Deep learning–derived CT body composition enhances survival risk stratification beyond the TNM system in locally advanced gastric cancer: a multi-modality cohort study

BACKGROUND: Survival outcomes in locally advanced gastric cancer remain heterogeneous despite standard treatment and outcome classifications. Visceral adiposity has increasingly emerged as a prognostic factor, yet its mechanistic and clinical utilization remains underexplored. METHODS: This retrospective cohort study evaluated 227 American Joint Committee on Cancer 8th edition stage III gastric cancer patients undergoing curative gastrectomy (2007-2022) at a tertiary referral center. A deep learning-enabled UNet++ model quantified computed tomography-based body composition (CTBC) metrics, validated against manual segmentation. Subsets of patients underwent plasma metabolomic ( n = 86) and tumor immune-metabolic profiling ( n = 40) using mass spectrometry, immunohistochemistry, and 35-color flow cytometry. Median follow-up was 33 months. RESULTS: Automated CTBC analysis showed excellent concordance with manual segmentation ( r2 > 0.85). Low subcutaneous adipose tissue (SAT) index and high visceral-to-subcutaneous adipose tissue (VAT/SAT) ratio independently predicted worse disease-free and overall survival (hazard ratio: 1.4-1.6). Incorporating CTBC metrics significantly improved survival models (likelihood ratio P < 0.03; ΔAIC > 4). A high VAT/SAT ratio correlated with increased plasma acylcarnitines and decreased phosphatidylcholines, indicating impaired mitochondrial fatty acid oxidation and altered lipid and membrane remodeling. Tumors from high VAT/SAT patients showed upregulated CPT1, downregulated CPT2/CACT, increased IDO1/AHR expression, and elevated immunosuppressive CD4 + EMRA and regulatory T cell infiltration (all P < 0.05). CONCLUSION: Deep learning-derived CTBC metrics, especially VAT/SAT ratio, enhance prognostic stratification beyond TNM staging in locally advanced gastric cancer. This ratio captures a systemic and tumor-level immunometabolic phenotype marked by mitochondrial dysfunction and immune suppression. Our findings highlight VAT/SAT as a noninvasive, clinically actionable biomarker to guide personalized therapy and risk-adapted algorithm in gastric cancer management.