ECP 2022 Abstract Book

Virchows Archiv (2022) 481 (Suppl 1):S1–S364 13 "Application of artificial intelligence to provide automation and standardiza- tion of the Gleason system in the diagnosis of prostate cancer" E-PS-14-017 An automated deep learning based mitotic cell detection and recognition in whole slide invasive breast cancer tissue images S. Cayir, G. Solmaz, H. Kusetogullari, F. Tokat, E. Bozaba, S. Karakaya, L. Iheme, E. Tekin, C. Yazici, G. Ozsoy, S. Ayalti, K. Kayhan, U. Ince, B. Uzel, O. Kilic, B. Darbaz* *Virasoft Corporation, Turkey Background & objectives: Nottingham Histologic Grading (NHG) is a prognostic indicator in early invasive breast cancer. NHG contains three factors which are pleomorphism, tubular formation and mitosis count. Mitosis recognition plays an important role for NHG estimation and accurate assessment of cancer prognosis. Methods: Two novel datasets, which include 139.124 nucleis and 9.816 mitoses with annotations, are created and presented. Moreover, a hybrid deep learning framework is proposed for mitosis recognition. To achieve the results, a modified scaled-YOLOv4 algorithm is first used to detect all nuclei in WSIs. Then, a modified-VGG11 model is employed to recognize and localize mitoses in the WSIs. Results: In comparison to various classification methods, the proposed framework provides the best results for both in-house and other datasets including MIDOG-21 and ATYPIA. In the first experiment, all algorithms are trained and tested on in-house dataset, the proposed framework obtains the best accuracy results than the other classification algorithms with F1-Score of 49. Moreover, the second experiment shows that the created novel dataset contains different characteristics and features than the MIDOG-21 and ATYPIA datasets. Conclusion: In this work, we propose a hybrid deep learning approach and introduce two new datasets. The proposed framework is compared with different algorithms on various datasets. The results prove that the proposed approach performs better than the other algorithms in both nuclei detection and mitotic cell recognition in WSIs. Further clinical validation studies are needed for clinical implementation of AI based mitotic count. E-PS-14-018 Comparison of different nuclear segmentation algorithms on the digital image of HER2 FISH labelled breast cancer tissue sample A. Csizmadia*, V. Jonas, R. Paulik, T. Krenacs, B. Molnar *Diagnostics Department, 3DHISTECH Ltd., Hungary Background & objectives: The FISH technique is a frequently used molecular method to visualize well known genetic aberrations. The method provides results of qualitative, quantitative measurements that can be used to identify the genetically affected cells and their proportion in the tumour population. Methods: The HER2 protein expression status in breast cancer provides a useful working example of tumour heterogeneity. The FISH technique can visualize the HER2 genetic heterogeneity in the breast tissue samples in situ. The fluorescence samples’ peculiarity of the signal burning out, therefore the fluorescence digitalization and the demand an objective FISH evaluation results have become a basic requirement in diagnostics. Results: The WSI imaging provides an opportunity to identify the different HER2 amplified nucleus. The base of the FISH image analysis is the nucleus segmentation. The classical 2D image analysis algorithms have a significant limitation regarding the detection of the overlapped nuclei. The intensity differences between the overlapped nuclei are not significant for define a proper cut line. The inadequate segmentation results incorrect signal assignment. The fluorescence Z layer scanning method opened a new opportunity for the digitalization and analysis of FISH samples. Slicing the sample section into different Z planes give an opportunity 3D image segmentation. AI based StarDist image segmentation algorithm was evaluated for achieve more precise nuclei segmentation. Conclusion: During our comparison image segmentation test, we found that the breast tissue intratumor HER2 heterogenetic results change considerably using different image processing algorithms. The deep-learning based segmentation algorithm is found to be more robust in versatile image environment resulting in more accurate object segmentation and separation. The novel approach of nuclei segmentation in turn improves the signal assignment process changing the number of signals within each nucleus, thus improving the reli- ability of the quotient values confirming the amplification. E-PS-14-019 Development and assessment of single-cell image classification systems for haematological cytomorphology using convolutional neural networks C. Matek* *Institute of Pathology, University Hospital Erlangen, Germany Background & objectives: Morphologic classification of leukocytes represents an important step in the diagnostic workup of blood and bone marrow samples for many haematological diseases. Frequency and importance of cytomorphologic evaluation motivate development and evaluation of diagnostic support systems for leukocyte classification. Methods: Publically available datasets of peripheral blood and bone marrow cytomorphology are described and used in order to train and evaluate sequential models and ResNeXt-based neural networks for leukocyte classification. Performance of the trained networks is evaluated using data from different sources, thus assessing generalizability of prediction performance. Finally, multiple strategies are presented for robustness assessment and explainability of algorithm predictions. Results: Convolutional Neural Networks developed using publicly available datasets of leukocyte cytomorphology in blood and bone marrow cytomorphology attain the performance of trained cytologists for tasks such as blast recognition in the diagnosis of AML. The overall classification performance critically depends on the training sample number per class. The pattern of deviation from ground truth is similar for algorithms and humans, with consecutive morphologic cell types within a continuous hematopoietic lineage most susceptible to confusion. Explainability methods show that the models map relevant image areas. Analysis of neural network predictions illustrates the importance of augmentation strategies to ensure classifier robustness against staining and scanner setting variabilities and avoid overfitting. Conclusion: Published expert-annotated datasets of leukocyte morphology allow training of state-of-the-art neural networks for single-leukocyte classification in the diagnostic workup of hae- matological diseases. For many diagnostically relevant classes, these networks attain a performance level comparable to human examiners. Using explainability methods, image areas relevant for classification can be illustrated, suggesting significant overlap with structures known to be relevant for cytomorphologic classification. Appropriate augmentation strategies allow hardening classifiers against different preanalytical and digitization parameters, thus ensuring robustness and generalizability of network predictions. S303