Preoperative prediction of lymphovascular and perineural invasion in locally advanced gastric cancer via CT habitat analysis and deep learning: A dual-center study.
Authors
Affiliations (4)
Affiliations (4)
- Department of Radiology, The First Affiliated Hospital of Zhengzhou University, Zhengzhou 450052, Henan, China; Henan Key Laboratory of Image Diagnosis and Treatment for Digestive System Tumor, The First Affiliated Hospital of Zhengzhou University, Zhengzhou 450052, Henan, China.
- Department of Blood Transfusion, The Fifth Affiliated Hospital of Zhengzhou University, Zhengzhou 450052, Henan, China.
- Department of Radiology, The Fifth Affiliated Hospital of Zhengzhou University, Zhengzhou 450052, Henan Province, China.
- Department of Medical Information, The First Affiliated Hospital of Zhengzhou University, Zhengzhou 450052, Henan, China; Institute of Interconnected Intelligent Health Management of Henan Province, The First Affiliated Hospital of Zhengzhou University, Zhengzhou 450052, Henan, China. Electronic address: [email protected].
Abstract
To develop and validate a combined model integrating CT-based habitat analysis and deep learning (DL) features for the preoperative prediction of lymphovascular and perineural invasion (LVI/PNI) in locally advanced gastric cancer (LAGC). This dual-center, retrospective study included 611 LAGC patients divided into training (n = 374), internal validation (n = 161), and external validation (n = 76) cohorts. Radiomics and DL features were extracted from venous-phase CT images of the whole tumor to identify the optimal intratumoral signature. Additionally, K-means clustering was utilized to partition the tumors into distinct habitat subregions, from which DL features were extracted to select the optimal habitat signature. A combined model integrating the best intratumoral and habitat signatures with significant clinical factors was constructed via multivariate logistic regression and visualized as a nomogram. Performance was assessed using the area under the curve (AUC) and decision curve analysis (DCA). The DL signature and the whole tumor (WT) habitat signature demonstrated superior discriminatory ability. Consequently, the combined model-integrating the DL signature, habitat WT signature, cT stage, and histological grade-achieved favorable predictive efficacy, yielding numerical improvements in AUC performance across all cohorts. AUC values were 0.891, 0.866, and 0.815 for the training, internal validation, and external validation cohorts, respectively. DCA confirmed that the model provided substantial net clinical benefit. The proposed combined model provides a promising and clinically relevant tool for preoperative LVI/PNI risk assessment in LAGC, facilitating individualized treatment strategies.