Multiscale Spatial Fusion Feature‐Driven Characterization of Gastric Cancer Invasive Margins: A Multicenter Cohort Study for Preoperative Accurate Differentiation Between T4a and T4b Subtypes

Accurate preoperative differentiation of gastric cancer T4a/b stages is crucial for surgical planning and prognosis. However, conventional CT assessments often yield suboptimal staging accuracy due to visual limitations and inadequate peritumoral microinvasion quantification. This study developed a multi-scale spatial feature fusion model based on extended regions of interest (eROI) for precise preoperative T4a/b differentiation. We proposed the GAVR model with a three-tier architecture: a Boundary-Augmented U-Net for eROI generation incorporating the peritumoral microenvironment; parallel pathways extracting conventional radiomics, 2D, and 3D deep learning features; and a Vision Transformer for global attention-weighted fusion and discriminative representation learning. The model was validated across a multicenter cohort of 1804 patients, including internal, external, and prospective sets. A blinded reader study involving 16 radiologists evaluated its clinical utility. GAVR demonstrated exceptional generalizability, achieving AUCs of 0.987 and 0.979 in two independent external sets and 0.987 prospectively. Ablation studies confirmed the necessity of multi-scale features. GAVR assistance significantly improved radiologists' diagnostic accuracy (0.609 to 0.795) and reduced reading time by 60%. By deeply fusing multi-scale spatial features, GAVR characterizes structural heterogeneity in complex gastric cancer invasive margins and mitigates overfitting. It demonstrates clear translational value as an embeddable decision-support tool for multidisciplinary gastric cancer management.