Objectives: Lung adenocarcinomas which manifest as ground-glass nodules (GGNs) have different degrees of pathological invasion and differentiating among them is critical for treatment. Our goal was to evaluate the addition of marginal features to a baseline radiomics model on computed tomography (CT) images to predict the degree of pathologic invasiveness.
Methods: We identified 236 patients from two cohorts (training, n = 189; validation, n = 47) who underwent surgery for GGNs. All GGNs were pathologically confirmed as adenocarcinoma in situ (AIS), minimally invasive adenocarcinoma (MIA), or invasive adenocarcinoma (IA). The regions of interest were semi-automatically annotated and 40 radiomics features were computed. We selected features using L1-norm regularization to build the baseline radiomics model. Additional marginal features were developed using the cumulative distribution function (CDF) of intratumoral intensities. An improved model was built combining the baseline model with CDF features. Three classifiers were tested for both models.
Results: The baseline radiomics model included five features and resulted in an average area under the curve (AUC) of 0.8419 (training) and 0.9142 (validation) for the three classifiers. The second model, with the additional marginal features, resulted in AUCs of 0.8560 (training) and 0.9581 (validation). All three classifiers performed better with the added features. The support vector machine showed the most performance improvement (AUC improvement = 0.0790) and the best performance was achieved by the logistic classifier (validation AUC = 0.9825).
Conclusion: Our novel marginal features, when combined with a baseline radiomics model, can help differentiate IA from AIS and MIA on preoperative CT scans.
Key points: • Our novel marginal features could improve the existing radiomics model to predict the degree of pathologic invasiveness in lung adenocarcinoma.
Keywords: Classification; Lung adenocarcinoma; Machine learning; Quantitative evaluation; Tumor microenvironment.