ABSTRACT
PURPOSE: The prognostic and predictive significance of pathologist-read tumor infiltrating lymphocytes (TILs) in head and neck cancers have been demonstrated through multiple studies over the years. TILs have not been broadly adopted clinically, perhaps due to substantial inter-observer variability. In this study, we developed a machine-based algorithm for TIL evaluation in head and neck cancers and validated its prognostic value in independent cohorts. EXPERIMENTAL DESIGN: A network classifier called NN3-17 was trained to identify and calculate tumor cells, lymphocytes, fibroblasts and "other" cells on hematoxylin-eosin stained sections using the QuPath software. These measurements were used to construct three predefined TIL variables. A retrospective collection of 154 head and neck squamous cell cancer cases was used as the discovery set to identify optimal association of TIL variables and survival. Two independent cohorts of 234 cases were used for validation. RESULTS: We found that electronic TIL variables were associated with favorable prognosis in both the HPV-positive and -negative cases. After adjusting for clinicopathologic factors, Cox regression analysis demonstrated that electronic total TILs% (p = 0.025) in the HPV-positive and electronic stromal TILs% (p < 0.001) in the HPV-negative population were independent markers of disease specific outcomes (disease free survival). CONCLUSIONS: Neural network TIL variables demonstrated independent prognostic value in validation cohorts of HPV-positive and HPV-negative head and neck cancers. These objective variables can be calculated by an open-source software and could be considered for testing in a prospective setting to assess potential clinical implications.