Liver MRI proton density fat fraction inference from contrast enhanced CT images using deep learning: A...

Area of Science:

• Radiology
• Artificial Intelligence
• Hepatology

Background:

• Metabolic dysfunction-associated steatotic liver disease (MASLD) is a prevalent global health issue.
• Accurate, noninvasive liver fat quantification is crucial for early diagnosis and management.
• Current gold standard, MRI Proton Density Fat Fraction (PDFF), faces accessibility and cost limitations.

Purpose of the Study:

• To explore the feasibility of inferring liver MRI-PDFF values from contrast-enhanced computed tomography (CECT) using deep learning.
• To develop and validate a deep neural network (DNN) for estimating liver fat content from CECT images.
• To assess the accuracy of DNN-predicted PDFF against MRI-PDFF ground truth.

Main Methods:

• Retrospective analysis of 94 living liver donor candidates with concurrent CECT and MRI-PDFF.
• A 3D U-Net deep neural network was trained using CECT as input and MRI-PDFF as output.
• Performance evaluated using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Kappa statistics for agreement.

Main Results:

• The DNN model achieved an MAE of 2.94 and RMSE of 4.27 when comparing predicted to reference MRI-PDFF medians.
• Substantial agreement (kappa = 0.75) was found between DNN-predicted and ground truth categorical steatosis grades.
• The model demonstrated robust categorical classification of lower-grade steatosis, outperforming prior methods.

Conclusions:

• Deep learning shows promise in estimating liver fat from CECT, offering a more accessible alternative to MRI-PDFF.
• While exact PDFF values were limited, categorical fat grade classification was reliable, especially for lower grades.
• This approach could potentially improve early MASLD diagnosis and patient care by leveraging readily available CECT data.