ECP 2022 Abstract Book

Virchows Archiv (2022) 481 (Suppl 1):S1–S364 13 Background & objectives: Anthracosis, the black granular pig- ment in lung tissue, is often wrongfully detected as nuclei by image analysis algorithms, leading to high false positive rates in immu- nohistochemical slides. Here, deep learning models for preemptive removal of anthracosis were evaluated. Methods: From n=8 CD8 stained whole slide images, 128 tiles (256x256 px at 0.2431 μm/px) were manually selected and annotated to reflect expected lung tissue heterogeneity. Tiles were used in 4-fold validation to comparatively evaluate a traditional U-Net vs. Xception-based U-Net model (31e6 vs 2e6 parameters). For image augmentation, we focused on colour augmentation. Model performance was assessed by Dice score. Results: The traditional U-Net outperformed the Xception-based model (0.78 ± 0.22 vs. 0.74 ± 0.22), with its unaugmented ver- sion performing the best (0.85 ± 0.16). Qualitative assessment showed that the models were more precisely segmenting individ- ual granules versus the coarse annotations provided in the ground truth, suggesting superior performance over those suggested by quantitative metrics. Faint anthracotic pigments on darker back- ground (condensate macrophages) and intensively stained CD8- postitve lymphocytes were common sources of error, with the dark combined hematoxylin and DAB at the nuclear membranes being detected as false positive anthracosis. Conclusion: We show that simple U-Net-based models are power- ful tools for localization of anthracosis. These models should likely be included in image analysis pipelines to help eliminate biologi- cally irrelevant artifacts, thus improving specificity of downstream analyses. Augmentation methods did not appear to improve the model in identifying potentially relevant morphological features, suggesting that colour is insufficiently discriminatory in many instances. Next steps will include the refinement of our anthra- cosis model, including more targeted augmentation methods, and combination with nuclei segmentation. Funding: MD-PhD scholarship 5088-06-2020 of Cancer Research Switzerland PS-03-027 Deep learning neural networks for real-time discrimination between osteosarcoma and fracture callus on conventional histological (H&E) sections S. Stavropoulos, K. Gourdoupis, S. Georgopoulos, V. Plagianakos, D. Papachristou* *University of Patras, School of Medicine, Unit of Bone and Soft Tissue, Patras, Greece Background & objectives: Histopathological discrimination between osteosarcomas and fractures can be extremely challenging. Herein, we aimed at developing a novel Machine Learning architecture (Convolutional Neural Network) for real-time discrimination between osteosarcoma and fracture callus based on conventional H&E sections under brightfield microscopy. Methods: We collected 2136 H&E images from 1154 osteosar- comas and 982 fractures from our archives (n=925) and the web. With the data collected, we trained a Convolutional Neural Net- work (CNN) based on the mobilenet version 3 architecture. Train- ing set included 2021 images (1086 osteosarcomas, 935 fractures), and test set 115 images (68 osteosarcomas, 47 fractures) from new/ unknown to the system cases. Results: We created a web application that allows microscopes to directly connect to either desktop/laptop computers or mobile phones, in order to facilitate easy access to the Neural Network classifier. The proposed system is simple to use and delivers real-time high accuracy classification regarding the separation of osteosarcoma from fracture, with sensitivity 91.3% and specificity 93.75%. Four (4) osteosarcomas were falsely classified as fractures; among them 2 were postchemotherapy osteosarcoma cases with 100% necrosis. Two (2) fractures were falsely classified as osteosarcomas. Conclusion: CNN trained on conventional H&E sections can significantly decrease the pathologist’s workload and serve as an additional tool towards accurate diagnosis of several pathologies in everyday-routine practice. Since this tool is based on conventional H&E images and not on WSI, it can be easily used by pathologists from remote areas and small hospitals for the diagnosis of osteo- sarcomas, but also for virtually every pathologic condition, after specific training. PS-03-028 Novel approach for adaptive colour normalisation based on tissue thickness and biological tissue type variability J. Jain*, P. Perugupalli, R. Gupta, D. Dodle, S. Krishna, V. Rao *Pramana, Cambridge, MA, USA/India Background & objectives: Colour variation in H&E and IHC slides poses a huge challenge for computational pathology algo- rithms. We present a novel method for colour normalisation based on tissue thickness and biological tissue variability to achieve sig- nificantly improved performance over existing methods. Methods: 30 H&E slides of endometrium tissue were scanned using Pramana WSI scanner (inline analysis of tissue thickness and colour) and data was collected for colour and tissue thickness variation using colour distribution map and tissue thickness graphs. Colour normalisation was done based on tissue thickness, biologi- cal tissue variability and colour differences between thick and thin tissue areas in the slide. Results: Visual evaluation by histopathologist revealed that our method yielded better results in comparison to existing methods. Our method does selective normalisation in the AOI as per the requirement after consideration of vectors for tissue thickness, biological tissue variability and colour distribution of the target AOIs instead of adopting the generic approach uniformly across the slide like existing methods. Reference pairing based on close matching of the colour distribution range of different colour in the target AOI to the reference AOI has improved the performance when biological tissue variability vector was also considered. Selective application of this approach based on tissue type and thickness reduces the over and under colour saturation. Conclusion: Inline analysis of tissue thickness variation and colour hue distribution for all the AOIs in a WSI can help in achieving better colour normalisation when biological tissue variability is taken into consideration. Target to Reference AOI pairing done on the basis of biological tissue type and colour distribution graph allows for selective colour normalisation in the target AOI if needed. Both tissue type and thickness variation across the AOIs determine the need of selective colour normalisation for optimal results. PS-03-029 Digital image analyses applied to HSIL cervical biopsies A. Alves*, V. Almeida, R. Almeida, M. Reis Silva, V. Sousa, L. Carvalho *Centro Hospitalar e Universitário, Portugal Background & objectives: Pathology has been integrating tech- nology into its workflow and there are several algorithms we can runout to generate reliable information. We aimed to determine if nuclear features of cervical HSIL differ among patients infected with distinct HPV genotypes. Methods: We randomly selected 57 HSIL (CIN II and CIN III) cer- vical biopsies with previous HPV genotypification and scanned one S82