Prediction of recurrence after resection in hepatocellular carcinoma via whole liver deep learning on preoperative contrast-enhanced CT.
Authors
Affiliations (9)
Affiliations (9)
- Research Center for Data Hub and Security, Zhejiang Lab, Hangzhou, China.
- Department of Radiology, Zhejiang Cancer Hospital, Hangzhou, China.
- Engineering Research Center of EMR and Intelligent Expert System, Ministry of Education, College of Biomedical Engineering and Instrument Science, Zhejiang University, Hangzhou, China.
- Department of Hepatobiliary and Pancreatic Surgery, the First Affiliated Hospital, Zhejiang University School of Medicine, Hangzhou, China.
- MOE Joint International Research Laboratory of Pancreatic Diseases, Zhejiang Provincial Key Laboratory of Pancreatic Disease, the First Affiliated Hospital, Zhejiang University School of Medicine, Zhejiang University Cancer Center, Hangzhou, China.
- Zhejiang Clinical Research Center of Hepatobiliary and Pancreatic Diseases, Hangzhou, China.
- Department of Computer Science, Rutgers University, Piscataway, USA. [email protected].
- Research Center for Data Hub and Security, Zhejiang Lab, Hangzhou, China. [email protected].
- Engineering Research Center of EMR and Intelligent Expert System, Ministry of Education, College of Biomedical Engineering and Instrument Science, Zhejiang University, Hangzhou, China. [email protected].
Abstract
This study aimed to develop a fully automated survival prediction (FASP) system that analyzes whole-liver regions from preoperative contrast-enhanced CT scans for predicting recurrence-free survival (RFS) after curative resection in Hepatocellular carcinoma (HCC). FASP comprised three consecutive components: automatic liver and tumor segmentation models, and a RFS prediction model, LiverSurv, all based on deep convolutional neural networks. FASP was compared against a clinical model leveraging clinical factors and three tumor-based methods using semantic, radiomic, or deep learning features. A total of 827 patients were included across the development, internal test, and external test sets. In the internal and external test sets, FASP achieved concordance indices (C-indices) of 0.646 (95% CI: 0.566, 0.725) and 0.786 (95% CI: 0.726, 0.846), respectively, outperforming the clinical model (both adjusted P<.05). Integrating clinical factors improved C-indices to 0.664 (95% CI: 0.591, 0.736) internally and 0.800 (95% CI: 0.750, 0.849) externally. FASP also surpassed the tumor-based models, which yielded C-indices ranging from 0.623 to 0.632 internally and 0.523 to 0.775 externally. Visualization analysis demonstrated that FASP captured prognostic information from both tumor and background liver regions. These findings suggest whole-liver-based deep learning provides a promising non-invasive approach to predict recurrence risk for HCC patients before surgery.