ECP 2022 Abstract Book

Virchows Archiv (2022) 481 (Suppl 1):S1–S364 13 PS-03-009 AI-aided assessment of HER2 status in primary and meta- static breast carcinoma C. Palm*, C. Connolly, R. Masser, B. Padberg Sgier, E. Diaman- tis Karamitopoulou, B. Bode, M. Tinguely Kovarik *Pathologie Institut Enge, Switzerland Background & objectives: Re-evaluation of HER2 receptor status is recommended following metastatic transformation of invasive breast carcinoma to assess for receptor status conver- sion. We examine the performance of artificial intelligence in the determination of HER2 status in metastatic compared to primary breast carcinoma. Methods: 60 slides of matched primary and metastatic invasive breast carcinoma were selected from our digitalised cohort of 476 cases (Roche Ventana DP200 Scanner). HER2 IHC scoring was performed by 3 pathologists, and followed with ISH when IHC score ≥ +1. Results were compared to those from uPath HER2 4B5 and Dual ISH algorithms for IHC and ISH respectively. Results: Overall, there was moderate agreement between pathol- ogist and AI results (Cohen’s κ 0.43, 95CI 0.25-0.61), with higher concordance on metastatic (κ 0.48, 95CI 0.24-0.73) ver- sus primary cases (κ 0.36, 95CI 0.07-0.64). Breakdown analyses revealed the lowest concordance with ISH on metastatic lesions (κ 0.23, 95CI 0.04-0.49) where unedited AI overestimated the number of positive and equivocal cases. Inter-observer-varia- bility of HER2 IHC scoring among pathologists was similar for primary (Fleiss’ κ 0.77, 95CI 0.68-0.86) and metastatic lesions (κ 0.73, 95CI 0.58-0.89). Conversion of HER2 status between primary and metastatic lesions was observed in one case, con- firmed by pathologists and AI on IHC and ISH. Conclusion: Moderate concordance was observed between AI and pathologists in the assessment of HER2 status on primary and metastatic breast carcinoma. Concordance between AI and pathologists was notably higher with IHC compared to ISH. Examination of larger cohort with a diverse range of metastatic sites, combined with increased operator input at the time of digi- tal analysis, may help to determine the feasibility of an auto- mated digitalised HER2 workup for metastatic breast carcinoma. PS-03-010 Evaluation of Ki67 by image-analysis-enhanced quantitative digital pathology M.J. Krämer*, B.V. Sinn, P. Jank, A. Grass, A. Litmeyer, M. Untch, D. Gerber, A. Schneeweiss, K. Saeger, M. Gleitsmann, J. Furlanetto, S. von Gerlach, B. Felder, A. Ramaswamy, S. Loibl, C. Denkert, W.D. Schmitt *Institute of Pathology, Philipps-University Marburg, Germany Background & objectives: Assessment of prognosis of breast cancer by Ki67 immunohistochemistry is an important element of personalized treatment strategies. A precise assessment is crucial for clinically relevant therapy decisions. We aimed to evaluate a Ki67 quantification tool using whole slide images (WSI). Methods: 61 Ki67 stained, pre- (preTx) and intra-therapeutic core biopsies, as well as corresponding surgical residual disease tissue from neoadjuvant GBG trials, were digitalized. Manual Ki67 scoring was performed by five individual pathologists on WSI (multi-observer), while semi-automated Ki67 scoring was done using VMscope’s Scan Connect (multi-tumour-area). Scan Connect analysed up to four 600x600dpi areas on WSI, followed by pathologist supervision. Results: The Pearson correlation between manual (man.) and computational (aut.) Ki67 assessment was r=0.781, with no significant differences in overall scoring or global precision (p=0.333, p=0.070). Semi-automated multi-area analysis on preTx tissues had a significantly lower variability than manual assessment (n=29, p<0.001), while showing no differences on intra-therapeutic samples (n=29, p=0.885). Using predefined cut offs, we observed that in Ki67 low and intermediate (int.) groups standard deviations (sd) were lower, while being higher for Ki67 high group, irrespectively of the assessment method (man. & aut.: low sd=±3.4 & ±2.0; int. sd=±4.3 & ±3.9; high sd=±18.2 & ±12.7). In comparison with full-automated assess- ment, the supervised semi-automated assessment improved the precision significantly (p=0.008). Conclusion: We found strong correlation between manual (multi- observer) and supervised semi-automated (multi-tumour-area) Ki67 assessment. In comparison to inter-observer variance in manual Ki67 assessment, improved precision was seen in semi-automated assessment considering intra-tissue variance. VMscope’s Scan- Connect software could be a useful tool in pathological cancer diagnostic. As a next step, results should be validated in a larger cohort and survival data could be included to examine prognostic differences between the two methods. Funding: This project is partly funded by joined BMBF/DLR pro- ject “CognoScan” (13GW0207). PS-03-011 Leveraging deep learning-based mitosis detection models for supporting automated breast cancer grading K. Korski*, K. Badowski, X. Li, Y. Nie, S. Abbasi-Sureshjani *F. Hoffmann-La Roche AG, Switzerland Background & objectives: The mitotic figure count within 10 HPF in a mitosis-dense region is the only mitosis-based metric used for tumour grading in breast cancer. We present an innovative spatial statistical analysis of mitoses driven by an automatic mitotic figure detection method. Methods: Mitotic figure candidates are detected using a detector neural network and are then filtered by a classifier neural network with customised architecture. The density and spatial distribution of detected mitotic figures and their relationship to tumour grade are investigated, using G-function that characterises probability distribution of nearest neighbour distances and Getis-Ord Gi sta- tistics that detects local hotspots. Results: A mitotic figure classification model with 83.06% validation accuracy, was applied on 131 test slides includ- ing 18, 47 and 66 slides with tumour grade I, II and III, respectively. Mitotic density showed 0.3987 Pearson (0.4660 Spearman) correlation with tumour grade. The area between observed and theoretic G-function suggested that nearest neighbours of mitotic figures were closer when tumour grade is higher. Group comparison p-values for tumour grade I vs II, II vs III, were 0.0069, 0.019 respectively. The hotspot ratio suggested more hotspots present in higher tumour grade. Group comparison p-values for tumour grade I vs II, II vs III, were 0.015, 0.0036 respectively. Conclusion: Automatic mitosis detection models enable identify- ing regions of highest mitotic activity and quantification of the tumour proliferation and aggressiveness based both on the pres- ence of mitotic figures and their relation to their neighbourhood. Our spatial analysis identifies metrics that quantify the mitotic figure spatial distribution and have statistical power to differenti- ate between low, intermediate, and higher grade tumours. Future work will focus on assessing whether these spatial metrics provide additional information over the current standard and ideally better prognostic value. S77