Multiparametric Deep and Radiomic MRI Features for Liver Stiffness Classification in Children and Adults with Chronic Liver Disease

Motivation: To improve access to liver stiffness assessment by offering an alternative to magnetic resonance elastography (MRE), which has limited availability in many geographic regions. Goal(s): Develop a deep learning method to classify liver stiffness as no/mild (<3 kPa) vs moderate/severe (≥3 kPa) using MRI and electronic health record (EHR) data. Approach: We used MRSegmentator for segmentation, the Swin-Transformer model and PyRadiomics for feature extraction, and combined these features with EHR data for classification. Our model was validated through internal and external experiments. Results: Our model achieved an AUROC of 0.88 and 0.90 during internal and external validation, respectively. Impact: Our model offers an alternative to conventional MR elastography, potentially expanding access and improving care for patients with chronic liver disease.