<p>Diagnostic delays in lung cancer compromise patient survival. This study aims to develop a deep learning-based framework that integrates rapid on-site evaluation (ROSE) cytomorphology and serum tumor markers to enable accurate pre-pathological subtyping, thereby facilitating earlier treatment initiation. A dataset of 156 matched cases with both ROSE cytology images and five serum biomarkers (squamous cell carcinoma antigen (SCCA), pro-gastrin-releasing peptide (ProGRP), carcinoembryonic antigen (CEA), neuron-specific enolase (NSE), cytokeratin-19 fragment (CYFRA21-1)) was retrospectively analyzed. Two deep learning models were developed: (1) a ROSE Image-Only Model (RIOM) using ResNet-50 with spatial attention, and (2) a ROSE-Serum Marker Model (RSMM) incorporating cross-modal feature alignment between ROSE images and serum biomarkers. A consensus strategy was implemented to stratify patients based on prediction agreement between the two models. Manual ROSE assessment achieved 84.0% overall diagnostic accuracy. The RIOM matched this performance (84.5% accuracy), while the multimodal RSMM significantly surpassed both, achieving 91.6% accuracy in five-class classification (Benign, lung adenocarcinoma (LUAD), lung squamous cell carcinoma (LUSC), large cell lung carcinoma (LCLC), small cell lung cancer (SCLC)). The consensus strategy identified 80.6% of cases as prediction-consistent, for which the accuracy for malignancy determination, non-small cell lung cancer (NSCLC) discrimination, and final subtype classification reached 98.4%. The proposed multimodal decision-support framework provides a rapid and reliable tool for pre-pathological lung cancer subtyping. By enabling high-confidence diagnoses immediately after biopsy, this approach could rationalize clinical workflows through risk-stratified management and significantly accelerate reflex molecular testing and therapy initiation, particularly for the majority of patients with prediction-consistent results.</p>