ECP 2022 Abstract Book

Virchows Archiv (2022) 481 (Suppl 1):S1–S364 13 Methods: Our protocol can process tissues up to 150-μm in thick- ness, allowing anti-PD-L1 staining of the entire tissue and pro- ducing high resolution 3D images. After 3D imaging process, the thick sections were recovered for epidermal growth factor receptor (EGFR) mutation analysis. We further developed artificial intel- ligence-assisted models to calculate the tumour proportion score (TPS) of the entire 3D tissue. Results: Artificial intelligence-assisted PD-L1 quantitation of these images revealed a marked variation of PD-L1 expression in 3D. In 5 of 33 needle-biopsy-sized specimens (15.2%), the TPS varied by greater than 10% at different depth levels. In 14 cases (42.4%), the TPS at different depth levels fell into different cat- egories (<1%, 1–49%, or ≥50%), which can potentially influence treatment decisions. The EGFR mutation analysis performed using pre- and post-processing tissue yielded identical results in all the tested cases, including 4 cases with EGFR L858R mutation, 6 cases with EGFR exon 19 deletion, 1 case with EGFR exon 20 insertion, 1 case with EGFR G719X mutation, and 8 cases without detectable EGFR mutation. Conclusion: Our novel method has the potential to increase the accuracy of tumour PD-L1 expression assessment and enable precise deployment of cancer immunotherapy. In addition to application in PD-L1 expression assessment in non-small cell lung cancer, 3D tissue imaging can also potentially apply to the evaluation of biomarkers in other cancer types. Importantly, our technology permits recovery of the processed tissue for subsequent mutation analysis, enabling holistic evaluation of the protein-level expression and genetic alterations in small specimens. Funding: Industry-academia collaboration grant from JelloX Biotech Inc. (grant number R-19005) OFP-11-008 CoNIC Challenge: large scale assessment of automated methods for identification and counting of Colon Nuclei S. Graham*, Q.D. Vu, M. Jahanifar, D. Snead, S. Raza, F. Minhas, N. Rajpoot *University of Warwick, United Kingdom Background & objectives: Identification of nuclei in histology images, such as those from epithelial and inflammatory cells, enables large-scale profiling of the tumour microenvironment. To help drive forward innovation for automatic nuclear recognition, we organised the Colon Nuclei Identification and Counting (CoNIC) Challenge. Methods: We created the largest dataset for nuclear recognition in computational pathology, containing around 550K labelled nuclei, and invited researchers to develop algorithms on the data, aimed at solving 2 tasks: 1) nuclear segmentation & classification and 2) prediction of cellular composition. Participants submitted model code to the challenge, which enabled the test data to remain com- pletely unseen. Results: In total, we had 323 and 50 submissions to the prelimi- nary and full test phases, respectively. Submissions were ranked by multi-class panoptic quality (mPQ+) and multi-class coefficient of determination (R2) for tasks 1 and 2, respectively. The top segmen- tation and classification submission achieved an mPQ+ of 0.501 and the top cellular composition prediction submission achieved an R2 of 0.764. Best models were able to successfully identify under- represented classes, such as neutrophils, eosinophils and plasma cells, helping them to achieve competitive mPQ+ and R2 scores. Conclusion: Nuclear morphology and co-localisation of different subtypes have shown to be important indicators of cancer prognosis and diagnosis. However, manual assessment is not feasible as each tissue sample typically contains thousands of nuclei. We organised the CoNIC Challenge to encourage the development of automatic approaches for nuclear recognition in computational pathology. We hope that the widespread participation will motivate research- ers to further develop methods on the provided data and acceler- ate the development of downstream cell-based models for clinical applications. Funding: Simon Graham, Mostafa Jahanifar, David Snead, Shan Raza, Fayyaz Minhas and Nasir Rajpoot are part of the PathLAKE digital pathology consortium, which is funded from the Data to Early Diagnosis and Precision Medicine strand of the governments Industrial Strategy Challenge Fund, managed and delivered by UK Research and Innovation (UKRI). OFP-11-009 Quantification of metabolic heterogeneity across multiple imag- ing modalities E. Smeets*, M. Trajkovic-Arsic, D. Geijs, M. van Zanten, L. Bros- ens, B. Feuerecker, M. Gotthardt, J. Siveke, R. Braren, F. Ciompi, E. Aarntzen *Department of Medical Imaging, Radboud UMC, The Netherlands Background & objectives: Radiomics based on clinical imaging are increasingly applied to classify tumours, but often lack sufficient biological rationale, hampering clinical implementation. We developed an [18F]FDG-PET/CT signature in pancreatic cancer based on MCT4-expression, quantified by texture analyses of whole tumour cross-sections. Methods: We developed a cross-modal image analyses pipeline using a cohort of pancreatic cancer patients (n=29) of which tumour cross-sections and PET-scans were available. We computed density maps of MCT4-expression on whole-slide images to extract texture features. Using k-means, we defined two subgroups with distinct MCT4-expression patterns. From corresponding [18F] FDG-PET scans, texture features were selected that associate with the pre-defined subgroups. Results: Clustering techniques, based on k-means, umap and heatmap analyses, revealed two distinct MCT4-expression patterns. MCT4-expression pattern A was dominated by a higher MCT4- expression level and more local variation. MCT4-expression pattern B was characterized by less MCT4-expression. Using MCT4-expression patterns as label, we investigated which [18F] FDG-PET derived texture features associate with these tumour characteristics. MCT4-expression pattern A was linked to a specific [18F]FDG-PET signature, characterized by higher tracer uptake values and second order features correlated to local variation of tracer uptake on the corresponding clinical scans. The MCT4-based [18F]FDG-PET signatures were applied to an additional cohort (n=71) pancreatic cancer patients who received palliative systemic treatment and showed prognostic value. Conclusion: The presented cross-modal image analyses pipeline allows to build PET-scan signatures based on a biological ration- ale using quantitative immunohistochemistry. As use-case we focussed on tumour glycolysis as hallmark of cancer, measured by MCT4-expression patterns on resected whole tumour sections and by [18F]FDG-PET scans. We show that a subgroup of pancreatic cancer patients with high and heterogenous MCT4-expression can accurately be identified in vivo using [18F]FDG PET-derived tex- ture features. This MCT4-based [18F]FDG-PET-signature associ- ated with worse prognosis. OFP-11-010 Artificial intelligence as a potential tool for pathologists to evaluate lymphocyte infiltration in melanoma S46