ECP 2022 Abstract Book

Virchows Archiv (2022) 481 (Suppl 1):S1–S364 13 PS-03-015 The use of digital pathology and artificial intelligence in the assessment of multiple myeloma M. McCabe, K. Sheehan*, S. Glavey *Royal College of Surgeons, Ireland Background & objectives: Plasma cell quantification in bone marrow trephines is an essential part of the diagnosis of multiple myeloma. Artificial intelligence (AI) is ideally suited for classifi- cation and quantification tasks but has not yet been employed in analysing the full myeloma microenvironment. Methods: Twenty-two trephines from patients with myeloma were retrieved from the archives and whole slide imaging performed at 40x (Objective Imaging). Using a deep learning convolutional neural network (HALO-AI, Indica), the algorithm was trained to segregate and quantify bone marrow tissue and cellular phenotypes. In addition, spatial analysis was performed to assess the relation- ship of plasma cells within the marrow microenvironment. Results: The trained classifier showed excellent segregation of marrow tissue elements as well as quantification of the different cell phenotypes on a Haematoxylin &Eosin stain. The mean number of plasma cells across the cases was 60,300 and mean density 1,637/mm2, occupying an average of 67% of the cell constituents. It was also noted that eosinophil numbers were increased (mean 4,900/trephine; 7% of all cells). Analysis of other marrow constituents demonstrated an inverse correlation between increasing plasma cell numbers and other cell types. Spatial analysis revealed a higher plasma cell density/mm2 with increasing distance from bone, although at a distance of 200uM, overall plasma cell numbers were reduced. Conclusion: In this study, we have successfully applied AI to the clas- sification of tissue types in multiple myeloma bone marrow biopsies. Using cellular phenotyping, it was possible to quantify plasma cells without immunohistochemistry, as well as other cell types in the mar- row microenvironment in a more reproducible manner. This is a useful diagnostic adjunct for Pathologists and enables us to further study the relationship between the pattern of disease burden and overall prog- nostic indicators for patients with this disease. PS-03-016 Optimization of automated tissue classif ication in histopathological images: use of a deep transfer learning approach on a pancreatic cancer cohort L. Haeberle*, R. Kronberg, M. Pfaus, H. Xu, K. Krings, M. Schlensog, T. Rau, A. Pandyra, K. Lang, I. Esposito, P. Lang *Institute of Pathology, Heinrich Heine University & University Hospital Duesseldorf, Germany Background & objectives: Neuronal networks (NNs) can assist with the analysis of digitalized histological slides. However, training of NNs can be hampered when training samples contain not one, but several tissue types. Our aim was overcome this problem by using deep transfer learning. Methods: Image tiles were extracted from tissue microarrays with samples from 223 pancreatic cancer patients. Tiles contained pan- creatic cancer, healthy pancreas, lymph nodes, but also confounders such as adiopse tissue. To purify the training data, we performed a data clean-up step using two communicating NNs (communica- tors). Subsequently, data was used to train NNs, which were then validated using an independent dataset. Results: By feeding pre-existing datasets containing confounders such as adipose tissue as well as our own training datasets into two communicating NNs (communicators), we received a selection of unequivocal tissue tiles for NN training. A ResNet-18-based NN re-trained with these data achieved a higher weighted accuracy over all tissue classes (94%) than after training with raw data (90%). Additionally, we tested 72 NNs using an independent dataset cre- ated from H&E-stained whole-slide images. NNs were able to distinguish between pancreatic cancer, healthy pancreas, lymph nodes and adipose tissue following training with purified data. Performances varied and depended on various factors, such as the learning rate and the optimizers used. Conclusion: Automated classification of histological tissue types in a pancreatic cancer cohort can be optimized by using communi- cator-driven data pre-processing. In the future, we aim to explore whether a similar approach can also be used to optimize other classification tasks, e.g., the distinction between pancreatic ductal adenocarcinoma and cholangiocellular carcinoma on digitalized H&E slides. PS-03-017 Inter-rater agreement of pathologists on determining cell-level PD-L1 status in non-small cell lung cancer L. van Eekelen*, E. Munari, I. Girolami, A. Eccher, J. van der Laak, K. Grünberg, M. Looijen-Salamon, S. Vos, F. Ciompi *Radboud University Medical Center, The Netherlands Background & objectives: Artificial intelligence (AI) based quan- tification of cell-level PD-L1 status enables spatial analysis and allows reliable and reproducible assessment of the tumour propor- tion score. In this study, we assess the cell-level inter-pathologist agreement as human benchmark for AI development and validation. Methods: Three pathologists manually annotated the centres of all nuclei within 53 regions of interest in 12 whole-slide images (40X magnification) of NSCLC cases and classified them as PD-L1 negative/positive tumour cells, PD-L1 positive immune cells or other cells. Agreement was quantified using F1 score analysis, with agreement defined as annotations less than 10 um apart and of the same class. Results: An average of 9044 nuclei (1550 negative, 2367 positive tumour cells, 1244 positive immune cells, 3881 other cells) were manually annotated by the three pathologists. The mean F1 score over pairs of pathologists at dataset level was 0.59 (range 0.54- 0.65). When split across classes, the mean per-pair F1 scores stay approximately the same, indicating the readers perform similarly regardless of cell type. Besides human variability in manual point annotations with respect to the centre of nuclei, lack of context contributed to disagreement: readers who reported they solely examined the ROIs tended to disagree more with readers that reported they also looked outside the ROIs for additional (mor- phological/density) information. Conclusion: Agreement on determining the PD-L1 status of individual cells is only moderate, suggesting a role for AI. By quantifying the inter-rater agreement of pathologists, we have created a human benchmark which may serve as an upper bound (and could be combined via majority vote) for the validation of AI at cell level, something not done previously. Cell-level AI-based assessment of PD-L1 may supersede slide level scoring, adding significant information on the heterogeneity and spatial distribution over the tumour. Funding: VIDI (F. Ciompi) PS-03-018 Nuclei detection with YOLOv5 in PD-L1 stained non-small cell lung cancer whole-slide images L. van Eekelen*, E. Munari, L. Dulce Meesters, G. Silva de Souza, M. Demirel-Andishmand, D. Zegers, M. Looijen-Salamon, S. Vos, F. Ciompi *Radboud University Medical Center, The Netherlands S79