A deep learning approach for predicting visceral pleural invasion in cT1 lung adenocarcinoma.

Background: Currently, patients with cT1 stage non-small cell lung cancer measuring less than 2 cm are recommended to undergo sublobar resection. However, it should be noted that there is tumor heterogeneity within these lung nodules. Potential visceral pleural invasion (VPI) is regarded as a significant factor that contributes to local recurrence and poorer prognosis after sublobar resection and postoperative upstaging of the T-stage. Currently, there are no effective techniques for preoperative or intraoperative prediction of the status of VPI in lung nodules. The primary objective of this study is to develop a machine learning model for the non-invasive prediction of VPI, thereby providing surgical decision-making support for surgeons during operations.

Methods: A total of 983 patients with nodules located within 5 mm distance from the pleura were included. Machine learning models were developed utilizing preprocessed 2D, 2.5D, and 3D computed tomography (CT) imaging data. These models were employed to predict the status of VPI, leveraging radiomics and deep learning techniques. The aforementioned three groups of data were further categorized into region of interest (ROI)-only (exclusively focused on the ROI) and ROI-rect (the minimum bounding rectangle or cuboid of the ROI) groups, based on whether they included images outside the ROI. Receiver operating characteristic (ROC) curves were created for the assessment of predictive accuracy across different models. Employing the Youden's index, patients were categorized into high-risk and low-risk groups based on the model's criteria, which was then followed by an in-depth analysis of overall survival rates among the distinct patient cohorts.

Results: This study included a training cohort of 786 patients and a validation cohort of 197 patients. In comparison to radiomic and radiological models, deep learning models, especially the 2D-rect model, demonstrated better predictive performance. Although the 3D-ROI-only model exhibited the highest areas under the curve (AUC) value (0.952), its predictive performance for the status of VPI was found to be inferior according to the decision curve, calibration curve, and survival analysis.

Conclusions: The developed deep learning signature offers a robust instrument for the precise prognostication of vascular invasion in stage cT1 lung adenocarcinoma, thereby enhancing stratification for prognostic evaluation. Moreover, the application of this advanced computational model aids in the refinement of therapeutic approach formulation for individuals diagnosed with cT1 lung adenocarcinoma.