Assessing Pancreatic Fat and Its Correlation with Liver Fat in Suspected MASLD Patients Using Advanced Deep Learning Techniques from MRI Images

Abstract

Pancreatic steatosis and metabolic-dysfunction-associated steatotic liver disease are characterised by fat accumulation in abdominal organs, but their correlation remains inconclusive. Recently proposed deep learning (DL) for proton density fat fraction (PDFF) estimation, which quantifies organ fat, has primarily been assessed for quantifying liver fat. This study aims to validate DL models for pancreatic PDFF quantification and compare pancreas and liver fat content. We evaluated three DL models—Non-Linear Variables Neural Network (NLV-Net), U-Net, and Multi-Decoder Water-Fat separation Network—against a reference PDFF measured using a graph-cut-based method. NLV-Net showed a strong correlation (Spearman rho) with the reference PDFF in the six-echo pancreatic head (slope: 1.02, rho: 0.95) and body (slope: 1.04, rho: 0.94) and a moderate correlation in the three-echo pancreatic head (slope: 0.44, rho: 0.40) and body (slope: 0.49, rho: 0.34). Weak correlations were found between liver and pancreatic body PDFF using graph cut in six-echo (slope: −0.041, rho: −0.12) and three-echo images (slope: 0.0014, rho: 0.073) and using NLV-Net in six-echo (slope: −0.053, rho: −0.12) and three-echo images (slope: −0.014, rho: −0.033). In conclusion, NLV-Net showed the best agreement with the reference for pancreatic fat quantification, and no correlation was found between liver and pancreas fat.

Keywords

• deep learning models
• proton density fat fraction
• chemical-shift-encoded MRI
• pancreatic steatosis
• metabolic-dysfunction-associated steatotic fatty disease