One hundred twenty-five pretreatment liver biopsies with AIH diagnosis from the biobank of the University Hospital Basel were selected and split into training (80%) and test (20%) datasets and utilised to train several convolutional neural network models in the Aiforia platform. Manuel annotations of target regions were created by a hepato-pathologist, and used to train and test AI models. Results: The liver microstructure detection model was trained to segment liver tissue into portal or, lobular areas and central vein compartments, while the necroinflammation model was trained for focal necrosis, interface hepatitis, or confluent necrosis. The immune cell classification model can detect, classify, and quantify lymphocytes, plasma cells, macrophages, eosinophils, and neutrophils. The bile duct model was trained for detecting the damaged bile duct. The four AI models are accurate and efficient in diagnosis of vari- ous morphological components of AIH biopsies. When evaluated on a separate test set(ratios of correct predictions) of 92.9%, 97.1%, 84.5%, and 99.5% on liver microstructures, necroinflammation fea- tures, immune cell classification and bile duct damage detection, respectively. Conclusion: AI(H) is a novel diagnostic tool for AIH histology. It demonstrates comparable results to manual hepato-pathologist assessment for several specific diagnostic tasks on AIH biopsies and classifies cell/tissue types a much shorter time. AI(H) is an intelligent, fast, accurate, and efficient diagnostic tool. Further planned development of AI(H) will allow for more functionalities such as portal and lobular necroinflammation, specific inflamma- tory cells, fibrosis, and bile duct damage. Funding: Study was partially sponsored by Novartis. PS-03-033 An assisting deep learning tool for accurate detection of colo- rectal cancer lymph node metastasis A. Khan*, N. Brouwer, F. Müller, H. Dawson, J. Thiran, I. Nagtegaal, A. Blank, A. Perren, A. Lugli, I. Zlobec *Institute of Pathology, University of Bern, Switzerland Background & objectives: Histopathological evaluation of lymph node metastasis (LNM) for colorectal cancer (CRC) patients can be laborious and time-consuming. We propose an assisting deep learning method for CRC-LNM detection by leveraging transfer learning with an ensemble model on hematoxylin and eosin slides. Methods: The proposed deep learning method consists of an LN segmentation (UNet) and an ensemble (Xception, Vision Trans- former) metastasis detection models. LN segmentation model was trained on one hundred annotated CRC slides. An ensemble metas- tasis detection model was trained first on the public breast cancer dataset, PatchCamelyon and then fine-tuned on CRC data. Results: The proposed method was validated on an internal and external CRC cohort by analysing AUC, sensitivity, and specific- ity on a whole slide level. The method achieved an AUC of 98.1% in the internal validation cohort (2803 slides) with a sensitivity of 99.5% and specificity of 96.7%. Around 0.5% of positive slides were incorrectly classified as negative due to small isolated tumour S83