Can Multiparametric Digital Whole-Slide Image Analysis Improve Recurrence Risk Stratification in Oral Tongue Squamous Cell Carcinoma?
文献摘要
Recurrence in oral tongue squamous cell carcinoma (TSCC) remains a prognostic challenge, which necessitates additional prognostic approaches reflecting tumor biology. Therefore, we developed an objective, observer-independent, multiparametric model incorporating features of the tumoral and stromal compartments. Digitized whole-slide images from 80 patients with TSCC were analyzed using a combination of semi-automatic and deep learning-based approaches. We extracted quantitative histomorphological features, including tumor/stroma ratio (TSR), lymphocyte/stromal cell ratio in the tumor stroma (LSCR), and the modified Polsby-Popper (MPP) score reflecting tumor disjointedness. Human trophoblast cell surface glycoprotein (Trop-2) expression was assessed by immunohistochemistry in central and invasive-front tumor regions sampled via tissue microarrays (TMA) and quantified using the H-score. A composite score (CS) was constructed by summing points from these parameters; as Trop-2 did not improve cross-validated discrimination, a final three-marker CS (TSR, LSCR, MPP) was adopted as the final model. Patients were stratified into prognostic groups and recurrence-free survival (RFS) was analyzed. While individual parameters showed only trends, the CS stratified patients into poor, intermediate and favorable prognosis groups with 5-year RFS probabilities of 64.6%, 75.3%, and 100.0%. Repeated five-fold cross-validation yielded a cross-fitted consensus concordance index of 0.587 (mean per-repetition concordance index 0.567 ± 0.039), compared with an apparent concordance index of 0.676, with no evidence of a difference from conventional prognostic factors. In exploratory multivariable Firth-penalized Cox regression adjusted for pT3-pT4a category, the favorable group showed a lower recurrence hazard than the poor group (HR 0.104, 95% CI 0.001-0.854, p = 0.0316); this non-cross-validated estimate is fragile given the small favorable group without recurrence events. This observer-independent whole-slide image approach may provide complementary prognostic information to conventional histological features but requires external validation.