Deep Learning-Based Cancer Region Segmentation from H&E Slides for HPV-Related Oropharyngeal Squamous Cell Carcinomas

Tumor region automated segmentation from the digitized hematoxylin and eosin stained histology image is a fundamental step for efficient tumor quantification and biomarker interrogation. In this study, we presented an automated deep learning-based tumor segmentation model for automated tumor extent delineation in whole slide tissue images of p16-positive oropharyngeal squamous cell carcinomas (n = 248). The employed ResNet model was trained using images with coarse annotations (i.e., polygon-style bounding box annotations). The model was trained using n = 194 images and validated using n = 49 images. Another cohort of five whole slide images was used as independent test purpose. The experimental result demonstrated that satisfactory segmentation results could be achieved, an accuracy of about 90% in both of the validation and test cohorts, even when using non-exhaustive tumor annotations for training the model. Such an accurate and efficient tumor detection model could be used for early detection of disease and the prediction of aggressiveness in oropharyngeal squamous cell carcinomas, which could improve the patients’ survival to manage their therapeutic strategies appropriately.