Prediction of Lymphovascular Invasion in Early-Stage Lung Adenocarcinoma Using Artificial Intelligence-Based...

Objectives: This study utilized artificial intelligence (AI)-based radiomics analysis of computed tomography (CT) images using a modified U-Net for lung nodule segmentation and convolutional neural network based on VGG-16 to predict lymphovascular invasion (LVI) in stage 0-I lung adenocarcinoma. Additionally, the study investigated whether combining radiomics data with serum microRNA (miR)-30d level as a potential biomarker could enhance predictive performance. Methods: A total of 1265 patients who underwent complete resection between 2008 and 2018 were included. AI-based CT analysis was performed, and logistic regression was applied to predict LVI using 35 imaging features. A risk score (RS) generated from 840 patients in the derivation cohort was used to identify a high-risk group, with validation performed using 425 patients. Additionally, 47 cases with extracellular vesicle (EV)-derived miR-30d level data were analyzed to evaluate the value of the integrated approach. Results: Among all the patients, 467 patients (36.9%) were LVI-positive, and LVI was independently associated with poorer overall survival. The receiver operating characteristic curve for LVI based on the RS yielded an area under the curve of 0.899. For LVI prediction, the sensitivity, specificity, and accuracy were 84.8%, 83.7%, and 83.9%, respectively, in the derivation group, and 82.3%, 79.4%, and 80.5%, respectively, in the validation group. The integrated approach with miR-30d enhanced the predictability of LVI, achieving a sensitivity of 93.3%, specificity of 70.5%, and accuracy of 85.1%. Conclusions: AI-based radiomics demonstrated high effectiveness for predicting LVI, with RSs showing broad clinical applications. The addition of EV-derived miR-30d modestly improved predictability.