ECP 2023 Abstracts
S232 Virchows Archiv (2023) 483 (Suppl 1):S1–S391 13 Despite that cell line controls are within two SDs of the mean, the NSCLC-TMA (n=40) illustrated a significant (p < 0.05) decrease in positive cases at 1:80 (20/25), 1:100 (17/25) and 1:150 (12/25) com- pared to standard 22C3 LDT. Furthermore, analysis of the dynamic range cell line revealed early detection of a faulty immunostainer. Conclusion: IHC quality control of PD-L1 using a dynamic range cell line combined with Qualitopix analysis showed variability in stain intensity and allowed early detection of technical issues. To reduce intra-laboratory variability and ensure reliable consistent immunohis- tochemical assays for PD-L1, a dynamic range cell line proved to be a better control compared to the conventional tonsil. In conclusion, a dynamic range cell line combined with Qualitopix showed an improve- ment in identifying optimal staining for cases near or at its analytical cut-off. E-PS-08-002 Improving breast cancer diagnosis through weakly supervised learning: promising results in malignancy detection and subtyping A. Alexander*, C. Mayer, N. Balint Lahat, I. Barshack, S. Ben Amitay *Sheba Medical Center, Israel Background & objectives: The application of AI in digital pathology for breast cancer diagnosis often requires manual and time-consuming annotation. To overcome this limitation, we propose a self-supervised learning and attention-based approach with case-level annotation to detect malignancy and enable subtype classification. Methods: We extracted tiles from 1,120 benign and malignant breast lesion slides and used them to train a self-supervised learning feature extractor with MoCo. The extracted features were fed into an attention- based deep multiple instance learning (MIL) model to detect malignancy in a weakly supervised manner. High-attention tiles selected by the MIL model were used for training subtype classification with RESNET-18. Results: Our malignancy detection model was trained and tested on 1186 and 297 breast slides, respectively, with a balanced malignant/ benign ratio. Our method achieved balanced accuracy of 0.91, AUC of 0.96 and average precision of 0.98 for the training and a balanced accuracy of 0.91, AUC of 0.96 and average precision of 0.99 for the testing. High-attention maps of 30 malignant slides were confirmed to contain tumours by a pathologist. For subtype classification, we employed a dataset of 561 slides (468 invasive ductal carcinoma, 93 invasive lobular carcinoma). We trained the classifier using 5-fold cross-validation and achieved accuracy of 0.88 and AUC of 0.91. All slides were stained with H&E. Conclusion: Our proposed approach has demonstrated promising per- formance in accurate malignancy detection and subtyping of breast cancer. The utilization of a case-level annotation approach helps to mitigate the annotation burden, facilitating scalability and applicability to larger datasets. The results of the subtype classification suggest that our framework can extract meaningful tiles that represent the tumour. We plan to expand our dataset size to enhance the robustness of our subtyping approach and extend our framework to predict biomarkers and hormonal status. E-PS-08-003 European laboratories capabilities for digital pathology and com- puter assisted algorithms N. Atkey*, P.D. Ramos Cirillo, O.D. McLeod, D. Vance, A. Capece, B. Bisaro, J. Tsiampali, R. Assemat, P. Garcia, Y. Pattni *Diaceutics PLC, United Kingdom Background & objectives: Digital pathology (DP) is a tool for labo- ratories to enhance accuracy and speed of diagnosis. Adoption of DP is subject to laboratory capabilities, awareness, and regulations. We reviewed the readiness of European laboratories for DP and computer assisted algorithms (CAA). Methods: The Diaceutics DXRX Diagnostic Network® contains real- world data from clinical laboratories worldwide. Technology capabilities from 124 European (including European Union member states France, Germany, Italy, and Spain) and UK pathology labs performing solid tumour testing were analysed, from January 2019 – August 2021, focus- ing on utilization of DP, whole slide image scanners (WSI) and CAA. Results: In all five markets, DP is utilized for clinical testing in aca- demic (33%), hospital (20%), and commercial labs (4%), while 18% of all labs use DP for research purposes and 25% do not use DP. For clinical use, DP is utilized for breast and lung cancer indications by the majority of labs (67%), and 33% of labs use DP for all solid tumour testing. Across labs with WSI platforms, Leica, Philips and Hamamatsu are the preferred choices (32%, 30%, and 20% respectively). Only 6% of all labs surveyed for CAA use it for clinical testing, 51% for research, 19% under validation, and 24% are not using CAA. Conclusion: DP and CAA are emerging solutions to aid pathologists on precision oncology testing. Over 50% of the labs use DP for clini- cal diagnosis, whilst the other half do not use it or use it for research purposes. CAA can enhance diagnostic accuracy of DP but it is not yet embraced in the wider clinical setting. The number of labs using DP with integrated CAA in clinical routine remains low, with most of labs using CAA only in a research setting. E-PS-08-004 Implementation of artificial intelligence assisted lymph node metas- tases detection for breast cancer M. Balkenhol*, G. Litjens, P. Bándi, K. Grünberg, J. van der Laak *Radboudumc, The Netherlands Background & objectives: Breast cancer diagnostics can be supported by artificial intelligence (AI) based detection of lymph node metastases. This project aims to integrate AI into routine pathology practice, to comply with the recent EU ‘in vitro diagnostic regulation’ (IVDR). Methods: In-house developed AI for detection of metastases in whole slide images (WSI) of H&E sections of sentinel lymph nodes (SN) was integrated in the clinical image management system. According to IVDR, performance evaluation was executed, including methods on scientific validity, analytical and clinical performance. A quality man- agement system with risk analysis was developed and key performance indicators (KPI) were defined. Results: AI was integrated using two operating thresholds for metas- tases detection: first processing all slides with high specificity, antici- pating that the majority of SNs are negative; and second, tuning the detections of the positive slides with higher sensitivity to ensure more accurate detections. Analytical performance was tested by cutting and staining tissue blocks on various moments and scanning them multiple times, after which WSI were processed and results on detections were compared. For clinical performance a real world set-up with a fully crossed, intermodal, multireader design was used. Results on perfor- mance studies and KPIs are currently evaluated. Conclusion: AI in pathology is rapidly developing and regulation on how to implement in clinical practice is not yet established. The recently introduced IVDR provides a framework for implementation of lab developed tests but does not specify in detail the needs for imple- mentation of AI in clinical practice. Our project describes procedures for implementing AI in routine pathology practice under IVDR, which may serve as a blueprint for labs wanting to use in-house developed AI while complying with IVDR. E-PS-08-005 Application of artificial intelligence in establishing uveal melanoma prognosis A. Bdioui*, F. Hamdaoui, A. Sghaier, O. Belkacem, N. Missaoui, S. Hmissa *Sahloul hôpital of sousse, Tunisia
Made with FlippingBook
RkJQdWJsaXNoZXIy Mzg2Mjgy