Integration of 2D/3D deep learning and radiomics for predicting lymphovascular invasion in T1-stage invasive lung adenocarcinoma: a multicenter study

We developed a combined model integrating radiomics,2D deep learning (DL), and 3D DL to predict lymphovascular invasion (LVI) status in patients with T1-stage invasive lung adenocarcinoma (LUAD). This retrospective study included 334 patients who underwent radical surgery from four academic medical centers. We constructed corresponding predictive models by extracting and analyzing conventional radiomic features, 2D DL features, and 3D DL features from the tumor regions in CT images. These features were then integrated to develop a combined model to identify LVI status in T1-stage invasive LUAD patients. The model's performance was evaluated by calculating the area under the receiver operating characteristic (ROC) curve (AUC). The established combined model demonstrated excellent performance in distinguishing LVI, with predictive capabilities superior to individual models, yielding AUC values of 0.958（95%CI :0.9294 - 0.9863）, 0.886（95%CI : 0.7938 - 0.9786）, and 0.884（95%CI : 0.8277 - 0.9401） for the training, internal validation, and external test sets, respectively. Decision curve analysis (DCA) indicated that the net benefit provided by the combined model surpassed that of other radiomic models, offering critical information for treatment decision-making.