Virtual Magnetic Resonance Elastography Using a Deep Generative Model for Liver Fibrosis Staging

ABSTRACT Background Liver biopsy is invasive, which presents many limitations in clinical settings. Magnetic resonance elastography (MRE) has significant value in non‐invasively diagnosing liver fibrosis. However, its use is currently uncommon because it requires specialized equipment. This study aimed to propose and assess the reliability of virtual MRE (vMRE) using diffusion weighted imaging (DWI) and evaluate its effectiveness in diagnosing liver fibrosis. Methods We proposed a Registration‐based Generative Adversarial Network‐convolutional Block Attention Model (RegGAN‐CBAM) to synthesize stiffness (cMap) and viscosity (phiMap) using DWI data acquired from 128 patients diagnosed with liver fibrosis or cirrhosis. Correlation and agreement between native MRE (nMRE)‐ and vMRE‐derived measurements were assessed using Spearman correlation coefficients and Bland–Altman analysis. Receiver operating characteristic curves were constructed to evaluate the diagnostic performance of these measures for cirrhosis. Results The proposed RegGAN‐CBAM model demonstrated favorable performance in image synthesis and estimation. vMRE measures had high consistency with nMRE for images; moreover, they correlated significantly with cMap ( r = 0.77) and phiMap ( r = 0.58) measurements. Staging based on predicted cMap and phiMap demonstrated excellent performance ( p < 0.01), with considerable accuracy for diagnosing cirrhosis (area under the curve: cMap = 0.75 and phiMap = 0.74) among the test set ( n = 40), which included 21 patients with histologically confirmed cirrhosis (52.5%). Conclusions Our study highlights the reliability of our proposed model for liver fibrosis diagnosis. Furthermore, the non‐invasive approach may serve as a practical alternative to conventional clinical MRE, particularly in healthcare facilities without access to MRE equipment.