CT-Based Deep-Learning Model for Spread-Through-Air-Spaces Prediction in Ground Glass-Predominant Lung Adenocarcinoma

Background Sublobar resection is strongly associated with poor prognosis in early-stage lung adenocarcinoma, with the presence of tumor spread through air spaces (STAS). Thus, preoperative prediction of STAS is important for surgical planning. This study aimed to develop a STAS deep-learning (STAS-DL) prediction model in lung adenocarcinoma with tumor smaller than 3 cm and a consolidation-to-tumor (C/T) ratio less than 0.5. Methods The study retrospectively enrolled of 581 patients from two institutions between 2015 and 2019. The STAS-DL model was developed to extract the feature of solid components through solid components gated (SCG) for predicting STAS. The STAS-DL model was assessed with external validation in the testing sets and compared with the deep-learning model without SCG (STAS-DL woSCG ), the radiomics-based model, the C/T ratio, and five thoracic surgeons. The performance of the models was evaluated using area under the curve (AUC), accuracy and standardized net benefit of the decision curve analysis. Results The study evaluated 458 patients (institute 1) in the training set and 123 patients (institute 2) in the testing set. The proposed STAS-DL yielded the best performance compared with the other methods in the testing set, with an AUC of 0.82 and an accuracy of 74%, outperformed the STAS-DL woSCG with an accuracy of 70%, and was superior to the physicians with an AUC of 0.68. Moreover, STAS-DL achieved the highest standardized net benefit compared with the other methods. Conclusion The proposed STAS-DL model has great potential for the preoperative prediction of STAS and may support decision-making for surgical planning in early-stage, ground glass-predominant lung adenocarcinoma.