Optimizing surgical decision-making for lung adenocarcinoma by preoperatively identifying pathological high-risk factors: retrospective development and prospective validation of a deep learning model

Background: Accurate identification of pathological high-risk factors (PHRFs) in early-stage lung adenocarcinoma (LUAD) is critical for optimizing surgical decision-making. However, reliance on intraoperative frozen section (FS) assessment is limited by insufficient sensitivity. This study aimed to retrospectively develop and prospectively validate a deep learning model (a knowledge-based graph convolutional network, KB-GCN) based on preoperative CT scans to identify PHRFs in LUAD. Methods: We retrospectively developed and externally validated a KB-GCN using two cohorts (A: 268 patients/297 lesions for training and internal validation; B: 68 patients/75 lesions for external validation). We then conducted a pre-registered, single-center, prospective observational validation in 200 consecutive surgical candidates with early-stage LUAD. Before prospective enrollment, the model architecture, weights, preprocessing pipeline, and the decision threshold (0.40, determined from the retrospective phase) were locked. For each patient, a preoperative prediction was generated before intraoperative FS and final pathology (FP); the clinical team was blinded to the model output. The performance of the locked model and FS was compared with FP. Results: In retrospective validation, the KB-GCN model achieved an area under the curve (AUC) of 0.92 (95% CI: 0.86–0.97) in the internal validation cohort and 0.88 (95% CI: 0.81–0.94) in the external validation cohort. The KB-GCN model outperformed all 6 compared classical 2D/3D CNN models (best comparative AUC: 0.79). In prospective validation, intraoperative FS achieved an overall sensitivity of 59% and accuracy of 77% for detecting PHRFs, misclassifying 40.6% (39/96) of PHRF-positive cases. In contrast, the KB-GCN model demonstrated significantly higher overall sensitivity (82%) and AUC (0.83), although with lower specificity (75% vs FS: 92%). The KB-GCN model showed superior performance in part-solid nodules (PSN, AUC: 0.86) and in 2–3 cm tumors (AUC: 0.86), with moderate performance in ≤1 cm tumors (AUC: 0.82) and in 1–2 cm tumors (AUC: 0.79). Conclusion: This study retrospectively developed and, for the first time, prospectively demonstrates that a deep learning model based on preoperative chest CT predicts PHRFs, achieving significantly higher sensitivity for identifying PHRFs in early-stage invasive LUAD than conventional intraoperative FS. Despite slightly lower specificity, the KB-GCN model effectively compensates for the critical sensitivity deficit of FS, particularly for tumors (1–3 cm) containing solid components. Preoperative deep learning assessment combined with intraoperative FS provides thoracic surgeons with more comprehensive and accurate information to optimize surgical decisions (e.g., extent of resection: lobectomy or sublobar resection). Future development requires integrating this preoperative model with intraoperative FS assessment into standardized workflows.