CT-based habitat radiomics and deep learning for preoperative prediction of spread through air spaces in early-stage lung adenocarcinoma

Abstract Objective Spread through air spaces (STAS) is an invasive pattern of lung adenocarcinoma associated with increased recurrence risk and poor prognosis. Preoperative identification of STAS may inform surgical strategy selection. This study aimed to develop and validate a fusion model combining CT-based habitat radiomics and deep learning for preoperative prediction of STAS in early-stage lung adenocarcinoma. Materials and methods We retrospectively analyzed 248 patients with stage I lung adenocarcinoma (120 STAS-positive and 128 STAS-negative), comprising a training cohort (173 cases) and a testing cohort (75 cases). CT-based habitat analysis was performed by applying K-means clustering to pooled voxel-level feature vectors from all tumors, yielding four habitat subregions. After feature selection, five model categories were constructed and compared: conventional radiomics, habitat radiomics, multichannel 2D deep learning, 3D deep learning, and a deep learning-habitat fusion model. All models were evaluated using the area under the receiver operating characteristic curve (AUC), calibration curves, and decision curve analysis. Results Among the conventional and habitat radiomics models, Habitat H4—one of the four CT-based clustered habitat subregions—achieved the highest testing-cohort AUC at 0.802 (95% CI: 0.700–0.904), outperforming the conventional whole-tumor radiomics model (AUC = 0.739, 95% CI: 0.626–0.853). For the deep learning models, the 3D deep learning model (testing-cohort AUC = 0.815, 95% CI: 0.715–0.915) outperformed the multichannel 2D deep learning model (testing-cohort AUC = 0.747, 95% CI: 0.633–0.862). Among all models, the fusion model combining Habitat H4 and 3D deep learning achieved the highest numerical performance. The AUCs in the training and testing cohorts were 0.913 (95% CI: 0.873–0.953) and 0.824 (95% CI: 0.727–0.920), respectively, and were significantly higher than those of the radiomics model. DeLong test revealed that the fusion model significantly outperformed the conventional radiomics model ( P < 0.05). Decision curve analysis revealed higher clinical net benefit of the fusion model, and calibration curves showed favorable consistency between predicted and actual probabilities. Conclusion The fusion of habitat radiomics and 3D deep-learning features showed favorable performance for noninvasive preoperative prediction of STAS in early-stage lung adenocarcinoma, with the two feature representations providing potential complementary value. This model may aid preoperative risk stratification, but its role in surgical decision-making requires further validation in external and prospective studies.