Simplifying
CT
/
MRI LI
‐
RADS
v2018 Maintains Diagnostic Performance: An Individual Participant
E. Lam, M. M. G. Leeflang, R. G. Adamo, C. B. van der Pol, M. Alabousi, J. P. Salameh, H. Naringrekar, M. R. Bashir, A. F. Costa, H. Osman, D. Ansari, R. Awad, M. K. Al‐Ghita, B. Levis, A. Polikoff, A. Furlan, A. Tang, A. S. Kierans, A. G. Singal, A. Arvind, A. Alhasan, B. C. Allen, B. Kim, C. S. Reiner, C. H. Lee, C. Clarke, C. Yang, D. R. Ludwig, D. Stocker, F. Diaz Telli, F. Piñero, G. Rosiak, H. Jiang, H. Kwon, H. Wei, H. J. Kang, H. Rhee, I. Joo, J. A. Hwang, J. H. Min, J. S. Song, J. Wang, J. Y. Choi, J. Podgórska, J. R. Eisenbrey, K. Sofue, K. Yasaka, K. Bartnik, L. Centonze, A. Vanzulli, L. D. Chen, M. Dioguardi Burgio, M. Ronot, M. Cerny, M. J. Kim, N. Seo, S. Rao, R. Cannella, S. H. Choi, T. J. Fraum, W. Wang, W. K. Jeong, X. Jing, Y. Y. Kim, M. D. F. McInnes ABSTRACT
Background
The Liver Imaging Reporting and Data System (LI‐RADS) diagnostic algorithm assesses hepatocellular carcinoma (HCC) probability of liver observations in high‐risk individuals, but complexity may limit its adoption.
Purpose
To evaluate whether LI‐RADS v2018 can be simplified without impairing diagnostic performance by removing threshold growth (TG) and reducing size cut‐offs.
Study Type
Retrospective meta‐analysis.
Population
Eight thousand five hundred eighty‐two patients (mean age: 59 ± 11 years; 6610 male) at risk for HCC with CT/MRI (986 CT, 6554 MRI, and 1042 CT/MRI).
Field Strength/Sequence
Field strength: 1.5T and 3.0T; sequences varied by institution per LI‐RADS technical requirements.
Assessment
LI‐RADS v2018 using major features only was compared with seven simplifications removing TG and/or applying single size cut‐offs (≥ 10, ≥ 15, and ≥ 20 mm) in individual participant data (IPD) of 64 studies (29 institutions, 11 countries) from the LI‐RADS Living Systematic Review and IPD Meta‐analysis Project. Risk of bias was assessed with QUADAS‐2.
Statistical Tests
Summary receiver operating characteristic curves estimated area under the curve (AUC) across LR‐3 to LR‐5; LR‐5 sensitivity, specificity, and positive predictive value (PPV) were estimated using a one‐step bivariate mixed‐effects model. Differences were assessed by cluster bootstrap. p < 0.05 indicated significance, uncorrected for multiple comparisons.
Results
Nine thousand four hundred six observations were included. Excluding TG with a 10‐mm cut‐off showed AUC similar to LI‐RADS v2018 (0.80 vs. 0.81, p = 0.44) and significantly higher LR‐5 sensitivity (73.8% vs. 71.8%) with similar specificity (72.5% vs. 73.7%, p = 0.13) and PPV (97.8% vs. 97.8%, p = 0.73). Removing TG with 15‐mm and 20‐mm cut‐offs had significantly lower AUC than v2018 (−0.03 [95% CI: −0.041, −0.019]; −0.043 [−0.054, −0.032], respectively) and sensitivity (−13.2% [95% CI: −14.5%, −12.0%]; −30.1% [−31.7%, −28.5%], respectively). Of the 64 studies, 17 had low‐risk of bias, and 56 had low applicability concerns.
Conclusion
Simplifying LI‐RADS v2018 by removing TG and using a 10‐mm size cut‐off preserved diagnostic table AUC while improving LR‐5 sensitivity without loss of specificity or PPV.
Protocol Registration
Open Science Framework (
Evidence Level
3.
Technical Efficiency
Stage 2.