Mining Whole-liver Information with Deep Learning for Preoperatively Predicting HCC Recurrence-free Survival

Area of Science:

• Hepatocellular Carcinoma (HCC) Research
• Medical Imaging Analysis
• Artificial Intelligence in Oncology

Background:

• Hepatocellular carcinoma (HCC) is a major global cause of cancer mortality.
• Predicting recurrence-free survival (RFS) non-invasively before surgery is crucial but challenging.
• Existing models often overlook the impact of overall liver health, focusing solely on the tumor area.

Purpose of the Study:

• To develop a novel convolutional neural network (CNN) named RFSNet.
• To utilize whole-liver information from contrast-enhanced computed tomography (CECT) for predicting RFS in HCC patients post-resection.
• To compare the performance of whole-liver analysis against tumor-only and clinical feature-based models.

Main Methods:

• A CNN (RFSNet) was designed to process entire liver regions from CECT scans.
• Cox proportional-hazards loss was employed for model training.
• A cohort of 215 HCC patients undergoing hepatic resection was analyzed, with data split into developing and testing subcohorts.

Main Results:

• RFSNet achieved high performance, with concordance indices (CIs) of 0.88 in the developing subcohort and 0.65 in the testing subcohort.
• Incorporating clinical features did not enhance RFSNet's predictive accuracy.
• The whole-liver approach demonstrated superior prognostic value compared to tumor-specific features and clinical indicators.

Conclusions:

• The RFSNet model effectively leverages whole-liver information from CECT for predicting HCC RFS.
• This whole-liver CNN approach offers a more comprehensive and accurate method for prognostication compared to traditional tumor-focused or clinical models.
• RFSNet shows promise for improving non-invasive pre-operative risk stratification in HCC management.