ECP 2023 Abstracts
S240 Virchows Archiv (2023) 483 (Suppl 1):S1–S391 13 H&E WSIs. Furthermore, we assessed methods to help mitigate overfit- ting to batch effects. Methods: WSIs from TCGA-LUAD originating from 22 separate TSSs were included in the study. Separate XGBoost models were trained to predict TSS using two sets of image features, one extracted using an Imagenet-pretrained ResNet50, and one using a ResNet50 pretrained on histology data. The effect of Boruta feature selection and ComBat batch correction for mitigating batch effects was evaluated. Results: Image features extracted using the SSL-histology-pretrained Resnet50 (AUROC=0.902±0.005) performed better (p=0.0017) at predicting TSS than using image features extracted using ImageNet- pretrained Resnet50 (AUROC=0.857±0.029). Boruta feature selec- tion and ComBat correction on SSL-histology image features had no impact on the performance of the model at predicting TSS (p=0.133 and p=0.255, respectively). Applied to the image features extracted with the ImageNet-pretrained ResNet50, Boruta feature selection had no impact on the performance of the model at predicting TSS (AUROC=0.839±0.026; p=0.195). However, a decrease in performance (p=6.68e-8) was observed when training the model with ComBat-corrected features extracted using an ImageNet-pretrained Resnet50 (AUROC= 0.536±0.024). Conclusion: Image features extracted using SSL-histology-pretrained Resnet50 may contain more site-specific information than image fea- tures extracted using an ImageNet-pretrained Resnet50. Future work should assess the impact of batch correction methods on the predict- ability of biomarkers directly from H&E WSIs. Outputs of model interpretability methods, such as class-activation maps and attention heatmaps, reviewed by pathologists should remain as gold-standard for ensuring that machine learning models in digital pathology do not exploit confounding batch effects in H&E WSIs. Funding: This presentation has emanated from research conducted with the financial support of Science Foundation Ireland under Grant number [18/CRT/6214 ] E-PS-08-032 PD-L1 IHC 22C3 pharmDx: scoring concordance on melanoma specimens J. Musser*, E. Olander, S. Tabuena-Frolli, E. Manna, K. Kersch, J. Christian, S. Hund *Agilent Technologies, USA Background & objectives: PD-L1 IHC 22C3 pharmDx is a qualitative assay used to detect PD-L1 expression in formalin-fixed, paraffin-embedded (FFPE) tissues. This study aims to establish equivalency in PD-L1 scoring on melanoma specimens between glass slides and whole slide images (WSIs). Methods: Thirty melanoma specimens were immunostained with PD-L1 IHC 22C3 pharmDx (Code SK006) and scored using a light microscope. Digital images using Aperio Scanner were scored using ImageScope software. Three observers scored blinded and randomized slides/images. Con- cordance of PD-L1 expression was analysed. A two-sided 95% confi- dence interval (CIs) for overall agreement (OA) was calculated using a percentile bootstrap method. Results: Samples analysed represented the dynamic range of PD-L1 expression and included 10 (33.3%) near cutoff specimens. The data was analysed for PD-L1 binary expression status (positive/negative) based on a MEL Score cutoff of 2 (≥1%). A total of 90 comparisons were made to the reference condition (glass slides). To meet the acceptance criteria (AC), the lower bound of the two-sided 95% percentile bootstrap CI com- puted on OA must meet or exceed 80%. The point estimate for OA was 91.1%. The 95% CI lower-bound for OA was 86.7%, meeting the AC. Conclusion: Digital scoring for PD-L1 expression in WSIs has been of great interest, especially in recent years. Binary PD-L1 expression status (positive/negative) concordance in melanoma specimens was achieved between glass slide and WSI scoring. These results support equivalency of PD-L1 scoring with MEL Score on both glass slides and WSIs. E-PS-08-033 Placental angiogenesis in foetal death E. Nardi*, E. Olivo, F. Ugolini, F. Castiglione *Section of Anatomic Pathology, Department of Health Sciences, Uni- versity of Florence, Italy Background & objectives: Placental dysfunction is one of the causes of intrauterine foetal demise (IUFD). The role of placental angiogenesis factors hasn’t been fully investigated. The aim of this study was to analyse the vascular burden through automated digital analysis in IUFD placentas. Methods: We morphologically evaluated 37 formalin-fixed and par- affin-embedded placental tissue samples from IUFD pregnancies (of which 32 were c-kit mutated) and 16 of healthy pregnancies. Repre- sentative sections were immunohistochemically stained with anti- CD31 (clone JC70, Ventana medical system). Positive vessels area was assessed by HALO digital image software. Data were expressed as posi- tive stained area. Statistical analysis was made with Mann-Whitney test. Results: IUFD placentas revealed the presence of hypoxic-ischemic state characterized by predominance of dysmorphic and hypo-vascular mature intermediate villi and thrombosis of staminal vessels while such features were not observed in healthy placenta specimens. These pathological changes were particularly highlighted in IUFD placentas characterized by c-KIT gene mutation. Digital analysis of immunohistochemical CD31 stained sections showed decreased positive stained area in IUFD placentas compared to healthy tissues (p = 0.0002). Conclusion: The present study showed that IUFD placentas showed morphological and phenotypical evidence of altered and decreased angiogenesis which may conduct to an altered placental structure and vascular development possibly leading to foetal death. E-PS-08-034 A fully automatic tumour infiltrating lymphocytes assessment tool R. Peyret*, A. Moreau, S. Sockeel, M. Petit, S. Touioui, B. Jean Jacques, E. Lanteri, M. Sockeel, J. Adam *Primaa, France Background & objectives: Tumour infiltrating lymphocytes (TILs) quan- tification has proven a reliable prognosis factor in breast cancer. Despite efforts to standardize TILs scoring, it is subject to inter-reader variability. In this context, we propose a fully automatic tool for TILs grading onWSI. Methods: The proposed processing pipeline includes three separate steps. The first one consists of a Deep Learning cancer localisation model. Then follows a combined segmentation of stroma and lymphocytes on cancer regions. This is performed using a convolutional backbone with two seg- mentation heads, that is trained through a custom multi-phase process. The final TILs score is computed from the resulting segmentations. Results: The proposed pipeline showed state-of-the-art performance TIGER dataset with a DICE of 84.9 ± 1.0 for stroma segmentation and 84.4 ± 0.4 for TILs nuclei segmentation. To prove reliability and generalizability, the algorithm was tested on publicly available TIGER dataset and on an in-house dataset of both biopsies and surgical sam- ples. We investigated the correlation between the final TILs scores and the pathologists scores on both datasets. Conclusion: The resulting algorithm can be integrated into a WSI analysis pipeline and, with such performance, provide pathologists with an automatic TILs assessment tool. E-PS-08-035 Reproducibility of serrated lesions and polyps on digital pathology, an interobserver study C. Ravaioli*, F. Ambrosi, C. Ricci, M.E. Maracci, A.G. Corradini, D. Malvi, F. Vasuri, M.L. Tardio, R. Gafà, A. Cadioli, F. Chiarucci, A. Zangrandi, M. Camponara, D.D. Di Nanni, M. Fiorentino
Made with FlippingBook
RkJQdWJsaXNoZXIy Mzg2Mjgy