ECP 2023 Abstracts

S238 Virchows Archiv (2023) 483 (Suppl 1):S1–S391 13 Funding: K_22 142604 grant of the National Research, Development and Innovation Office (Hungary) E-PS-08-024 Semantic segmentation of ductal carcinoma in situ in breast cancer histopathology whole slide images with deep learning K. Ledesma Eriksson*, J. Ho, S. Kang Lövgren, B. Su, Y. Wang, P. Weitz, S. Robertson, J. Hartman, M. Rantalainen *Stratipath, Sweden Background & objectives: Detection of invasive cancer(IC) in breast cancer whole slide images (WSI) is a pre-requisite in many com- putational pathology methods. Due to the high interclass similarity between IC-cells and ductal carcinoma in situ(DCIS), explicit distinc- tion between IC and DCIS regions is needed. Methods: This study investigates deep learning methods to detect DCIS in haematoxylin and eosin stained WSIs of cancer resection specimens from 346 female primary breast cancer patients diagnosed in Sweden. Regions of DCIS and invasive cancer were annotated by a clinical pathologist. DeepLabV3+ models were trained and evaluated using 5-fold cross validation at magnifications 0.625X, 1.25X, 2.5X, 5X, 10X. Results: At the best performing magnification of 5X, the AUC of the validation folds were 0.972, with a median slide-level sensitivity of 0.71 at a median specificity of 0.99 and a median slide-level accuracy of 0.91. Conclusion: The study shows the possibility to detect and segment DCIS in H&E stained WSIs with reasonably high accuracy using mod- els that operate at a single magnification level. However, due to the high interclass similarity of DCIS and invasive cancer cells, multi-resolution models considering the spatial distribution of the data has the potential to improve DCIS segmentation. E-PS-08-025 HistoBlur: a general deep learning tool for flexible and accurate blur detection on whole slide digital pathology images P. Liakopoulos*, L. Padayachy, M. Kreutzfeldt, S. Köfler, M. Tihy, R. Nair, M.A. Cuendet, O. Michielin, R. Langer, D. Merkler, A. Janowczyk *Precision Oncology Center, Department of Oncology, Lausanne Uni- versity Hospital, Switzerland Background & objectives: Blur artifacts may be introduced to whole slide images (WSI) by scanners during the digitization of glass slides. Here, an open-source deep learning tool is presented that allows for rapid, precise, annotation-free detection of blurry regions in WSI. Methods: One non-blurry skin H&E WSI at 10x was used to train a DenseNet model. Low/medium/high levels of Gaussian smoothing were applied to patches, yielding a supervised training task for blurri- ness identification. Medium/high blur was defined as levels at which tissue characteristics were indiscernible. For validation, n=111 slides of various tissue types (70 non-blurry, 41 blurry) had their percent blurriness computed. Results: Despite being trained on a single H&E image, HistoBlur con- sistently returned higher blurriness percentages for blurry WSIs regard- less of tissue type. For blurry WSIs, the median blur value was 54,95% (min=8.807%, max=100%). For non blurry WSIs, the median blur value was 0,9% (min=0%, max=15.25%). With a cutoff point of 8% blurriness, sensitivity for the detection of blurry slides was 100% and the specificity was 98,5%. The estimated computational throughput was 26mm2/s, which translates to ~30s for a typical biopsy. Preliminary results further show that due to its deep-learning backend, HistoBlur is likely to be stain/tissue agnostic, requiring only a single non-blurry representative slide of the target tissue/stain combination for training. Conclusion: Histoblur is an open-source tool for easily training and employing a Deep Learning model for detection of blurry regions on WSI via a simple Command Line Interface. These preliminary results suggest Histoblur enables rapid identification of poor-quality slides in clinical workflows autonomously, reducing technician overhead and improving diagnostic efficiency, for any stain/scanner/organ combina- tion. From a research perspective, the automatic detection and exclu- sion of blurry regions can mitigate adverse effects in building image- based biomarkers. HistoBlur is freely available (histoblur.com ). E-PS-08-026 Halo Breast AI, a deep learning workflow for clinical scoring of HER2, ER, PR &Ki67 immunohistochemistry (IHC) in breast cancer tissue M. Lodge*, A. Graham, A. Ironside, A. Polonia, S. Reinhard, W. Solass, I. Zlobec, P. Caie *Indica Labs, United Kingdom Background & objectives: The assessment of ER, PR, HER2, and Ki67, although associated with observer variability, is the cornerstone of treatment stratification for invasive breast cancer. Automated bio- marker quantification through HALO Breast AI aims to increase the speed and standardization of their quantification. Methods: The algorithm was trained using 107,328 pathologist-reviewed annotations to identify and threshold DAB-positive tumour cells within automatically segmented tumour regions. Technical performance was evaluated on 60,012 pathologist-reviewed annotations from unseen cases. Clinical performance was assessed by comparing the algorithm scores across whole slide images to either a pathologist consensus score (n=80) or the clinical report (n=200) respectively from two institutes. Results: The median image F1-score for tumour classification was 0.91, while the median image F1-score for cell-level validation was 0.96. The internal validation showed agreement between HALO Breast AI and three expert pathologists across the biomarkers: 95% for ER, 85% for PR, 85% for Ki67 and 80% for HER2. Performance on WSI obtained from an independent, external institute showed agreement between the scores obtained from HALO Breast AI and the clinical report: 96% for ER, 94% for PR, 84% for Ki67 and 84% for HER2 (91% 0/1+, 60% 2+ and 100% for 3+ scores). Conclusion: Combined immunohistochemical assessment of HER2 sta- tus, ER, PR, and Ki67 forms part of the routine clinical prognostic and predictive pathway for invasive breast carcinomas. Pathologist scoring of IHC at the microscope is time-consuming and prone to observer vari- ability. HALO Breast AI detects tumour regions and tumour cells within breast cancer tissue with high accuracy and clinical agreement when scoring routine diagnostic IHC. This product can support pathologists by improving workflow efficiency and standardizing results. E-PS-08-027 Robust and generalisable classification systems for haematologic malignancies in peripheral blood and bone marrow C. Matek*, R.M. Umer, C. Marr *Institute of Pathology, University Hospital Erlangen, Institute of AI for Health, Helmholtz Munich, Germany Background & objectives: Morphologic evaluation of leukocytes and their distribution in peripheral blood and bone marrow remain key steps in the diagnostic workup of haematologic malignancies. AI- based decision support algorithms have to be tested for robustness and generalizability prior to wider routine application. Methods: We tailored state-of-the art deep-learning methods for clas- sification support systems for both single and multiple leukocytes in patients with acute myeloid leukaemia (AML) and non-malignant con- trols. For single-cell classifiers, we assessed domain adaptation and domain generalization training strategies that use multi-domain data to develop more robust classification algorithms. Results: Both single- and multi-cell classifiers attain high performance in answering clinically relevant diagnostic questions such as identifying