ECP 2023 Abstracts
S235 Virchows Archiv (2023) 483 (Suppl 1):S1–S391 13 100%/100%, respectively. The adenocarcinoma subtype classifica- tion model achieved a normalized discounted cumulative gain score of 81.8% (Top k+1). By comparing with the pathological reports, we found the model could accurately identify the cancerous areas and give pathological subtypes. Conclusion: We have established a complete AI-powered platform for lung cancer histopathological diagnosis and subtyping based on deep learning. The platform holds great promise for cost-effective and effi- cient clinical applications. The clinical application of the AI-assisted platform will help pathologists relieve workloads, avoiding missed diag- noses, and producing consistent reports. AI-assisted pathology diagnosis is expected to be widely adopted in hospitals in the near future. E-PS-08-012 Investigation of the applicability of mass spectrometry in patho- logical differential diagnosis B.A. Deák*, F. Csiza, G. Horkovics-Kováts, B. Gyöngyösi *Department of Pathology, Forensic and Insurance Medicine Sem- melweis University, Hungary Background & objectives: Metabolomics become one of the fastest growing research areas in recent years. Our aim was to investigate whether samples obtained through routine pathological processing are suitable for mass spectrometry measurements, and whether there are significant differences between different tissue types. Methods: First we examined 10 μm thick sections of pork liver in every step during the routine pathological processing. After that we analysed ccRCC and non-tumour renal parenchyma, obtained from the Department archive. Mass spectrometry analyses were performed with REIMS mobilization technique in negative ion mode, in the 50-1200 m/z range. Results: During processing of pork liver samples, fresh frozen sec- tions showed a rich and highly peaked lipid profile, which was not significantly influenced by formalin fixation. During routine patho- logical processing, we observed signal changes, decreases, and losses in the detected lipid spectrum at every further step of dehydration. Nevertheless, measurable signal was still obtained from FFPE samples, and the remaining peaks provided assessable signals. Principal compo- nent analysis could differentiate tumour and non-tumour renal tissue in human FFPE clear cell renal cell carcinoma samples. Conclusion: Mass spectrometry is a novel method for investigating the lipid profile of human samples. The gold standard of the method is measurement of fresh frozen samples, but formalin fixation alone does not cause significant signal reduction. Based on our preliminary results, spectra obtained through routine pathological processing may also be useful for distinguishing renal cancer and normal renal tissue. E-PS-08-014 The development and evaluation of a novel H&E stain quantifica- tion method C. Dunn*, D. Brettle, D. Treanor *National Pathology Imaging Co-operative, Leeds Teaching Hospitals NHS Trust, United Kingdom; University of Leeds, United Kingdom Background & objectives: Tissue sections are stained using histo- chemical stains for histopathological analysis. The staining process can introduce colour variability, yet current methods of quality control are subjective. We propose an objective method of H&E stain assessment to improve quality control in pathology. Methods: A stain quantification method is proposed, using stain assessment slides that objectively quantify haematoxylin and eosin (H&E) stains. To validate the use of stain assessment slides, they were characterised with a range of H&E staining durations and were imple- mented within eight clinical laboratories across a period of two weeks to analyse variation. Results: The stain assessment slides H&E stain response was linear with respect to increasing stain duration (r = 0.99). Clinical imple- mentation of stain assessment slides quantified intra-laboratory stain variation (average = 11%) and inter-laboratory stain variation (average = 26%). The impact of stain variation on automated nuclear counts was measured with and without colour normalisation. Conclusion: Stain assessment slides offer a quantitative method for measuring H&E staining in pathology laboratories. The results show strong linearity to H&E stain duration and the clinical utility in objec- tively quantifying stain variation in laboratories. This is important to improve quality control and standardisation of slides in laboratories, but also crucial for supporting the quality of digital pathology images and the growth of artificial intelligence in digital pathology. Funding: This research is part of the National Pathology Imaging Co- operative, NPIC (project no. 104687) supported by a £50m investment from the Data to Early Diagnosis and Precision Medicine challenge, man- aged and delivered by UK Research and Innovation (UKRI). FFEI Ltd and Futamura Chemical UK Ltd are partners on the NPIC program, and are providing in-kind contributions to the research activities in NPIC. E-PS-08-015 Construction of an extensive human skin dataset for artificial intel- ligence development J. Frias Rose*, A. Bodén, J. Van Der Laak *Department of Clinical Pathology, and Department of Biomedical and Clinical Sciences, Linköping University, Sweden Background & objectives: The Bigpicture consortium consists of members from both private and public organizations. Bigpicture’s main goal is to create the first European General Data Protection Regula- tion (GDPR) compliant platform, where quality-controlled whole slide images (WSI) and advanced AI algorithms will co-exist. Methods: In order to help achieve the goal of 3MWSI with their associated metadata, we chose to participate as contributors to Bigpicture, since the department of clinical pathology in Region Östergötland has a digital image archive of >2 Petabytes. We developed a protocol for dataset extraction, that complies with all applicable regulations, ensuring high quality content. Results: We were tasked with gatheringWSI from skin samples. A human skin dataset was designed, mirroring daily-basis clinical cases and their WSI as our contribution to the Big picture repository. After ethical approval for using patient data for research, we selected skin cases from patients 18 years and older, from 2019-2022 including cases with only one diagnosis amongst melanoma, other melanocytic lesions, squamous cell carcinoma, basal cell carcinoma, dermatofibroma, seborrheic keratosis, actinic keratosis and scar tissue. Metadata was partly preserved (patient age, anatomical site, acquisition time, laboratory related data and diagnosis/observations). All data anonymization, conversion and extraction was automated. A dataset of 45,000 WSI with their associated metadata was compiled. Conclusion: We succeeded in the compilation of an extensive clini- cally relevant dataset for Bigpicture’s repository, which will be useful for research purposes and development of relevant AI-solutions. The increasing adoption of digital pathology is an enabler for the develop- ment of AI-based tools that support histopathological diagnostics. The main limiting factor when developing and implementing AI-tools is availability of data, which can be attributed to challenges with data quality, storage and regulations for patient data protection, and Big- picture helps overcome said challenges. Funding: IMI Funding and EFPIA in kind E-PS-08-016 Automated quantification of stromal tumour infiltrating lympho- cytes is associated with prognosis in breast cancer M. Gonzalez Farre*, J. Gibert, P. Santiago, J. Santos, M.P. Garcia, J. Masso, B. Bellosillo Paricio, J. Albanell, B. Lloveras Rubio, I. Vázquez de las Heras, L. Comerma Blesa
Made with FlippingBook
RkJQdWJsaXNoZXIy Mzg2Mjgy