Clinical Validation of AI‑Based Automated 3D Liver Segmentation for Whole‑Liver Fat Fraction Quantification Using Mdixon-Quant MRI
Abstract
Purpose
Quantitative MRI is a reliable, non‑invasive method for assessing liver fat content. Because manual whole‑liver ROI delineation is impractical in clinical practice, fat fraction (FF) is typically estimated from a single‑slice manual ROI (carefully avoiding non liver tissue). In this study, we applied a published AI model to automatically segment 3D whole‑liver for FF calculation and compared the results with a manual ROI‑based approach.
Methods
Forty‑five patients referred for liver fibrosis assessment(1.5T) were included. A retrospective sub‑analysis of mDIXON‑Quant–derived FF was performed. IRB approved the waiver of individual consent. mDixon-Quant: Philips 3D six-points axial (whole liver covered) gradient-echo, TR/TE1/ΔTE=5.3ms/0.9ms/0.7ms, Voxel size was 2.5x2.5x6.0mm3. Data analysis: 3D mDIXON‑Quant derived water and FF images were exported for offline analysis. A 3D whole‑liver mask was automatically segmented from the water images using a published open‑source AI model (https://doi.org/10.1148/ryai.240777). Liver FF was computed using a Histogram‑Gaussian method with three ROI approaches: manual single‑slice ROI excluding non‑hepatic tissue (FFman), 2D ROI extracted from the 3D mask at the same slice (FF2D), and full 3D mask–based ROI (FF3D). The liver FF (FFman, FF2D, and FF3D) for each subject calculated from the three ROI methods were processed using custom-build MATLAB script. Bland-Altman method was used to assess the agreement between FFman and FF2D, FF3D.
Results
The measured liver FF values range at 1.7%-42.4% (Fourteen had normal liver FF and 31 had elevated FF). The Bland-Altman analysis demonstrated strong agreement: FFman vs FF2D and FFman vs FF3D with negligible bias and narrow limits of agreement (bias±1.0SD: -0.1±0.2%, -0.1±0.4% respectively).
Conclusion
Automated 3D whole‑liver FF estimation using this AI‑based segmentation model is feasible and demonstrates strong agreement with manual ROI analysis. By removing the need for expert‑prescribed ROIs and enabling larger liver volume assessment, this approach has the potential to increase diagnostic confidence in hepatic steatosis evaluation.