Multimodal deep learning for preoperative invasiveness stratification of lung adenocarcinoma spectrum nodules

Preoperative ternary stratification of lung adenocarcinoma-spectrum nodules remains challenging, particularly for minimally invasive adenocarcinoma (MIA). We developed and externally validated a multimodal deep-learning framework integrating three-dimensional CT nodule patches with structured radiological semantic features to stratify atypical adenomatous hyperplasia/adenocarcinoma in situ (AAH/AIS), MIA, and invasive adenocarcinoma (IAC). Consecutive patients with surgically resected, pathologically confirmed nodules were retrospectively enrolled from three centers, with postoperative pathology as the reference standard. The Center 1 development cohort included 2004 patients/2208 nodules and was split at the patient level into training (1603/1764) and internal validation (401/444); external validation used Center 2 (446/483) and Center 3 (276/378). Internally, the multimodal model achieved an AUC of 0.914 (95% CI, 0.894–0.932), exceeding CT image-only (0.874; absolute gain, 0.040; P < 0.001) and semantic-only (0.867; absolute gain, 0.047; P < 0.001) models. External AUCs were 0.879 (0.856–0.902) in Center 2 and 0.895 (0.868–0.917) in Center 3, with significant improvements over CT image-only and semantic-only models. These findings support structured radiological semantic features as a clinically interpretable complementary input to CT representations and could inform preoperative invasiveness stratification in decision-support workflows for preoperatively suspected and surgically considered lung adenocarcinoma-spectrum nodules.