Integrating CT Radiomics and Machine Learning for Preoperative Assessment of STAS Grading and Prognostic Stratification in Lung Adenocarcinoma.
Authors
Affiliations (6)
Affiliations (6)
- Postgraduate Cultivation Base of Guangzhou University of Chinese Medicine, Panyu Central Hospital, Guangzhou, China (Q.C., S.Q., S.L., P.W.); Department of Radiology, The Affliated Panyu Central Hospital, Guangzhou Medical University, Guangzhou, China (Q.C., S.Q., S.L., Q.H., Z.A., P.W., Z.X.).
- Department of Pathology, The Affliated Panyu Central Hospital, Guangzhou Medical University, Guangzhou, China (F.C.).
- Cancer Institute of Panyu District, The Affliated Panyu Central Hospital, Guangzhou Medical University, Guangzhou, China (Z.Y.).
- Department of Radiology, The Affliated Panyu Central Hospital, Guangzhou Medical University, Guangzhou, China (Q.C., S.Q., S.L., Q.H., Z.A., P.W., Z.X.).
- CT Imaging Research Center, GE Healthcare China, Guangzhou, China (K.M.).
- Department of Radiology, The Affliated Panyu Central Hospital, Guangzhou Medical University, Guangzhou, China (Q.C., S.Q., S.L., Q.H., Z.A., P.W., Z.X.). Electronic address: [email protected].
Abstract
The distance of spread through air spaces (STAS) dissemination is associated with prognosis in lung adenocarcinoma (LUAD). This study aimed to develop a computed tomography (CT) radiomics-based machine learning model to predict STAS grade. Retrospective data included 293 LUAD patients for training/internal validation and 92 from the National Lung Screening Trial for external validation. STAS, defined as tumor cells beyond the tumor edge, was stratified by a 2500 µm threshold into grade I and II. Cox regression assessed associations between STAS grade and recurrence-free survival (RFS). Radiomics features were selected by minimum redundancy maximum relevance (MRMR), and 11 machine learning classifiers predicted STAS presence and grade. Independent clinic-radiological predictors were identified by univariable and multivariable analyses, and logistic regression combined these with the best radiomics model to build a combined model. Performance was evaluated by area under the curve (AUC) and accuracy (ACC). STAS grade II was independently associated with poorer RFS (P = 0.037), while grade I was not. For STAS detection, the random forest model achieved AUCs of 0.888, 0.886 and 0.779 with accuracies of 81.5%, 85.2%, and 69.6%; the combined model performed similarly (AUCs 0.877, 0.894, 0.802; accuracies 77.6%, 85.2%, 72.8%). For STAS grading, logistic regression yielded AUCs of 0.845, 0.852, and 0.798 (accuracies 78.2%, 70.4%, 77.1%), with the combined model comparable (AUCs 0.845, 0.855, 0.796; accuracies 78.2%, 70.4%, 77.1%). STAS grading improves prognostic stratification in LUAD. CT-based radiomics models allow noninvasive prediction of STAS grade, aiding surgical decisions and prognosis.