Automated Computed Tomography–based Liver Steatosis Risk Stratification of Deceased Organ Donors Using Real-world Data

Background. Macrovesicular steatosis in liver transplant donors is associated with increased risk of poor outcomes. Prerecovery computed tomography (CT) is widely available but has substantial variability and requires expert interpretation. Automating CT-based steatosis assessment could streamline evaluation. Methods. We analyzed CT scans from a single organ procurement organization between 2019 and 2024 with biopsy reports. The primary outcome was macrovesicular steatosis ≥30%. The liver and spleen were automatically segmented. Predictive approaches included (1) clinical features only; univariable models using (2) liver attenuation, (3) liver-to-spleen ratio, and (4) liver-spleen difference; (5) radiomics (1130 features) with extreme gradient boosting; (6) a 2-step method combining liver-spleen difference with radiomics; and (7) a 3-dimensional convolutional neural network. Performance was evaluated with repeated stratified train/validation/test splits for >100 random seeds, using area under the receiver operating characteristic (AUROC) as the primary metric. Results. A total of 147 CTs were included; 25.9% had ≥30% biopsy-proven macrovesicular steatosis, 11.6% were contrast-enhanced only, and 40.8% showed at least grade 1 fibrosis. Model performance ranged from an AUROC of 0.65 (interquartile range [IQR], 0.57–0.72) for the clinical-only model to 0.87 (IQR, 0.81–0.92) for the 2-step approach ( P = 0.028 relative to the clinical model). Univariable models achieved AUROCs of 0.81 (IQR, 0.78–0.82) for liver attenuation, 0.84 (IQR, 0.77–0.89) for liver-to-spleen ratio, and 0.83 (IQR, 0.78–0.90) for liver-spleen difference. Conclusions. An automated 2-step approach using radiomic features achieved strong performance in predicting donor macrovesicular steatosis. Such models could help support decision-making and expedite evaluation in the donor-offer process.