Integrated radiomics-deep learning modeling of hepatic vasculature for enhanced noninvasive prediction of portal hypertension and high-risk gastroesophageal varices.
Authors
Affiliations (9)
Affiliations (9)
- Shanghai Institute of Medical Imaging, Shanghai, China.
- Department of Radiology, Zhongshan Hospital, Fudan University, Shanghai, China.
- Institute of Science and Technology for Brain-Inspired Intelligence, Fudan University, Shanghai, China.
- Key Laboratory of Computational Neuroscience and Brain-Inspired Intelligence (Fudan University), Ministry of Education, Shanghai, China.
- Human Phenome Institute and Shanghai Pudong Hospital, Fudan University, Shanghai, China.
- Institute of Science and Technology for Brain-Inspired Intelligence, Fudan University, Shanghai, China. [email protected].
- Key Laboratory of Computational Neuroscience and Brain-Inspired Intelligence (Fudan University), Ministry of Education, Shanghai, China. [email protected].
- Shanghai Institute of Medical Imaging, Shanghai, China. [email protected].
- Department of Radiology, Zhongshan Hospital, Fudan University, Shanghai, China. [email protected].
Abstract
To evaluate whether CT-derived radiomics and deep learning features can improve non-invasive risk stratification of clinically significant portal hypertension (CSPH) and high-risk gastroesophageal varices (HRVs) in cirrhotic patients. This retrospective study included 454 cirrhotic patients from center A and 77 from center B. CSPH labels were available only for patients from center A. Patients in both centers with available endoscopic findings were classified as HRVs or non-HRVs. A dual-path modeling framework integrating radiomics and deep learning-derived features extracted from multiple major hepatic vessels was developed, and was compared with single-path (radiomics-only and deep learning-only), single-vessel, vascular geometric, and clinical-feature-based models. Model performance was assessed using the area under the receiver operating characteristic curve (AUC). For CSPH prediction, the dual-path model achieved the highest AUC (0.892 vs. 0.678-0.870 for the other models; P < 0.05 for most comparisons). For HRVs prediction, the proposed model achieved AUCs of 0.790 and 0.782 in the internal and external test set, respectively. The proposed model consistently outperformed the single-path models (0.625-0.648; all P < 0.05), single-vessel models (0.526-0.642; all P < 0.05), the vascular geometric model (0.528; P < 0.01), and the clinical-feature-based models (0.327-0.728; P < 0.001 for the APRI-based model) in the external test set. A dual-path model integrating multi-vessel radiomics and deep learning features demonstrates robust predictive performance for CSPH and HRVs in cirrhotic patients. This approach may provide a new non-invasive alternative for risk stratification and clinical management of portal hypertension-related complications.