Perioperative CT fractal analysis for early recurrence risk stratification in advanced gastric cancer: development and validation of a prognostic model

Early recurrence after curative resection occurs in approximately 25–30% of patients with advanced gastric cancer (AGC) and is associated with dismal prognosis. However, reliable preoperative tools for identifying high-risk patients remain unavailable. This study aimed to develop and validate a perioperative prediction model integrating CT-based fractal features and clinicopathological variables for early recurrence risk stratification in AGC. This multicenter retrospective study included 750 AGC patients who underwent R0 gastrectomy (training cohort, n = 413; internal validation, n = 178; external validation, n = 159). Early recurrence was defined as locoregional recurrence, peritoneal seeding, lymph node metastasis or distant organ metastasis confirmed by imaging or pathology within 24 months after R0 curative gastrectomy; isolated tumour-marker elevation without imaging or pathological confirmation was not counted as recurrence. Three-dimensional fractal dimension (FD_final) was extracted from preoperative contrast-enhanced CT to quantify tumour structural complexity. A combined model integrating FD_final, serum CA19-9 and postoperative histological differentiation was developed. Model performance was assessed by AUC, calibration, and decision curve analysis (DCA). Early recurrence rates were 25.9%, 25.8%, and 29.6% in the three cohorts. FD_final, a three-dimensional multiscale morphologic descriptor of tumour structural complexity, was significantly higher in patients who developed early recurrence (all p < 0.001). On multivariable analysis, FD_final (per 1-SD increase: OR = 2.98, 95%CI:2.00–4.43), poor histological differentiation (OR = 6.34), and elevated CA19-9 (OR = 5.71) were independent predictors. The combined model achieved AUCs of 0.833, 0.803, and 0.801 in the three cohorts and outperformed the two-variable clinical model comprising histological differentiation and CA19-9 (paired ΔAUC CIs excluding zero). Exploratory decision-curve analysis suggested positive net benefit across illustrative threshold probabilities. CT‑quantified fractal dimension provides independent prognostic information beyond conventional clinicopathological factors. This perioperative model may serve as a non‑invasive reference for early recurrence risk stratification in AGC patients undergoing upfront R0 D2 gastrectomy; prospective external validation is required before clinical application.