Abstract PO-036: A deep learning alternative to regional molecular testing for HPV status

Abstract Introduction Current evidence demonstrates that HPV-positive disease in oropharyngeal cancers is indicative of a more favourable prognosis compared with HPV-negative tumors and is recognised as a separate entity in TNM8 staging guidelines. Immunohistochemistry for p16 can indicate HPV-related disease; however, regional access to sophisticated molecular assays may be required to facilitate patient stratification by obtaining definitive HPV status using second-line tests. In this study, we develop and test if artificial intelligence could be robustly applied to diagnostic H&E stained slides in the absence of HPV-specific testing at local institutions. Methods Representative digital images of H&E stained oropharyngeal cancers (Σn=765) from the UK, Europe, South America and the USA were imported into QuPath, an open source image analysis software, for tumor annotation. After annotation, regions of interest with >70% tissue present were extracted for model development, validation and independent testing. Image augmentation was applied to all extracted tiles to reduce effects of staining variability and scanner variance. Transfer learning was used with a pre-trained EfficientNet-V1-B4 architecture for model development. Model performance and explainability was assessed using both computational and biological approaches to support pathological review of model predictions. Results Using a 70:15:15 patient split for training, validation and test data from a population representative cohort we developed a novel model, which demonstrated 90% accuracy within the withheld test set of Northern Irish oropharyngeal cancer patients. Our model achieved similar accuracy when tested using independent cohorts from the UK (84%), Europe (88%), South America (79%), and the USA (84%). Use of explainable artificial intelligence methodology indicated that contributions from both stromal and tumor epithelial compartments were equally important when determining HPV status when using our model. Review of block-like prediction heatmap’s generated by the model were found to be associated with heterogeneous histology and focal p16 expression. Conclusion Defining HPV status based on morphological assessment has historically been complicated by subjective assessment. This work presents a novel model that accurately determines HPV status from H&E stained samples by enabling reproducible and robust qualitative assessment of the tissues using artificial intelligence to predict HPV status in the absence of molecular testing. Our study demonstrates that use of p16 immunohistochemistry to validate our model significantly enhanced biological understanding of model performance in situ. We anticipate that our process of model development and validation using epidemiological datasets and orthogonal sources of biomarker classification in this study to be a starting point for the development of clinically explainable, AI-derived patient stratification strategies in tissue-based histopathological assessments which aim to be employed remotely as an aid to clinical diagnostics. Citation Format: Stephanie G. Craig, Richard Gault, Kristopher D. McCombe, Andrew Moyes, Youcheng Sun, Tao Wang, Andrew Schache, Terry M. Jones, Janet M. Risk, Philip Gunning, Valerie Gaborieau, Paul Brennan, Behnoush Abedi-Ardekani, Jacqueline A. James. A deep learning alternative to regional molecular testing for HPV status [abstract]. In: Proceedings of the AACR-AHNS Head and Neck Cancer Conference: Innovating through Basic, Clinical, and Translational Research; 2023 Jul 7-8; Montreal, QC, Canada. Philadelphia (PA): AACR; Clin Cancer Res 2023;29(18_Suppl):Abstract nr PO-036.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.