Transformer‐Based PET/CT Fusion Enables Preoperative Prediction and Prognostic Stratification of Spread Through Air Spaces in Lung Adenocarcinoma

Spread through air spaces (STAS) is associated with recurrence and unfavorable outcomes in lung adenocarcinoma, yet reliable preoperative identification remains challenging. This multicenter retrospective study includes 212 patients with 219 tumors across three institutions and develops a transformer-based multimodal framework integrating eight complementary 2.5D PET and CT representations for preoperative STAS risk assessment. The model achieves an area under the receiver operating characteristic curve of 0.822 (95% CI, 0.725-0.920) in the independent external test cohort, with higher discrimination than the evaluated single-modality and conventional fusion approaches. Beyond STAS classification, the continuous model-derived risk score is associated with progression-free survival after adjustment for pathological T stage, N stage, and maximum tumor diameter (HR, 4.24; 95% CI, 1.47-12.25; P = 0.008), although the limited number of progression events warrants cautious interpretation. Exploratory transcriptomic analyses across two public cohorts further nominate candidate programs involving cell adhesion, vesicle trafficking, complement activity, and metabolic remodeling. Together, these findings highlight the potential of multimodal PET/CT fusion to integrate morphological and metabolic information for noninvasive STAS risk stratification and support further prospective multicenter evaluation toward individualized surgical decision support.