ECP 2022 Abstract Book

Virchows Archiv (2022) 481 (Suppl 1):S1–S364 13 E-PS-14-020 A reinforcement learning approach for automated scoring of immunohistochemically stained mismatch repair genes M. Raza*, R. Awan, R.M.S. Bashir, T. Qaiser, D. Snead, N. Rajpoot *University of Warwick, United Kingdom Background & objectives: Assessment of the Mismatch Repair (MMR) status is critical for determining the treatment of colorectal cancer (CRC) patients. We propose a novel reinforcement learning (RL) approach that mimics histopathologists for automated scoring of immunohistochemically stained MLH1 and PMS2 slides. Methods: The proposed framework consists of two components where initially, a soft attention module analyses WSIs at lower magnification (5×) and identifies potential regions of interest (ROIs) for tile extraction. The tiles are fed into the second com- ponent, an RL-based retina-inspired hard-attention classification model which further extracts and processes a series of high-magni- fication patches from each image tile for target prediction. Results: The proposed method mimics a pathologist analysing only relevant parts of the IHC slides. We compare our approach against random-sampling with RL-classifier and sliding window based tile extraction. We aggregate results using average probabilities of the top 15 tiles from each WSI for final score. The proposed model achieved superior performance for both biomarkers with F1-scores of 0.88, 0.82 and AUROC of 0.92, 0.88 for MLH1 and PMS2, respectively. Competitive perfor- mance was achieved by ResNet-18, achieving F1-scores of 0.72, 0.71 and AUROC of 0.92, 0.93 for MLH1 and PMS2 while worst performance was achieved by random selection method for F1-score 0.66, 0.62 and AUROC 0.67, 0.64 for MLH1 and PMS2. Conclusion: We demonstrate the effectiveness of our approach, which mimics a pathologist analysing a fraction of a given IHC slide, for scoring MMR markers in CRC. Our method demonstrates comparable performance to standard computational-pathology methods while only processing a tiny fraction of the WSI. To the best of our knowledge, this is the first RL approach to automatically score the MMR markers. We further intend to improve the model by determining the optimal number of image tiles required for each WSI. E-PS-15 | E-Posters Molecular Pathology E-PS-15-001 Intratumoural heterogeneity with different resistant subclones in a treatment resistant gastrointestinal stromal tumour revealed through a novel tissue masking technology X.F. Chen*, T.K.Y. Tay, G.S. Tan, R.C.H. Goh, J. Butler, T.K.H. Lim *Department of Anatomical Pathology Singapore General Hospital, Singapore Background & objectives: A 73-year-old man with a history of metastatic small bowel gastrointestinal stromal tumour (GIST) who have undergone liver and small bowel resections was found to have recurrent peritoneal nodules on follow-up despite being on Imatinib. He subsequently underwent tumour resections. Methods: On histological examination, the nodules composed of sheets of spindle cells and focal epithelioid cells with vesicu- lar nuclei and prominent nucleoli. The tumour was positive for DOG1 and CD117 on immunohistochemistry which confirmed the diagnosis of GIST. It was then sent for next generation sequencing (NGS). NGS revealed KIT exon 17 N822K, exon 17 N822Y and exon 11 N564_P573delinsGSMET mutations. Results: The heterogeneous molecular findings in conjunction with the variable tumour morphology had raised the possibil- ity of intratumoral heterogeneity (ITH). A repeat sequencing was performed to confirm the findings. Quantumcyte Onco- mask technology was used to assist in the sampling of tar- geted areas for NGS. Three separate lysates on three areas of interest on a tumour section were created based on morphol- ogy. The repeated NGS results verified the findings of ITH. Both areas 1 and 2 with spindle cells revealed KIT exon 11 N564_P573delinsGSMET and exon 17 N822K mutations. Area 3 with epithelioid cells showed KIT exon 17 N822Y and exon 17 G803D mutations. Conclusion: This case demonstrated the utilization of Quantumcyte Oncomask, a novel tissue masking technology in detection of heterogeneous resistant subclones in a treatment resistant GIST. It enables the sampling of targeted areas in a tumour section to increase the cellular purity for molecular sequencing especially for the analysis of tumour heterogeneity which is known to be associated with therapeutic resistance. The presence of different mutation subclones in GIST is also important in guiding oncologists on the treatment choices. E-PS-15-002 Cellular and cell-free targetable mutations in non-small cell carcinoma: a comparative study N. Kumari*, R. Paturu, R. Lingaiah, S. Singh, N. Krishnani *All India Institute of Medical Sciences Raebareli, India Background & objectives: Therapeutic relevance of molecular alter- ations in non-small cell lung carcinoma (NSCLC) is long established with detection in tissue being gold-standard. Insufficient tissue neces- sitates use of non-invasive approach as an alternative. Targetable mutations in 100 paired (tissue, plasma) samples were evaluated. Methods: 100 patients with paired (formalin fixed paraffin embed- ded tissue and plasma) samples from treatment naïve NSCLC cases were tested for EGFR (exons 18, 19, 20, 21), ALK, ROS1 and MET through Realtime PCR. Interpretation was based on the difference in cycle threshold (CT) of reference and mutation (CT mutation – CT reference = ΔCT). Results: Tissue samples showed mutations in 60 cases [EGFR mutation - 47, ALK rearrangement -12 (one case had concurrent EGFR and ALK alterations), ROS1 fusion - 2]. Plasma samples showed only EGFR mutation in 43 cases. Overall concordance between tissue and plasma was 62% which dropped to 44% in EGFR mutation-specific sub-type. 38% discordance was recorded for EGFR positive tissue samples while 17 cases of plasma samples showed EGFR mutation where tissue was wild. Two cases showed ALK rearrangement in tissue but EGFR mutation in plasma. Over- all sensitivity and specificity for all molecular alterations in plasma were 46.7% and 62.5% which increased to 55.3% and 67.9% with respect to EGFR mutation. Conclusion: Cell free testing for targetable mutations are com- plementary but not a surrogate marker to tissue biopsies. Realtime PCR is quite sensitive in detecting cell-free mutations, however more sensitive techniques like next generation sequencing and appropriate collection and processing of cell free samples may be be used to increase the detection rate of targetable mutations in NSCLC. S304