A fusion model based on tumor and peritumoral CT radiomics for differentiating bronchiolar adenoma from lung adenocarcinoma.

PURPOSE: This study aimed to develop and validate a combined clinical and tumor-peritumoral CT radiomics model to differentiate bronchiolar adenoma (BA) from lung adenocarcinoma (LUAD), thereby improving preoperative diagnostic accuracy and guiding individualized treatment strategies.

METHODS

A total of 362 patients with pathologically confirmed BA or LUAD were retrospectively analyzed. Data from Medical Center 1 (n = 281) were divided into training and test sets (7:3 ratio), and data from Medical Center 2 (n = 81) served as an external validation set. Clinical characteristics, CT morphological features, and tumor-peritumoral radiomics features were extracted. Five machine learning algorithms were applied to construct and compare predictive models.

RESULTS

Lung lobe distribution, density, vacuolar sign, tumor-associated vessels, distance to pleura, and nodule diameters differed significantly between BA and LUAD. Among radiomics models, the tumor-peritumoral MLP model achieved the best performance (AUCs: 0.918, 0.912, 0.888). The clinical-radiomics fusion model outperformed single models, with AUCs of 0.935, 0.939, and 0.910 and accuracies of 0.862, 0.847, and 0.864 in the training, test, and validation sets, respectively.

CONCLUSION

The proposed fusion model enables accurate, non-invasive differentiation between BA and LUAD, offering valuable support for personalized clinical decision-making.