ECP 2022 Abstract Book

Virchows Archiv (2022) 481 (Suppl 1):S1–S364 13 PS-03-012 Computer-aided algorithm and 3D imaging technology are sen- sitive methods in the diagnosis of HER2 expression-low breast cancers Y. Lee*, Y. Hsieh, S. Chang, Y. Chen, Y. Lin, Y. Lin *National Taiwan University Hospital, Taiwan Background & objectives: The developed antibody-drug con- jugates (ADC) show promising results especially in HER2-low expression breast cancer defined as immunohistochemically 1+ or 2+ with no gene amplification. More sensitive methods can be helpful in HER2-low samples diagnosis compared with traditional light microscopy examination. Methods: Two approaches were used to determine HER2 expres- sion. The computer-aided algorithm on digital pathological slides can determine HER2 expression levels. The 3D (three-dimensional) imaging approach used 100-μm thickness slide labelled by HER2 antibody with fluorescence and acquiring image under confocal microscopy with optical clearing method. Both methods can be successfully applied to retrospective and forward clinical studies. Results: In computer-aided approach, we developed a workflow including tumour recognition and HER2-positive cell counting based on 70 WSIs (Whole slide imaging) training dataset. Using 68-ROIs (Region of interest) cropped from 15-WSIs as validation, this method reached 86.7% accuracy and 94.34% sensitivity. 2 ROIs originally categorized as HER2-negative were reclassified as HER2-low by this method. In 3D imaging approach, HER2 fluorescent stained slides from the same 15 validation cohort were acquired under confocal micros- copy. One of four HER2-negative specimens originally categorized by IHC report was reclassified as HER2-low. Using 3D image, 2 of 15 specimens showed heterogeneous HER2 expression in different depth of their thick slides. Conclusion: This study demonstrated both computer-aided algorithm and 3D imaging technology were able to identify more HER2-low samples than light microscopy. These sensitive methods can detect cases with very low HER2 expression which are considered as HER2-negative by traditional light microscopy. For patients with very low HER2 expression identified by these sensitive methods, more studies are needed to see their clinical response to HER2 antibody-drug conjugates. PS-03-013 Differential diagnosis of Crohn’s disease and Ulcerative Colitis with deep learning based on hyperspectral infrared images T. Arto, F. Großerüschkamp*, T.M. Müller, A. Mosig, M.F. Neurath, S. Zundler, K. Gerwert *Ruhr University Bochum, Center for Protein Diagnostics (PRODI), Biospectroscopy, Germany Background & objectives: Differential diagnosis of inflamma- tory bowel disease (IBD) can be challenging but is important for treatment decisions and follow-up strategies. We aimed to establish label-free quantum cascade laser (QCL)-based infrared imaging combined with deep learning as a tool for differential diagnosis. Methods: Infrared imaging was used to analyse IBD and non- IBD cases. It is based on the interaction of electromagnetic waves with molecules within the tissue creating specific molecular fin- gerprints. A two-step deep learning approach, based on a modified U-Net (CompSegNet), was applied. The first instance differentiates IBD from non-IBD. The second instance differentiates between Crohn’s disease (CD) and Ulcerative Colitis (UC). Results: The cross-sectional sample set consisted of formalin- fixed, paraffin-embedded (FFPE) tissue sections from biopsies of CD (n=102), UC (n=52), and control cases (n=70). These cases were equally separated in the sub-cohort train (n=99), test (n=66), and validation (n=59). With the first instance CompSegNet, we achieved a validation area under receiver operating characteristic (AUROC) of 0.99 (sensitivity 95%, specificity 91%) in distinguish- ing between non-IBD and IBD. The subsequent second instance differentiation between CD and UC provided a AUROC of 0.89 (sensitivity 93%, specificity 83%). Conclusion: Our approach provides an objective and label-free tissue diagnosis in IBD distinguishing between CD and UC. With an increased number of patients and longer training phases, we expect a more accurate and robust classification. The combination of spatial and biochemical information encoded in the infrared images allows to track changes on the molecular level. Overall, this approach has the potential to become a widely applicable diagnos- tic tool for IBD and maintains intact tissue for further molecular analysis. Funding: This work was founded by a synergy award of the Ken- neth Rainin Foundation. PS-03-014 Generalisation of deep learning for breast cancer metastasis detection S. Jarkman*, M. Karlberg, M. Pocevičiūtė, A. Bodén, P. Bándi, G. Litjens, C. Lundström, D. Treanor, J. van der Laak *Department of Clinical Pathology, and Department of Biomedical and Clinical Sciences, Linköping University, Linköping, Sweden and Center for Medical Image Science and Visualization (CMIV), Linköping University, Sweden Background & objectives: For clinical adoption of AI, models must be robust when applied to new settings. Objective of this study was to test the generalization potential of a pretrained deep learning (DL) model to unseen data and to a new diagnostic domain. Methods: Whole slide images from Linköping, Sweden, with exhaustively annotated tumour regions, and publicly available CAMELYON data were used. Previously developed DL for breast cancer metastases detection in sentinel lymph nodes, developed using CAMELYON data, was used as baseline. The model was tested on sentinel nodes and lymph nodes from axillary dissections (n =51 and n=17, respectively; both Linköping). Results: Base model showed decreased performance on Linköping data (AUC 0.929; 95%CI 0.800-0.998 and FROC 0.744; 95%CI 0.566-0.912), compared to the performance on CAMELYON data. A large FROC decrease was found for the base model applied to axillary nodes. The model was retrained, using both CAMELYON and Linköping WSI, resulting in increased performance for both sentinel nodes and axillary nodes (both AUC p<0.05). Pathologist qualitative evaluation of the outputs of the retrained model showed no missed positive slides and in 21 of 24 positive slides slide- level diagnoses matched the clinical ground truth. False positives and false negatives were observed. One previously undetected micrometastasis was identified as a result of using DL. Conclusion: The study highlights the generalization challenge (even when using DL trained on multi-centre data), both as a result of applying DL in a new diagnostic setting for the initial indication, but even more remarkably, a slight change in indication impacted the model´s performance even more. Retraining the model, includ- ing data from target application, could mitigate the problem. Fur- ther studies are required to explore strategies to overcome the generalization challenge and evaluate what model performance is needed for different clinical applications. S78