3D Radiomics Analysis Based on Liver CT Imaging for Predicting Hepatic Encephalopathy in HBV-ACLF Patients.

Area of Science:

• Hepatology
• Medical Imaging
• Artificial Intelligence

Background:

• Hepatic encephalopathy (HE) is a significant mortality risk in hepatitis B virus-related acute-on-chronic liver failure (HBV-ACLF).
• Accurate prediction of HE is vital for managing HBV-ACLF patients.

Purpose of the Study:

• To evaluate the efficacy of 3D deep learning using CT imaging for predicting HE in HBV-ACLF patients.
• To compare the performance of 3D deep learning models with traditional radiomics and clinical predictors.

Main Methods:

• A retrospective study involving 222 HBV-ACLF patients from two medical centers.
• Utilized ResNet50 and DenseNet201 architectures for 3D deep learning feature extraction from CT scans.
• Extracted classical radiomics features and analyzed clinical data using multivariate logistic regression and machine learning algorithms (XGBoost).

Main Results:

• The 3D deep learning model, particularly ResNet50, demonstrated strong predictive performance for HE (AUCs ranging from 0.833 to 0.844 across cohorts).
• Classical radiomics, especially with XGBoost, also showed predictive potential.
• Independent clinical predictors for HE included INR and ammonia levels.
• A combined nomogram model integrating 3D deep learning, radiomics, and clinical features achieved the highest predictive accuracy.

Conclusions:

• 3D deep learning is highly effective in predicting HE in HBV-ACLF patients.
• Combining 3D deep learning with radiomics and clinical features further enhances predictive performance, offering a promising tool for clinical decision-making.