Deep Learning‐Accelerated
MRI
Assessment of Hepatic Proton Density Fat Fraction: Agreement With Conventional Acquisition Across Manual and Automated Measurement Approaches
Seo Yeon Youn, Bohyun Kim, Hyun‐Soo Lee, Jungwoo Lee, Seohyun Choi, Dominik Nickel, Sabine Mollus, Stephan Kannengiesser, Joon‐Il Choi, Yu Ri Shin, Soon Nam Oh, Sung Eun Rha ABSTRACT
Background
Deep learning reconstruction can shorten breath‐hold MRI for liver proton density fat fraction (DL‐PDFF), but agreement with conventional PDFF (Conv‐PDFF) and the impact of measurement approach (regional vs. whole‐liver segmentation) remain unclear.
Purpose
To evaluate linearity and agreement of DL‐PDFF with Conv‐PDFF and MR spectroscopy PDFF (MRS‐PDFF), and reconstruction‐dependent whole‐liver distribution metrics.
Study Type
Retrospective.
Population
Fat phantom three vials per nominal fat mass fraction (27.86%, 18.32%, and 9.12%); 96 adults (53 males; median age, 64).
Field Strength/Sequence
3 T; Chemical shift‐encoded multi‐echo gradient‐echo; acquisition time, 10 s (DL‐PDFF; acceleration factor [AF], 6) and 14 s (Conv‐PDFF; AF, 4).
Assessment
Central region of interest (ROI) for phantom PDFF; In vivo, single ROI by radiology technologists, multiple ROIs by a resident and a medical student, and automated whole‐liver segmentation, yielding voxel‐based histograms and residual fitting error; inter‐observer agreement.
Statistical Tests
Wilcoxon signed‐rank tests, linear regression (slope/intercept/ R 2 ), Bland–Altman bias and 95% limits of agreement (LoA) with proportional‐bias slope ( β ), Lin's concordance correlation coefficient (CCC), Hodges–Lehmann method, and Spearman ( ρ ). Two‐sided p < 0.05.
Results
DL‐PDFF versus Conv‐PDFF showed excellent phantom linearity ( R 2 = 1.00). In patients, within the same approach: single ROI (bias, −0.24%; widest LoA, −2.76 to 2.28; CCC, 0.94); multiple ROIs (bias, −0.002; LoA, −1.04 to 1.04; CCC, 0.99); and whole‐liver segmentation (bias, 0.12; LoA, −0.54 to 0.78; CCC, 0.99). Cross‐method comparison (DL‐PDFF segmentation vs. Conv‐PDFF multiple ROIs) showed larger bias (0.57; LoA, −0.77 to 1.91; CCC, 0.97). Against MRS‐PDFF, bias was similar (−0.05 to −0.02; LoA, −3.00 to 2.96). For DL‐PDFF, whole‐liver SD (Hodges–Lehmann difference, 0.55%; ρ = 0.95) and residual fitting error significantly higher (median, 2.80 vs. 1.92). Inter‐observer agreement was high (CCC, 0.99).
Data Conclusion
DL‐PDFF demonstrated high agreement with Conv‐PDFF while reducing acquisition time. However, measurement approach (segmentation vs. ROI) contributes larger systematic differences relevant to longitudinal and cross‐study comparisons.
Evidence Level
3.
Stage of Technical Efficacy
2.