ECP 2022 Abstract Book

Virchows Archiv (2022) 481 (Suppl 1):S1–S364 13 Conclusion: The diversity in opinions illustrates variety in barriers and facilitators in computational pathology implementation. A next step would be to quantitatively determine important influencing factors among all relevant stakeholders. Simultaneously, prospec- tive validation studies may be developed and initiated, to collect evidence on the most effective way of implementation. This will further propel the use of computational pathology into clinical practice. Funding: This study received funding from the Dutch Cancer Soci- ety (grant number 2017-10602). This project has received funding from the Innovative Medicines Initiative 2 Joint Undertaking under grant agreement No 945358. This Joint Undertaking receives sup- port from the European Union’s Horizon 2020 research and innova- tion program and EFPIA. CP-02-004 The prognostic value of deep learning based mitotic count for breast cancer molecular subtypes M. Balkenhol*, C. Mercan, L. Tessier, D. Tellez, A. Niendorf, A. Wegscheider, P. Bult, F. Ciompi, J. van der Laak *Radboudumc, The Netherlands Background & objectives: Breast cancer grading was introduced decades ago and its prognostic value has not yet been studied in the context of contemporary molecular classification. This study uses automatic mitotic count to evaluate its prognostic value for different breast cancer molecular subtypes. Methods: A previously developed artificial intelligence (AI) algo- rithm detected mitoses and assessed mitotic counts in H&E-stained whole-slide images from a multicentre cohort of 846 breast cancer patients. Stratified analyses based on hormonal receptor (HR) and HER2 status were performed to study potential different prognostic mitotic cut off values. Multivariable Cox regression survival mod- els were used to study its independent prognostic value. Results: We found that the mitotic count, assessed by AI was prog- nostic in univariate Cox analysis for HR positive / HER2 negative breast cancers, applying the widely used Nottingham cut-offs, both for recurrence free and overall survival. Prognostic value could be optimized applying a cut-off of 10 mitoses per 2 mm2 (recur- rence free survival hazard ratio = 2.05 (1.14-3.68; p=0.02); overall survival hazard ratio = 1.84 (1.09-3.11; p=0.02) in multivariable analysis). However, for HER2-positive tumours, no mitotic cut off was found to be prognostic. Conclusion: This study shows that automatic mitotic count yields different prognostic information for specific subtypes of breast can- cer, suggesting the need for a molecular subtype specific grading assessment in clinical practice. In addition, it showed the potential of AI to automate part of the pathologists’ workflow, as well as the feasibility of applying modern AI technologies to re-assess widely used histopathological features by evaluating large numbers of cases in a systematic, accurate and reproducible manner. CP-02-005 Pathologist validation of a machine learned biomarker for risk stratification in colon cancer V. L’Imperio, E. Wulczyn, M. Plass, H. Muller, N. Tamini, L. Gianotti, N. Zucchini, R. Reihs, G.S. Corrado, L.H. Peng, P. Cam- eron Chen, M. Lavitrano, Y. Liu, D.F. Steiner, K. Zatloukal, F. Pagni* *Department of Medicine and Surgery, Pathology, ASST Monza, San Gerardo Hospital, University of Milan-Bicocca, Monza, Italy Background & objectives: Identifying new prognostic features in colon cancer may refine histopathology review. While prog- nostic artificial intelligence (AI) systems have demonstrated sig- nificant risk stratification in several cancer types, studies have not yet shown that the machine learned features are interpretable by pathologists. Methods: This retrospective study utilized de-identified, archived colorectal cancer cases from 2013 to 2015 from University of Milano-Bicocca (UNIMIB). Histologic slides from 258 consecu- tive colon adenocarcinoma cases were reviewed at UNIMIB by two institutional pathologists. The pathologists conducted semiquanti- tative scoring for Tumor Adipose Feature (TAF), which was previ- ously identified via a prognostic deep-learning model developed using an independent colorectal cancer cohort. Results: 258 colon adenocarcinoma histopathology cases from 258 patients (median age 67 years; interquartile range 65-81; 47% female) with stage II (n=122) or stage III (n=139) cancer were included. TAF was identified in 120 cases (widespread n=63; multifocal n=31; unifocal n=26). For OS analysis adjusting for tumour stage, TAF was independently prognostic: Hazard Ratio (HR)=1.55 (95%CI 1.07- 2.25; p=0.02) for TAF as a binary feature (presence vs. absence); and HR=1.87 (95%CI 1.23-2.85; p<0.005) for the highest TAF category (widespread) when evaluating semiquantitative scoring. Inter-pathologist agreement for widespread TAF vs. lower categories (absent/unifocal/multifocal) was 90%, corresponding to kappa at this threshold of 0.69 (95%CI 0.58-0.80). Conclusion: Pathologists were able to learn and reproducibly score for TAF providing significant risk stratification on this independent dataset. While additional work is warranted to understand the biological significance of this feature and to establish broadly reproducible TAF scoring, this work represents an important milestone as the first validation of human expert learning from machine learning in pathology. This validation demonstrates that a computationally identified histologic feature can represent a human-identifiable, prognostic biomarker with the potential for integration into pathology practice. CP-02-006 Predicting genetic variation from quantitative tissue pheno- types using explainable machine learning J. Connelly*, J. Luft, C.J. Anderson, P. Bankhead, F. Connor, P. Flicek, N. López-Bigas, C.A. Semple, D.T. Odom, S.J. Aitken, M.S. Taylor *MRC Human Genetics Unit, Institute of Genetics and Cancer, University of Edinburgh, United Kingdom Background & objectives: Most human cancer genomes exhibit multiple mutational signatures, reflecting the complex milieu of damage and repair occurring during carcinogenesis. We used a robustly controlled, highly powered in vivo experiment to investi- gate genotype-phenotype correlates. Methods: Inbred mice were exposed to a single dose of diethylni- trosamine shortly after birth. Resultant liver tumours were isolated and submitted for WGS, total RNAseq, and histopathology. This cohort was used to discover lesion segregation, which drives can- cer genome evolution. We used deep learning to segment nuclei in these images, computed quantitative morphometric features, and modelled these using machine learning. Results: We find that supervised learning of quantitative nuclear morphology robustly predicts (i) germline variation between ances- trally divergent mouse strains, (ii) germline heterozygosity within strain, and (iii) somatic mutations in driver oncogenes. We apply a game-theoretic approach to uncover morphometric features which explain the inference, identifying nuclear geometry as key to infer- ring driver gene mutations, and nuclear histochemical staining most S57