Artificial Intelligence-Assessed Tumor Central Tumor-Infiltrating Lymphocytes Predict Recurrence-Free Survival in Transurethral Resection of Bladder Tumor Specimens

초록

Background Tumor-infiltrating lymphocytes (TILs) represent a critical component of the tumor microenvironment, but objective assessment in transurethral resection of bladder tumor (TURB) specimens remains challenging due to artifacts and inter-observer variability. Traditional prognostic factors including pathologic T category, carcinoma in situ, lymphovascular invasion, and histologic variants require systematic re-evaluation in TURB specimens where sampling limitations may affect accurate assessment. We developed an AI-based TIL quantification system to investigate prognostic significance of spatially-resolved TIL densities alongside traditional clinicopathologic parameters. Design We analyzed 84 consecutive TURB specimens from bladder cancer patients (39 non-muscle invasive, 45 muscle invasive with complete clinicopathological data and long-term follow-up. External validation cohort comprised 200 patients from independent institution. Our AI pipeline integrated a custom encoder-decoder segmentation network achieving 82% IoU for tumor and 81% for stroma identification. YOLOv8-based detection algorithm was employed for lymphocyte detection, achieving mAP50 of 81.71%. TIL densities were calculated separately for tumor center (Central TIL) and invasive front (Stromal TIL). Statistical analysis included Kaplan-Meier survival analysis, Cox proportional hazards regression with backward elimination. Results With median follow-up of 34.7 months, 49 patients (58.3%) experienced recurrence and 49 died. Central TIL dichotomized at median (12.5×103) significantly stratified recurrence-free survival (RFS) (log-rank p=0.029), with 5-year RFS rates of 62.4% (high) versus 36.8% (low). Central TIL ratio demonstrated independent protective effect for RFS in univariate and multivariate analysis (HR 0.964, p=0.037), representing 3.6% reduction in recurrence risk per unit increase. Stromal TIL ratio showed borderline significance for overall survival (HR 0.966, p=0.059). Among traditional factors, pT category emerged as strongest predictor for overall survival (multivariate HR 2.376, 95%CI: 1.288-4.384, p=0.006). Preliminary external validation confirmed Central TIL's prognostic value in 200 patients.Conclusions AI-assessed Central TIL emerged as independent RFS biomarker, likely reflecting better tumor center preservation versus cautery-affected margins. Traditional factors lost significance in TURB specimens, emphasizing Central TIL as superior prognostic marker in this context.