ECP 2022 Abstract Book

Virchows Archiv (2022) 481 (Suppl 1):S1–S364 13 learning technologies, that will improve the speed and quality of diagnosis processes. iPATH is a multidisciplinary project aiming to create useful solu- tions for pathologists and Pathology labs. PS-03-022 Upconversion nanoparticles as labels for histopathological tissue evaluation T. Nilsson*, K. Krawczyk, M. Mickert, S. Andersson-Engels *Lumito, Sweden Background & objectives: H&E staining and DAB-labelling are the gold standard in pathology, suffer from narrow dynamic range, difficulties in quantification and limited possibilities regarding multiplexing. We present an upconversion-nanoparticle (UCNP)-based technique that allows to overcome problems asso- ciated with commonly used labelling techniques. Methods: Formalin-fixed paraffin-embedded breast cancer cell line and human breast cancer tissue were sectioned and labelled. Upconversion imaging of the human tissue sections was con- ducted in our prototype device and compared with a standard DAB-based IHC. The combination of UCNP and haematoxylin counterstaining on the same slide was investigated. Results: Images obtained with our novel device demonstrate that our UCNP bioconjugates are excellent labels for the detec- tion of cancer markers in tissue sections. Brightfield images prove that UCNPs do not interfere with the standard tissue evaluation by a pathologist. Additionally, brightfield and lumi- nescent images can be merged to provide a better understand- ing of tissue morphology. Conclusion: The emerging field of UCNP-based labelling techniques provides new possibilities for more accurate diagnosis. Staining solutions and a novel device developed by us keep the advantage of H&E staining and combine it, in one image, with the UCNP luminescent data. The high-contrast images of the UCNP labelling – generated by our scanning device – set the foundation for generating ground truth for machine learning algorithms. PS-03-023 Computer-assisted diagnosis of early-stage lung adenocarci- noma using deep learning T. Trandafir*, J. Wolf, F. Akram, Y. Li, A. Dingemans, A. Stubbs, J. von der Thüsen *Department of Pathology and Clinical Bioinformatics, Erasmus MC, Rotterdam, The Netherlands Background & objectives: Early-stage lung adenocarcinoma growth patterns strongly associate with disease progression. Tumour biopsies are subtyped regarding microenvironment alterations and growth patterns: lepidic, acinar, papillary, micro- papillary, and solid. We developed a deep learning (DL) pipe- line to sub-classify adenocarcinomas to improve pre-operative assessments. Methods: We developed a multi-class DL classification model for the prognostically relevant patterns. A retrospective cohort of 129 whole-slide images of needle-biopsy sections of stage I and II lung adenocarcinomas stained with haematoxylin and eosin and their corresponding annotations for the regions of interest were used to train and validate our DL classification models. Results: In preliminary experiments, we designed a three-class DL model to classify normal tissue, tumour area of combined growth patterns, and tumour microenvironment. Compared to the ground truth, we reported a high overall accuracy of 0.86. Next, we designed a nine-class DL model to individually classify all patterns, yielding an overall accuracy of 0.77 and a Dice similar- ity coefficient of 0.66. We believe the results are influenced by class imbalance, introduced by dominant normal tissue, stroma, lepidic and solid patterns. Furthermore, high intra-class variability and inter-class similarity might have also influenced the results. For example, acinar patterns cover heterogeneous morphologies ranging from glandular to cribriform patterns, that may resemble lepidic patterns. Conclusion: Our results indicate the potential aggressiveness of early-stage lung adenocarcinomas from small biopsy sam- ples by sub-classifying growth patterns using DL. As we have a limited cohort, further training of the model on enriched data- sets is required. This analysis can guide the extent of the surgical approach to maximise the preservation of healthy adjacent tissue and increase the patient’s quality of life. However, biopsies might misclassify the dominant growth pattern due to sampling error. PS-03-024 A novel machine learning pipeline to analyse unstained liver biopsies and automatically quantify tumour-related structures R. Scodellaro*, D. Panzeri, E. Pagani, L. D’Alfonso, M. Bouzin, M. Collini, G. Chirico, D. Inverso, L. Sironi *Department of Physics "G.Occhialini", University of Milano- Bicocca, Italy Background & objectives: Tumour diagnosis is usually performed through the visual inspection of stained biopsies. Here, we propose a pipeline to support the pathologists’ clinical routine: it avoids staining procedures and provides novel quantitative insights to improve the diagnosis accuracy. Methods: Images of unstained liver biopsies, acquired by a whole- slide scanner, are virtually H&E-stained through a convolutional neural network (CNN). Relevant biological structures (e.g. dead hepatocytes, cell nuclei, collagen) and tissue dis-architecture (e.g. steatosis) are retrieved by exploiting semantic segmentation and texture analysis procedures, coupling the phasor approach with clustering and semi-supervised machine learning techniques. Results: The pipeline accuracy has been evaluated on 20 liver murine biopsies (10 affected by hepatocarcinoma and 10 healthy controls) by comparing the algorithm output with the pathologist’s quantification. In each biopsy, the amount of dead hepatocytes, cell nuclei and the extension of steatosis-affected regions have been automatically provided, with a mean accuracy > 95%. Moreover, the CNN performance in the virtual staining procedure has been evaluated for H&E samples. 2000 real and virtually stained tissue patches have been pixelwise compared, resulting in a colour content discrepancy < 5%. Conclusion: These preliminary results demonstrate the capabil- ity of the proposed pipeline to extract quantitative and objective information, expanding the tumour feature dictionary, assisting pathologists with a more accurate diagnosis and overcoming the limitations due to inter-observer variability. Moreover, time and resources consumptions due to staining procedures are avoided by the H&E virtual colouring provided by the CNN in the images pre-processing steps. PS-03-025 Segmentation of anthracosis – a needed first step in digital analysis of lung tissue P. Zens*, E. Baumann, A. Janowczyk, M. Kreutzfeldt, D. Merkler, S. Berezowska *Institute of Pathology, University of Bern, Switzerland S81