ECP 2023 Abstracts

S52 Virchows Archiv (2023) 483 (Suppl 1):S1–S391 13 OFP-13-011 Comparison of cell of origin (COO) classification of DLBCL patients using HTG EdgeSeq platform with conventional Han’s algorithm S. Feldstein (Davydov)*, J. Ben-Ezra, D. Hershkovitz, A. Erental *Tel Aviv Medical Center, Israel Background & objectives: This study compares the RNA-based HTG EdgeSeq assay with Immunohistochemistry (IHC)-based Han’s algo- rithm for subtyping diffuse large B-cell lymphoma (DLBCL) into GCB and ABC types, where ABC subtype correlates with shorter survival and poor conventional treatment response. Methods: The HTG EdgeSeq DLBCL COO Assay was used to evalu- ate 20 cases, including 15 previously analysed by Han’s algorithm and 5 reference samples. After RNA extraction and automatic library prepara- tion, sequencing was performed on the IonTorrent platform. Eventually, the HTG EdgeSeq System was used to analyse the data. The assessment of agreement between techniques was performed by Fisher’s exact test. Results: The study found 100% agreement in the 5 reference samples. Among the 15 cases previously analysed by Han’s algorithm, 7 were classified as ABC and 8 as GCB. Of the seven IHC-based ABC cases, 6 were confirmed by the HTG assay, but one was discordant and clas- sified as GCB. Of the 8 IHC-based GCB cases, 3 were discordant and classified as ABC by the RNA expression assay. FISH analysis showed double-hit involving c-MYC and BCL-6 in 2 of the 4 discordant cases, which are more commonly found in ABC, suggesting possible mis- classification by Han’s algorithm. Overall, 11 out of 15 (73%) samples showed concordant results. Conclusion: The combined diagnostic approach is used for DLBCL patients including immunophenotyping, FISH and COO classification, which are required for tailored treatment. The widely used Han’s algo- rithm reported 78% concordance with the gold standard Nanostring assay for COO classification. Our study demonstrated similar concordance with the HTG assay. As NGS platforms become more widely available in pathology, RNA expression-based assays are expected to become more significant in diagnostics and replace IHC-based methods of classification. CP-01 | Computational pathology – Where do we stand? CP-01-004 Deep learning-based multi-organ cancer and site classification using pathologic whole slide images of public datasets Y. Chong*, K. Yim, K.J. Seo, M. Alam, B. Kim, G. Hwang, D. Kim, O.R. Shin *The Catholic University of Korea, Republic of Korea Background & objectives: Accurate classification of cancer and its primary site is crucial for effective diagnosis and treatment planning. However, current artificial intelligence (AI) application in pathology focuses on the classification and subtyping within a certain organs or cancers. Methods: This study aims to develop a deep learning-based AI model for multi-organ/tumour-origin cancer and subtype classifica- tion using TCGA and CPTAC datasets, including 14 types of cancer samples. The dataset was divided into training, validation, and test sets for each organ and cancer subtype. Model performance was assessed using sensitivity, specificity, and accuracy metrics for each organ and cancer subtype. Results: Our model achieved an overall class classification accuracy of 89.98% and a cancer site classification accuracy of 97.83% (Ranged 97.23-99.60%). Overall sensitivity and specificity of cancer site classi- fication were 88.46% and 99.88%, respectively (Ranged 63.93-97.00% and 97.90-99.73%, respectively). Conclusion: Our model exhibits promising performance for most organs and cancer subtypes, assisting pathologists in diagnosing and treating cancer patients. However, further improvements can bemade for specific organs with lower performance and external validationwith real-world data should follow. This model can be incorporated into the general AI system as an ensemble with many preexisting or upcoming organ-specific/task-based models. CP-01-005 Metastatic melanoma immunotherapy response prediction from routine histopathology slides using digital pathology J. Bonjour*, A. Wicky, A.H. Seipel, S. Latifyan, P. Liakopoulos, M. Beaussart, O. Michielin, M.A. Cuendet, A.R. Janowczyk *Precision Oncology Center, Department of Oncology, Lausanne University Hospital, Switzerland Background & objectives: A clinically actionable biomarker iden- tifying which patients with metastatic melanoma may benefit from immunotherapy is sorely needed. Tumour-infiltrating lymphocyte (TIL) patterns may lead to such a biomarker. Here, we precisely characterized them at scale using digital pathology. Methods: In n=48 metastatic melanoma biopsies from patients treated with immunotherapy, deep-learning models identified TILs in pre-treat- ment 40x magnification Hematoxylin and Eosin (H&E) whole slide images. Hand-crafted TIL spatial organization features were extracted and assessed against overall survival (OS). Using leave-one-out-cross- validation, feature selection using Maximum Relevance Minimum Redundancy (MRMR) led to development of a predictive immuno- therapy benefit score via logistic regression. Results: Several features showed significant OS association: kurtosis of lymphocyte cluster areas (dichotomized at the median: HR=0.225, p=0.001, 95% CI=0.089-0.568), fraction of 500x500-pixel tiles with at least one lymphocyte (HR=0.308, p=0.006, 95% CI=0.127-0.742), and median distance between non-lymphocyte cells and nearest lymphocyte cluster (HR=0.390, p=0.023, 95% CI=0.168-0.907). These computer- only derivable features performed significantly better than those typically assessable visually by pathologists, such as TIL density (HR=0.500, p=0.089, 95% CI=0.220-1.130) and TILs to-all-cells ratio (HR=0.539, p=0.133, 95% CI=0.238-1.223). In a multivariate setting with clinical characteristics, TIL features remained independently predictive. The logistic regression-based prediction of immunotherapy benefit using six MRMR-selected features yielded an AUC of 0.78 (95% CI=0.61-0.93). Conclusion: Initial findings suggest that H&E TIL-based biomarkers hold promise in stratifying patient overall survival after immunotherapy. Com- putational digital pathology allows precise quantification of complex TIL features inaccessible via visual assessment, offering a novel opportunity for inexpensive, rapid, and non-destructive image-based biomarkers. These promising results indicate that an image-based biomarker using routine histopathology slides may aid in clinical treatment optimization, though further large-scale multi-site validation is necessary. CP-01-006 A roadmap for single-cell annotation using artificial intelligence algorithms – analysis of PD-L1 IHC expression analysis in gastric cancer biopsies S. Badve 1 , G. Kumar 2 , J. Ruschoff, H. Schildhaus, T. Lang, F. Faber, K. Daifallla, M. Karasarides 1 Emory University School of Medicine, USA, 2 Bristol Meyers Squibb, USA Background & objectives: There are no clear guidelines at a sin- gle cell level to annotate tumour, immune, and stromal cells in gastric biopsies for PD-L1 AI-assisted combined positive scoring. To resolve this, we have developed a roadmap to guide scoring consistency and accuracy.