Automated tumor-stroma ratio estimation for improved prognostic stratification of squamous cell carcinoma of the oral tongue

Abstract

Squamous cell carcinoma of the oral tongue (SCCOT) represents an aggressive malignancy characterized by high metastatic potential and significant heterogeneity in its tumor microenvironment. The tumor-stroma ratio (TSR) has emerged as a prognostic biomarker, with higher stromal content frequently correlating with worse survival outcomes. Traditional approaches using the standard 50% TSR cutoff may not be optimal for SCCOT, and visual TSR estimation introduces variability during TSR region annotation. This study aimed to develop and validate a dedicated TSR estimation model for SCCOT by incorporating representative TSR regions from the invasive tumor front of whole slide images and to determine the optimal TSR threshold for prognostic stratification. Using hematoxylin and eosin-stained images from The Cancer Genome Atlas as a discovery cohort and whole slide images from Norrland's University Hospital Umea, Sweden (NUS) as a validation cohort, we developed a computational model to estimate TSR. The model demonstrated a high correlation with pathologist-based TSR estimation in both discovery (R = 0.848, p < 0.01) and validation (R = 0.783, p < 0.01) cohorts. The optimal 55% cutoff identified by the model improved prognostic accuracy over the traditional 50% threshold, with patients having high stroma within the tumor invasive front showing worse overall (log-rank p = 0.006) and disease-specific (log-rank p = 0.016) survival. Our computational TSR model for SCCOT demonstrates that automated TSR estimation enhances prognostic accuracy at an optimal cutoff of 55%, contributing to more precise risk stratification and potentially enabling personalized treatment strategies in SCCOT management.