GAST-NET: A multi-modal and multi-task deep learning framework for preoperative prediction of perineural invasion and prognostic risk in gastric cancer.
Authors
Affiliations (5)
Affiliations (5)
- School of Computer Science and Technology, Harbin University of Science and Technology, Harbin, China.
- Department of Interventional Medicine, the First Affiliated Hospital, Harbin Medical University, Harbin, China.
- Department of General Practice, the Second Affiliated Hospital, Harbin Medical University, Harbin, China.
- Computer Science School of Computing, Engineering & Intelligent Systems Ulster University, NI, UK.
- Department of Internal Medicine, Harbin Medical University Cancer Hospital, Harbin Medical University, Harbin, China. Electronic address: [email protected].
Abstract
Preoperative imaging prediction of perineural invasion in gastric cancer (GC-PNI) mainly relies on tumour characteristics and clinical variables, while the potential of non-tumour-derived multimodal features remains underexplored. We retrospectively enrolled 777 patients from three medical centers and divided them into a training cohort, an internal testing cohort (I-T), and two external testing cohorts (E-T1, E-T2). We developed an end-to-end multimodal and multitask deep learning framework, termed GAST-NET, that integrates tumour CT features, visceral adipose tissue characteristics, and clinical variables to jointly predict perineural invasion (PNI) and five-year prognostic survival risk (PR). The model incorporates an Adaptive Multi-scale Feature Fusion Module (AMFM) and a Cross-Scale Fusion Pooling (CSF Pooling) module to capture hierarchical semantic information and enhance discriminative cross-modal representation. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), and decision curve analysis (DCA). Furthermore, five radiologists were invited to participate in the image reading experiment to verify the clinical interpretability and diagnostic gain of the model. The proposed model achieved AUCs of 0.923 (95% CI: 0.865-0.969), 0.868 (95% CI: 0.791-0.934), and 0.871 (95% CI: 0.806-0.930) for PNI prediction across the internal and two external cohorts, respectively. For prognostic risk prediction, the AUC reached 0.873 (95% CI: 0.835-0.922). When used as a decision-support tool, GAST-NET significantly improved diagnostic accuracy and reduced misclassification compared with radiologists. GAST-NET demonstrated strong generalizability and potential clinical utility in predicting perineural invasion (PNI) and prognosis in gastric cancer. Notably, visceral adipose tissue features provided complementary value for PNI prediction beyond conventional tumour characteristics.