ECP 2023 Abstracts

S68 Virchows Archiv (2023) 483 (Suppl 1):S1–S391 13 immunohistochemistry assays requiring expert visual assessment. We aim for standardized and automated tissue-saving COO prediction using H&E whole-slide-images (WSIs) and interpretable machine learning methods. Methods: Multiple field-of-view (FOV) images were randomly extracted from tumour regions in WSIs. For each FOV, cellular features characterizing tissue morphologies were extracted and used to train a random-forest (RF) model for FOV prediction. The averaged FOV prediction is reported per WSI. As a comparison, an attention-multi- instance-learning (AMIL) network using FOV image embeddings gen- erated by a pre-trained ResNet50 is also investigated. Results: The RF and AMIL models were developed using a 40X WSI dataset, which contained 120 activated-B-cell-like (ABC) and 236 germinal-centre-B-cell-like (GCB) subtype cases determined by genomic testing. The RF model achieved an average ROC-AUC of 0.675 on a 5-fold cross validation, and the AMIL model achieved an average ROC-AUC of 0.687. On a separate test dataset containing 22 ABC and 32 GCB cases, the RF model achieved a ROC-AUC of 0.715, and the AMIL model achieved a ROC-AUC of 0.737. In addition, RF model’s feature importance revealed that tumour cell spatial distri- bution features consistently had high impact on the model prediction across training and testing. Conclusion: In this work we demonstrate that tissue morphological features are correlating to the biology driving COO determination. A RF model employing explicit morphological features and an AMIL model using image embeddings from a convolutional neural network can achieve comparable COO prediction performance. While the AMIL model provides interpretability through locating high attention regions in a WSI, the RF model enables identifying impactful cellular level features and provides insights for further investigation of model trustworthiness. This study was funded by F. Hoffmann-La Roche AG, Basel, Switzer- land. Support with compliance and the abstract submission process was provided by PharmaGenesis Cardiff, Cardiff, UK, and funded by F. Hoffmann-La Roche AG. PS-02-013 Automated PD-L1 tumour proportion scoring algorithm in non- small cell lung cancer for multiple companion diagnostic assays D. Rodrigues*, C. Neppl, D. Dorward, T. Losmanová, R. Wyatt, D. Mulkern, S. Pattle, R. Oberson, S. Reinhard, T. Waldburger, I. Zlobec, P. Caie *Indica Labs, USA Background & objectives: High interobserver disagreement when reporting programmed cell death ligand 1 (PD-L1) expression, may result in suboptimal treatment decisions. HALO PD-L1 AI aims to support pathologist quantification for PD-L1 companion diagnostics SP263 and 22c3 assays in non-small cell lung cancer (NSCLC). Methods: HALO PD-L1 AI was trained with 146984 expert annota- tions to identify PD-L1 tumour-positive cells, within segmented tumour regions. The algorithm’s Tumour Proportion Score (TPS) was validated on 203 SP263-stained whole slide images (WSI), assessing its agreement with the TPS scores of three pathologists. For the 22c3 clone, we gathered agreement metrics from the algorithm and the reported TPS score. Results: For SP263-stained images (n=203), pairwise pathologist agreement ranged from 74.9% to 77.3%. Algorithm agreement with the pathologists’ mode was 75.4%, with agreement at the clinically relevant cut-offs (<1%, 1-49% and >50%) ranging from 0.71 to 0.78. Intraclass correlation coefficient (ICC) between the algorithm and pathologists’ TPS scores was 0.95 (95% CI 0.93 – 0.97). Preliminary results on 22c3-stained slides (n=243), show that the over- all percent agreement with the reported TPS score was 74.1%, with the agreement at the clinically relevant cut-offs ranging from 0.68 to 0.82. The ICC agreement was 0.95 (95% CI 0.94 – 0.96). Conclusion: Immunotherapy has revolutionized advanced NSCLC treatment and several companion diagnostic assays are available to determine eligibility for this therapy. However, reporting of PD-L1 expression suffers from high interobserver disagreement. We developed HALO PD-L1 AI to support pathologists PD-L1 scoring with the aim of saving pathologists time and ensuring consistency in the reported results. The algorithm is highly concordant with the pathologist TPS scores for SP263 and 22c3 companion diagnostic assays. PS-02-014 An automated deep learning artifact detection tool for quality control of whole-slide digital pathology images D. Rodrigues*, E. Burlingame, C. Babcock, V. Ovtcharov, S. Reinhard, T. Waldburger, D. Martin, S. Couto, I. Zlobec, P. Caie *Indica Labs, USA Background & objectives: Mechanical and digital artifacts have a negative impact on digital pathology workflows. Image focusing issues can be a critical bottleneck during slide digitisation. We developed SlideQC to automatically segment tissue artifacts in haematoxylin and eosin (H&E) and immunohistochemistry (IHC)-stained slides. Methods: SlideQC was developed using 2499 artifact annotations across 302 H&E and IHC stained slides, alongside 2048 synthetically gener- ated out-of-focus images. SlideQC performance was evaluated on 432 annotations across external H&E (HistoQC Repo) and IHC (LYON19) test images. SlideQC’s ability to distinguish in-focus from out-of-focus was assessed on 4954 patches from the TCGA@Focus dataset. Results: For the external H&E and IHC test sets, SlideQC showed high precision, recall, and F1-score with average values of 0.94, 0.90, and 0.91, respectively, over pixel-level annotations. Recall per artifact type was 0.84 for air bubbles, 0.91 for debris/dust, 0.84 for folds, 0.98 for pen marker, and 0.97 for out-of-focus regions. For the TCGA@ Focus dataset, the median percent of artifact reported for the 2461 out-of-focus labelled patches was 76.7 [IQR 41.3 – 97.6] and for the 2493 in-focus labelled patches was 3.3 [IQR 0.8 – 8.4]. Conclusion: SlideQC can alleviate the bottleneck of manual qual- ity control in both clinical and research based digital pathology workflows, thereby bringing efficiency gains to both fields. Slide QC achieved high precision, recall, and F1-score in H&E and IHC external test cohorts. Furthermore, SlideQC showed a good ability to distinguish out-of-focus from in-focus patches in the TCGA@Focus dataset. By identifying and reporting the percentage of artifacts on each slide, SlideQC could provide an automated, measurable quality control procedure. PS-02-015 Interpretable artificial intelligence to predict lymph node metas- tasis in early gastric cancer Y. Sung*, Y.J. Lee, S.H. Lee, S. Ahn *Republic of Korea Background & objectives: When early gastric cancer (EGC) shows high-risk features, current guidelines recommend surgery due to the risk of lymph node metastasis (LNM). We aimed to develop machine learning algorithm that can predict LNM status using H&E-stained histopathology images from multiple institution. Methods: Our pipeline consists of two sequential approaches; a feature extractor and a risk classifier. For the feature extractor, segmentation network (DeepLabV3+) was trained on 243 WSIs across 5 datasets to differentiate each histologic subtype. UMAP was employed to visual- ize the quality of encoder. After that, risk classifier was trained with XGBoost using 70 morphologic features inferred from trained feature extractor.