Attention mechanism-based habitat analysis for predicting pleural invasion and prognosis of pulmonary nodules.

Background: The use of segmental resection in pulmonary adenocarcinoma is increasing, yet visceral pleural invasion (VPI) remains a critical risk factor impacting overall survival (OS). The benefits of segmental resection for these patients are unclear, and non-invasive methods to predict VPI need further development. This study aims to develop a predictive model for VPI and OS, aiding surgeons in preoperative and intraoperative decision-making.

Methods: A retrospective study was conducted using data from the Sun Yat-sen Memorial Hospital, the Fifth Affiliated Hospital of Sun Yat-sen University and an external dataset (named NSCLC Radiogenomics from The Cancer Imaging Archive) of cT1 stage pulmonary nodules. Original computed tomography images were enhanced using generative adversarial networks. Habitat analysis identified tumor subregions, which were clustered. Radiomics and vision transformer features were extracted and integrated using attention-equipped transformers to develop prediction models. Performance was evaluated using receiver operating characteristic (ROC) curves, net reclassification improvement, and integrated discrimination improvement.

Results: The study included 742 patients, comprising 338 males and 404 females, with a mean age of 61±10.2 years. Data from the Fifth Affiliated Hospital of Sun Yat-sen University were divided into training and validation cohorts, while data from the Sun Yat-sen Memorial Hospital and the NSCLC Radiogenomics dataset formed the test cohort. The Rad-adjacent model had an area under the curve (AUC) of 0.822 for predicting VPI, while the combined model achieved an AUC of 0.819. For predicting 5-year OS, the combined model's AUC was 0.821, compared to 0.775 for the Rad-adjacent model.

Conclusions: The developed models show strong predictive capabilities for VPI and OS in cT1 stage lung adenocarcinoma, providing valuable non-invasive support for surgical decision-making.