ECP 2022 Abstract Book

Virchows Archiv (2022) 481 (Suppl 1):S1–S364 13 database, that can be accessed by PC, iOS and Android. The application of this study was carried out in 102 medical students from University of Fortaleza, being applied to a questionnaire through Google Forms seeking to evaluate the effectiveness of the learning strategy developed. Results: In a class of 102 students, 63 of them answered the form. More than 90% of the students considered that the flashcards were useful to the consolidation of the contents of Pathology and Radi- ology, and more than 90% marked that the flashcards contributed to a better revision of the same contents. The flashcards would be stored by 83% of the students to review the same contents in the future, letting the algorithms determine the dates of the next revi- sions. More than 90% of the students would use this tool to study other modules, and 50 of them would use it in tutoring, another methodology that requires a long-term revision. Conclusion: From the recent study, it was possible to demonstrate that the methodology used was relevant and useful for developing clinical reasoning based on automated spaced repetition learning focused on pathology and radiology exercises, which were made available on several mobile platforms, making the experience even more dynamic and easy to access. Furthermore, the systematization employed is a facilitator for the long-term consolidation of student knowledge, which in practice was confirmed by the data analysis obtained during the study evaluation. E-PS-14-004 Application of image analysis based algorithm for quality con- trol of immunohistochemical staining O. Greenberg, D. Hershkovitz*, R. Hagege, I. Hayun *Pathology Department, Tel Aviv Sourasky Medical Center, Tel Aviv, Israel Background & objectives: Immunohistochemistry (IHC) is a main tool in today’s pathology routine. Common applications include disease classification and biomarkers for targeted therapy. Positive controls are included on each IHC slide, misinterpretation of the staining in the control may lead to inaccurate diagnosis. Methods: To develop an algorithm for quality control of IHC we trained the algorithm to identify 3 patterns of IHC staining (nuclear, cytoplasmic and membranous). Fifty slides from each category (membranous Her-2-neu, nuclear TTF1, and cytoplasmic Cytokeratin7) were scanned. From these images, we captured 1,645 images, out of which 1,174 images were used for training phase and 471 for validation phase. Results: The algorithm was able to detect an average 87% of all the relevant stains. Specifically, the algorithm accurately identi- fied 98%, 85% and 78% of membranous, nuclear and cytoplasmic staining, respectively. Misclassification of cytoplasmic staining was mainly due to clear cytoplasm. Misclassification of nuclear staining was due to poor segmentation of adjacent nuclei. Conclusion: We develop an algorithm that shows high success in classifying staining patterns in scanned slides. Although the algorithm requires mild improvement, we estimate it could be a significant addition to the toolbox of digital pathology and be applied for quality control of IHC in the daily routine practice. Moreover with the correct usage of the algorithm, misinterpretation of the staining and inaccurate diagnosis could be easily prevented. E-PS-14-005 Homology-based approach for pathological diagnosis of pro- static cancer Y. Nomura*, K. Uchida, K. Nakane, M. Nishio, M. Ishizawa *Ise Municipal General Hospital, Japan Background & objectives: Prostatic adenocarcinoma (PCa) detection by image analysis from pathological specimens has been attempted. Although the systems based on the AI algorithm are very popular, we propose a newly established unique idea that is called “the homology profile method”. Methods: The digital data from prostate needle biopsy speci- mens at Ise Municipal General Hospital were binarized for each grayscale (0 ‐ 255) and calculated the homology index (the Betti number). The several profile, including its maximum value was featured. The staining condition was not in a consideration. The Betti number was calculated by an ordinally note type computer. Results: The Betti number was calculated for 100 JPEG format images for each of normal, Gleason pattern 3, 4, and 5. From the results of the t-test, the Betti number of PCa was significantly higher than that of normal prostatic glands. (p<0.0001). There was also a significant difference in Betti number of each Gleason pat- tern compared to normal (p<0.0001). The Betti number of Gleason pattern 5 was significantly higher than others. Conclusion: The homology is a mathematical concept that meas- ures “the contact degree” from the images. The homology profile method has not been applied to detect PCa using pathology images. This method is not only useful for detecting adenocarcinoma, but also for detecting Gleason pattern 5 PCa alone. Unlike AI algo- rithm, our idea is expected to be applicable to medical practice because of its small data size and reduced computation. E-PS-14-006 Efficacy and efficiency of a mitosis detection tool in invasive breast cancer C. Simmat*, L. Guichard, S. Sockeel, N. Pozin, M. Lacroix-Triki, C. Miquel, M. Sockeel, S. Prévot *Primaa, France Background & objectives: Mitosis counting is part of the Scarff- Bloom-Richardson (SBR) scoring for invasive breast cancer (IC). We evaluate the benefit of using an AI-based mitosis detection algorithm in the pathologist’s workflow. Methods: Our algorithm has two steps. A RetinaNet detector was first trained on a specially designed mitotic dataset of 4132 patches from 162 Whole Slide Images (WSI) to detect mitosis. An EfficientNetB0 classifier refines these results by removing a part of false positives. Results are displayed in our interactive viewer, Cleo. Results: 4 pathologists have used the solution on 50 WSI, with and without AI-results displayed. We evaluated the performance (F1-scores, Precisions and Recalls) and the time spent by pathologists on the WSI in both cases. The clinical study is ongoing and final results will be available in June. Actual algorithmic metrics are promising, but do not capture the value brought to pathologists as a proper interactive user interface enriched with the detector output will ease practitioners’ task and increase both recall and precision. Conclusion: To meet pathologists’ needs we developed a mitotic detection algorithm trained with routine data along with an interface designed with and for practitioners. Our clinical study will assess whether this tool can help pathologists in their daily practice. E-PS-14-007 Unsupervised stain adaptation in invasive carcinoma classifica- tion for breast histopathology using CycleGANs N. Nerrienet*, R. Peyret, M. Sockeel, S. Sockeel *Primaa, France S299