Document resource
Background Patients with resectable colorectal liver metastasis (CRLM) show heterogeneous outcomes after neoadjuvant chemotherapy (NAC) and resection. Overall survival (OS), the gold-standard endpoint for assessing surgical benefit, requires prolonged follow-up, delaying treatment evaluation. Currently, no guideline-recommended prognostic indicators exist for this population, and existing pathology-based surrogates lack standardized quantification and validated cut-offs. Residual viable tumor percentage (RVT%) offers a quantitative, objective, and reproducible alternative. This study developed a deep learning (DL) system to automatically derive RVT% from whole-slide images (WSIs) and evaluated its value as a surrogate OS marker.Methods A total of 511 CRLM patients who underwent resection after NAC were retrospectively included across three cohorts: discovery (n=122, Sun Yat-sen University Cancer Center [SYSUCC]), internal validation (n=136, SYSUCC), and external validation (n=253, Beijing Cancer Hospital). A UNI encoder-based model classified WSIs into six tissue classes. RVT% was defined as tumor area divided by regression bed area, with regression bed defined as regions having a combined proportion of tumor, stroma, lymphocyte, mucus, and debris ≥90% and a hepatocyte proportion ≤50%. The optimal RVT% cut-off (20%) was determined using survival analysis. Prognostic surrogate value was assessed using Kaplan-Meier and Cox regression.Results We developed a DL-based WSI analysis framework that enabled fully automated six-class tissue classification and RVT% quantification as a surrogate endpoint for OS. Using the 20% cut-off optimized, high RVT% was consistently associated with worse OS across all cohorts (discovery: HR 2.32, 95% CI 1.41–3.82, P=0.0009; internal: HR 1.85, 95% CI 1.18–2.90, P=0.0074; external: HR 2.79, 95% CI 1.47–5.29, P=0.0017). After adjustment for Clinical Risk Score, RVT% remained an independent surrogate predictor of OS in all cohorts (all P<0.001), supporting the robustness and cross-cohort generalizability of this artificial intelligence-driven biomarker.Conclusions This multicenter study established a DL-based framework for automated RVT% quantification and validated RVT% as a robust, independent surrogate prognostic biomarker for resectable CRLM after NAC. By enabling objective, reproducible, and early prognostic stratification, this approach supports more fine-grained pathological response assessment, improved risk stratification, and may inform risk-adaptive trial design, accelerate treatment evaluation, guide subsequent clinical decision-making, and help optimize follow-up intensity.