ECP 2022 Abstract Book

Virchows Archiv (2022) 481 (Suppl 1):S1–S364 13 Background & objectives: Nuclei detection in histopathology images is an important prerequisite step of downstream research and clinical analyses, such as counting cells and spatial interac- tions. In this study, we developed an AI-based nuclei detector using the YOLOv5 framework in whole-slide NSCLC cases. Methods: Our dataset consisted of 42 PD-L1 stained cases (30 training, 12 test). Four trained (non-expert) readers manually annotated all nuclei (both positive/negative) within regions of interest (ROIs) viewed at 40X magnification. We trained a YOLOv5(s) network on annotations of one reader. Performance was measured using F1 score analysis; hits were defined as being less than 10 um away from annotations. Results: We evaluate YOLOv5 on the test set by pairing it against all four readers separately. There, YOLOv5 performs excellently, falling within the interrater variability of the four readers: the mean F1 score over algorithm-reader pairs is 0.84 (range 0.76-0.92) while the mean F1 score over pairs of readers is 0.82 (range 0.76- 0.86). When we determine the cell count (number of annotations/ predictions) per ROI in the test set, agreement of algorithm-reader pairs and reader pairs is equally well aligned: 0.93 (range 0.90- 0.97) versus 0.94 (range 0.92-0.96). Visual inspection indicates YOLOv5 performs equally well on PD-L1 positive and negative cells. Conclusion: We have trained a nuclei detector that performs within the interrater variability of four human readers. In future work, we could extend this detector to additional tissues and immuno- histochemistry stainings. Moreover, this detector could be used as a AI-assisted manual point annotation tool: while human read- ers perform the (context-driven) task of delineating homogeneous regions (e.g. clusters of PD-L1 positive stained cells), the detector performs the (local, yet laborious) task of identifying individual nuclei within these regions, providing labelled point annotations. Funding: VIDI (F. Ciompi) PS-03-019 Automated annotation of digital H&E/SOX10 dual stains gen- erates high-performing convolutional neural network for cal- culating tumour burden in H&E-stained cutaneous melanoma P.S. Nielsen*, J.B. Georgsen, M.S. Vinding, L.R. Østergaard, T. Steiniche *Dept. of Pathology, Aarhus University Hospital and Dept. of Clinical Medicine, Aarhus University, Denmark Background & objectives: Deep learning for analysis of H&E stains requires a large annotated training set; a labor-intensive task that often involves highly skilled pathologists. We aim to develop and evaluate computer-assisted annotation based on digital dual stains of the same tissue section. Methods: H&E stains of primary (n=48) and metastatic (n=48) melanoma were digitized, re-stained with SOX10, and re-scanned. Images were aligned, and automated annotations of SOX10 stains based on thresholding and a trained convolutional neural network (CNN) were thus directly transferred to H&E stains of the training set (n=37). Training of the final CNN for calculating tumour burden included 1,221,367 annotated nuclei. Results: For primary melanomas, nuclei-annotation precision was 99.7% (95%CI=99.4%;99.9%) for tumour cells and 99.2% (95%CI=97.7%;99.7%) for normal cells. With a mean differ- ence of 7.9% (95%CI=6.1%;9.7%), precision for normal cells was markedly reduced for metastases compared with primary melanomas (p<0.001). Associated false-positive annotations were predominantly related to SOX10-negative tumour cells. Corre- spondingly, mean SOX10 intensity (red chromaticity) was 0.37 (95%CI=0.35;0.39) for primary melanomas and subcutaneous metastases but 0.32 (95%CI=0.30;0.34) for lymph-node and organ metastases (p=0.002). Accuracy of trained CNN for calculating tumour burden in primary and subcutaneous lesions was 92.6% (95%CI=83.6%;96.8%). Compared with stereological counting, mean difference in tumour burden was 5.8% (95%CI=-1.2%;12.9%, p=0.10) for CNN and 16% (95%CI=3.7%;28.3%, p=0.02) for rou- tine eyeballing. Conclusion: With this annotation technique, a large annotated H&E training set with high quality was created within a reasonable timeframe for primary melanomas and subcutaneous metastases. For these lesion types, the training set generated a high-performing CNN for calculating tumour burden, which was superior to routine eyeballing. Yet, due to low or missing tumour-cell SOX10 positiv- ity, advantages were limited in lymph-node and organ metastases. To include other cancer types or objects of interest, immunohisto- chemistry of the technique may easily be modified. Funding: Health Research Foundation of Central Denmark Region PS-03-021 Collaborative web platform: lab organisation, research and digital pathology practice V. Sousa*, R. Almeida, V. Almeida, M. Reis Silva, R. Jesus, L. Bastião Silva, D. Gonzalez, T. Adão, J. Carias, C. Costa, L. Carvalho *Institute of Anatomical and Molecular Pathology, Faculty of Med- icine of the University of Coimbra; CIMAGO – Research Center for Environment, Genetics and Oncobiology, Faculty of Medicine, University of Coimbra; University Hospital Anatomical Pathology Coimbra, Portugal Background & objectives: Digital pathology has become increas- ingly important in Pathology laboratory practice. The combination of whole-slide-based imaging techniques and machine learning has been greatly applied to develop visual analytics tools, with great contribute to a more accurate and time-effective practice. Methods: A consortium between BMDSoftware, Computer Graphics Center and Institute of Anatomical Pathology and Molecular Pathology - Faculty of Medicine/University of Coimbra, is developing a collaborative Web platform for digital pathology. It is a three years project that began in January of 2020 and joins the efforts of pathologists, computer science researchers and software developers. Results: The result is a cloud-based platform known as iPATH. It allows the visualization of whole-slide images, easy navigation through the images, annotations, delimitation of areas of interest, and a wide range of measurements. It aims the informatization of laboratory routine, tracking of all steps, from the reception of samples to the final pathologic report, including the management of response time. The platform allows the management of the datasets annotation process for production but also for supporting the development and integration of artificial intelligence tools. Cur rently, Helicobacter pylori identification and quantification instrument to apply to gastric biopsies and mitoses identification in different neoplasms are the challenges being considered as case-study. Conclusion: Digital pathology is already a reality in many Pathology labs across the world, perhaps driven by Covid-19 pandemic crisis, as it allows remote work. We believe that digi- tal pathology is just the beginning and a platform for the crea- tion of decision aid tools through artificial intelligence and deep S80