ECP 2022 Abstract Book

Virchows Archiv (2022) 481 (Suppl 1):S1–S364 13 Background & objectives: Generalization is one of the main challenges of computational pathology. Slide preparation heterogeneity leads to poor models’ performance on unseen data. This issue is addressed here by comparing approaches based on different cycleGAN usages. Methods: We build a stain translation device using unsupervised cycleGANs image-to-image translation. Three approaches were compared to a baseline model, their performance were assessed using invasive carcinoma patch classification. The first two approaches use the translation device at inference or training respectively, leading to stain specific models. The last method uses it for stain augmentation to produce a stain invariant model. Results: Baseline metrics are set by training and testing a model on a reference stain with colour augmentation. The first approach showed improved performance on different stains by 20% AUC compared to the baseline without requiring task specific model re-training. Secondly, we demonstrate that using the stain trans- lation device before model training allows for labelling knowl- edge transfer between stains. This results in high performance without using any labels for a particular target stain. Finally, the translation device is used for stain augmentation during training, resulting in a stain invariant model with equally good perfor- mance on every stain. Conclusion: Every modality tested in this study improves the baseline without needing labelled data on target stains. We assessed the performances using three medical centres with H&E and H&E&S stainings. The study shows that cycleGAN based domain adaptation methods are solutions to the generalization challenge in computational histopathology. Such a framework can be used in other applications such as auto- matic segmentation or object detection. E-PS-14-008 Detection of microcalcifications in whole slide images: a comparison between image processing and deep learning approaches M. Clavel*, N. Pozin, S. Sockeel, M. Sockeel, C. Miquel *Primaa, France Background & objectives: Microcalcifications are calcium deposits in breast ducts. If the mammographic abnormality reveals microcalcifications, the pathologist should make every effort to identify them. As a help, we present an automatic microcalcification detection pipeline in Whole Slide Images (WSI). Methods: Epithelial regions are parsed from the WSI into patches that are fed to a classifier. A first proposed image processing-based classifier (Classifier1) detects blurry dark objects as this is the typical aspect of microcalcifications on WSI. The second one (Classifier2) is a convolutional network trained on 164835 patches from a total of 1615 WSI, including 1633 microcalcifications patches. Results: Classifiers are evaluated based on their balanced accu- racy (BAcc), precision (Pr), recall (Re). However, Those metrics are insufficient to capture the value brought to the pathologist. Not every microcalcification must be found, the mere detection of their presence is enough to determine whether the biopsy was well located. To evaluate the pipeline performance, we propose additional metrics. Objects classified as microcalcifi- cations are sorted by classifier’s confidence and we compute: the average rank of the first detected microcalcification (Aver- age_Rank), the average number of slides for which a microcalci- fication is detected among the top 16 objects (Microcal_in_top). Results are: BAcc: 0.79-0.88, Pr: 0.07-0.73, Re: 0.67-0.76, Average_Rank: 8.5-0.7, Microcal_in_top: 0.75-0.83 (format metric_name: metric_classifier1-metric_classifier2) Conclusion: Our automatic microcalcification detection pipeline could be helpful to pathologists. Two different classifiers can be plugged into the pipeline. The first one is based on image pro- cessing techniques and needs a small amount of labelled data to be set up. Although its performance is good, it is outperformed by the deep learning based classifier. Both solutions can be used depending on the availability of labelled data. E-PS-14-009 Understanding batch variation within a cohort before digi- tal pathology analysis of multiplex immunofluorescence in colorectal cancer I. Carretero Del Barrio*, A. Viratham Pulsawatdi, K. McCombe, V. Bingham, J. James, M. Salto-Téllez, S. Craig *Patrick G Johnston Centre for Cancer Research, Queen’s Uni- versity Belfast, UK, Spain Background & objectives: To apply two optimised multiplex-immu- nofluorescence (mIF) protocols (Panel 1-DAPI, Cytokeratin, CD4, CD8, CD3, CD20; Panel 2-DAPI, Cytokeratin, CD4, CD68, FOXP3) to a cohort of colorectal-cancer (CRC) in order to determine effect of batch variation within the cohort using digital pathology. Methods: mIF protocols were applied to 518 FFPE-CRC sections using an Leica Bond RX autostainer. Each batch of slides was ran with a tonsil control and fluorescent scanning was performed using a Vectra Polaris. Using QuPath, we performed annotations, cell-segmentation, epi- thelium-stromal classification and cell classification. Cell features and summary statistics for each mIF channel were exported for quantitative analysis using RStudio. Results: Mean greys per simulated cell for each biomarker of inter- est were reviewed in order to determine batch variability. Values for Cytokeratin, DAPI, CD4, CD8 and FOXP3 demonstrated limited variability across the cohort for both panels. In contrast, values for CD20, CD3 in panel 1 and CD68 in panel 2 were found to be batch dependent, affecting 7/15 and 5/15 of panel-specific staining runs. On review, the same florescent dye (Opal570) was used to visual- ise CD20 and CD68 across both panel designs and therefore most likely to influence batch-dependent, dye-specific background stain- ing. Whilst tumour-stroma classification was adequate in affected batches, biomarker-dependent cell classification was positively skewed compared to adjacent tonsil controls. Conclusion: This study found that batch artefacts did not sig- nificantly impair tissue-specific, mIF epithelium-stromal classi- fication, due to use of an artificial neural network. In contrast, biomarker-specific cell classification using pre-defined intensity thresholds was found to be vulnerable to positive skew in tissue sections from batches where high Opal570 background staining was present. Whilst use of pre-defined thresholds was acceptable for most optimised antibody-opal pairs in both panels, this work indicates the importance of reviewing dye-specific bias prior to mIF image analysis. Funding: This study was supported by Cancer Research UK. Grant Numbers: C11512, /A20256 E-PS-14-010 Abnormal differentiation of follicular helper CD4 T (TFH) cells in systemic lupus erythematosus; an imaging perspective K. Ioannidou, D. Comte, A. Clottu, L. de Leval, C. Petrovas* *CHUV, Switzerland S300