Inter‐Reader Agreement for
CT
and
MRI LI
‐
RADS
Version 2018: An Individual Participant Data Meta
S. Das, E. Lam, O. Muhn, S. Mahbub, D. Ansari, H. Osman, R. Awad, M. K. Al‐Ghita, R. Thornhill, C. Walsh, M. M. G. Leeflang, H. Naringrekar, C. B. van der Pol, M. Alabousi, M. R. Bashir, A. F. Costa, B. Levis, J. Salameh, H. Kang, C. Clarke, D. R. Ludwig, M. D. Burgio, S. Lewis, T. J. Fraum, M. D. F. McInnes ABSTRACT
Background
The Liver Imaging Reporting and Data System (LI‐RADS) is a classification system to diagnose hepatocellular carcinoma (HCC) noninvasively. The reported LI‐RADS inter‐reader variability in single‐center studies highlights the need to better understand the variability on a broader scale.
Purpose
Evaluate inter‐reader agreement for LI‐RADS CT/MRI v2018 categorization and imaging features using individual participant data (IPD) meta‐analysis for MRI and CT separately.
Study Type
Retrospective meta‐analysis.
Population
805 patients (542 male, mean age 60.8 ± 11.6 [SD] years) at risk of HCC.
Field Strength/Sequence
1.5 and 3 T, Sequences varied across institutions but followed standard LI‐RADS protocols.
Assessment
Studies from the LI‐RADS Living IPD Meta‐Analysis Project were used. Inclusion criteria: observations evaluated by > 1 reader using LI‐RADS CT/MRI v2018. Risk of bias was assessed for the index test domain using QUADAS‐2; sensitivity analysis was conducted excluding high‐risk observations. The developed protocol is available on the Open Science Framework (
Statistical Tests
A two‐step meta‐analytic approach was used to determine pooled kappa ( κ ) with heterogeneity assessed using τ 2 and I 2 . A post hoc univariable meta regression was conducted using the following variables: number of readers, reader experience, feature prevalence, and study country. Statistical significance was defined as two‐tailed p < 0.05.
Results
Five studies (three CT and MRI, two MRI) comprising 975 observations were included. Four studies had two readers, and one had three, most readers were specialized abdominal radiologists (3–33 years' experience). Inter‐reader agreement ( κ ) for major features was 0.21–0.74 on MRI and 0.11–0.75 on CT. For Ancillary features, κ was −0.02 to 0.68 for MRI and −0.02 to 0.58 on CT, and LR‐M agreement was −0.07 to 0.58 on MRI and −0.07 to 0.66 on CT. LI‐RADS categorization showed agreement of κ = 0.76 for MRI and κ = 0.50 for CT.
Data Conclusion
MRI consistently reported higher agreement than CT for shared features, but overall inter‐reader agreement was variable.
Evidence Level
3.
Technical Efficacy
Stage 2: diagnostic accuracy, considered by most as clinical efficacy.