To construct a CT radiomics model combined with clinical-imaging features based on the tabular prior-data fitted network (TabPFN) algorithm to enhance the preoperative prediction accuracy for the invasiveness of subsolid lung adenocarcinoma.

This retrospective study enrolled 330 patients with postoperative pathologically confirmed subsolid lung adenocarcinoma who underwent surgical treatment at multiple centers between January 2019 and January 2023. Patients were randomly divided into a training set (n=231) and a validation set (n=99) at a 7:3 ratio. Preoperative thin-slice high-resolution CT images were segmented to extract 1,239 quantitative radiomics features. A feature engineering pipeline was applied to optimize the feature set, involving: batch t-test for screening differential features, Pearson correlation analysis to remove redundant features (|r|≥0.9), and LASSO regression to construct a radiomics score (Radscore). Clinical-imaging features including patient age, sex, nodule size, consolidation-to-tumor ratio (CTR), lesion location, and morphological signs were also extracted. Multiple machine learning models were built for the binary classification task using “Radscore + clinical-imaging features” as input, including TabPFN, random forest (RF), extreme gradient boosting (XGBoost), and light gradient boosting machine (LightGBM). Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, precision, recall, and F1 score.

Among the 330 patients, pathology confirmed 101 cases (30.61%) as non-invasive adenocarcinoma and 229 cases (69.39%) as invasive adenocarcinoma. The TabPFN model demonstrated the best performance, achieving an AUC of 0.839, accuracy of 0.798, precision of 0.853, recall of 0.877, and an F1 score of 0.865. The interpretation of the model based on SHAP analysis indicated that Radscore and CTR are the most reliable features for predicting the invasiveness of subsolid lung adenocarcinoma in the TabPFN model.

The analysis combining CT radiomics and clinical-imaging features based on the TabPFN model can effectively achieve preoperative prediction of invasiveness in subsolid lung adenocarcinoma. This provides an objective reference for clinical decision-making regarding individualized surgical strategies, such as lobectomy versus sublobar resection, and holds significant clinical application value.