Using Artificial Intelligence to Predict Cirrhosis From Computed Tomography Scans.

Area of Science:

• Radiology and Medical Imaging
• Hepatology
• Artificial Intelligence in Medicine

Background:

• Undiagnosed cirrhosis is a significant clinical challenge.
• Accurate prediction of cirrhosis is crucial for patient management and treatment.
• Existing diagnostic methods may have limitations in certain patient populations.

Purpose of the Study:

• To develop and validate an automated liver segmentation tool for cirrhosis prediction.
• To assess the performance of imaging features extracted from CT scans.
• To evaluate the combination of imaging features with clinical data for improved diagnostic accuracy.

Main Methods:

• Trained an automated liver segmentation model (3D-U-Net, DeeplLabv3+) on 1,590 CT scans.
• Extracted imaging features from an external cohort of 351 patients with paired CT and liver biopsy data.
• Developed multivariate models using gradient boosting decision trees to predict histologic cirrhosis.
• Evaluated model performance using 5-fold cross-validation and area under the receiving operating characteristic (aUC).

Main Results:

• The combination of liver morphomics with laboratory and demographic data achieved an aUC of 0.85 (0.81-0.90).
• This combined model significantly outperformed FIB-4 alone (aUC 0.76) (P < 0.001).
• Liver morphomics alone showed comparable performance to FIB-4 (aUC 0.71 vs 0.76).

Conclusions:

• Automatically extracted CT features combined with electronic medical record data improve cirrhosis prediction.
• This tool shows potential for detecting undiagnosed cirrhosis in both pre- and post-transplant patients.
• This proof-of-principle study highlights a promising approach for liver disease management.