High content image analysis in routine diagnostic histopathology predicts outcomes in HPV-associated oropharyngeal squamous cell carcinomas

ObjectiveRoutine haematoxylin and eosin (H&E) photomicrographs from human papillomavirus-associated oropharyngeal squamous cell carcinomas (HPV+OpSCC) contain a wealth of prognostic information. In this study, we set out to develop a high content image analysis workflow to quantify features of H&E images from HPV+OpSCC patients to identify prognostic features which can be used for prediction of patient outcomes. MethodsWe have developed a dedicated image analysis workflow using open-source software, for single-cell segmentation and classification. This workflow was applied to a set of 567 images from diagnostic H&E slides in a retrospective cohort of HPV+OpSCC patients with favourable (n = 29) and unfavourable (n = 29) outcomes. Using our method, we have identified 31 quantitative prognostic features which were quantified in each sample and used to train a neural network model to predict patient outcomes. The model was validated by k-fold cross-validation using 10 folds and a test set. ResultsUnivariate and multivariate statistical analyses revealed significant differences between the two patient outcome groups in 31 and 16 variables respectively (P<0.05). The neural network model had an overall accuracy of 78.8% and 77.7% in recognising favourable and unfavourable prognosis patients when applied to the test set and k-fold cross-validation respectively. ConclusionOur open-source H&E analysis workflow and model can predict HPV+OpSCC outcomes with promising accuracy. Our work supports the use of machine learning in digital pathology to exploit clinically relevant features in routine diagnostic pathology without additional biomarkers.