Body of the Abstract

Introduction
Liver Imaging Reporting and Data System (LI-RADS) was developed to standardize diagnosis and reporting for liver observations in patients at-risk for hepatocellular carcinoma (HCC). Despite accepted as a reliable diagnostic system with over 90% specificity¹, the application of LI-RADS is hampered by substantial complexity² and suboptimal diagnostic sensitivity^1,3. Previous works have explored possible solutions to address these issues^4-9. However, many of them were based on investigator experiences, and limited evidences are available regarding the performances of the many available combinations of imaging features. In addition, although there are strong data supporting the LI-RADS major features as the main determinants of HCC diagnosis^1,10,11, the assignment of a lesion to a given LI-RADS risk stratum (LR-3 to LR-5) has not been rigorously explored on large patient samples.
Therefore, in a data-driven manner, we aimed to assess the risk of HCC in lesions with a variety of combinations of LI-RADS major features and propose revisions to the LI-RADS diagnostic table (rLI-RADS) on gadoxetate disodium (EOB)-enhanced MRI to improve simplicity and sensitivity while maintaining high specificity.
Methods
Consecutive at-risk patients with LR-3 to LR-5 observations were analyzed retrospectively from a prospective observational registry collected between July 2015 to September 2018 [Clinical trial registration No: ChiCTR1900026668].
In the training set, LI-RADS version 2018 (v2018) major features evaluated by three independent abdominal radiologists were used to develop rLI-RADS. Specifically, the LI-RADS v2018 CT/MRI diagnostic table was first converted into a decision tree, and the proportion of HCC (%HCC, equivalent to positive predictive value) of each leaf node was computed. Based on the associated %HCC, the rLI-RADS categories (rLR) were then defined as the following: 1) leaf nodes with %HCC ≥ 90% were categorized as rLR-5; 2) those with %HCC < 20% as rLR-3; 3) and those with %HCC between 20% and 90% as rLR-4. On a per-lesion basis, a composite reference standard according to either pathology or imaging was established.
In an independent testing set, diagnostic performances were computed using a generalized estimating equation model to account for clustering effects due to reader and observation multi-collinearity and compared to the LI-RADS v2018 with McNemar’s test.
Results
224 consecutive patients (180 males; median age, 51 years) were included (Fig. 1). Among them, 196 (88%) patients had chronic HBV infection, and 92 (41%) had established cirrhosis. Of the 742 included liver observations (median size, 13 mm; range, 3–148 mm), 498 (67%) were HCCs, 26 (4%) were non-HCC malignancies, and 218 (29%) were benign lesions.
In the training set, the rLI-RADS diagnostic table was developed with 6 cells, compared to 16 cells for LI-RADS v2018 (Fig. 2C). The %HCC for categories LR-3, LR-4 and LR-5 based on LI-RADS v2018 was 33%, 88%, and 95% in the training set (Fig. 2A) and 27%, 79%, and 94% in the testing set (Fig. 2B), respectively. These measures based on rLI-RADS was 12%, 51%, and 94% in the training set (Fig. 2A) and was 55%, 25%, and 93% in the testing set (Fig. 2B), respectively. The training and testing set weighted kappa value for LI-RADS v2018 categories were 0.807 and 0.798, respectively, and were 0.782 and 0.823 for rLI-RADS, respectively.
In the testing set, rLI-RADS demonstrated a significantly higher sensitivity than LI-RADS v2018 for all observations as a whole (76% vs. 61%, p<0.001), those <10mm (40% vs. 0%, p<0.001), and those between 10-19mm (78% vs. 62%, p<0.001). Likewise, rLI-RADS also yielded a significantly higher accuracy for all observations as a whole (82% vs. 74%, p<0.004), those <10mm (74% vs. 60%, p<0.001), and those between 10-19mm (85% vs. 77%, p<0.001). However, despite with comparable specificities for individual size subgroups (p = 0.125-0.500), the specificity of rLI-RADS was slightly lower than LI-RADS v2018 for all observations as a whole (91% vs. 94%, p=0.008).
Discussions
In a large prospectively-collected cohort, we modified the LI-RADS v2018 diagnostic table while maintaining LI-RADS v2018 feature definitions and overall algorithm design to develop an rLI-RADS system on EOB-MRI. Compared with LI-RADS v2018, rLI-RADS was completely data-driven, remarkably simpler, and demonstrated significantly optimized diagnostic sensitivity and accuracy while maintaining reasonable specificity for HCC.
In this study, we optimized the diagnostic performances by dictating the most appropriate feature combinations of LI-RADS v2018 major features from hard data. Furthermore, system simplicity was also improved driven primarily by the exclusion of lesion size as a criterion from the diagnostic table. Although representing a substantial discordance with LI-RADS v2018¹² and other diagnostic algorithms^13-16, one possible explanation is that the visibility of either washout and/or capsule at imaging may reflect both the underlying physiology and lesion size. Therefore, the associated risk can be captured in the visualization of these findings without requiring size as a separate feature^17,18.
Conclusion
Developed based on hard data, rLI-RADS demonstrated superior simplicity, sensitivity and overall accuracy for HCC than v2018 LI-RADS without substantial loss of specificity. Thus, it may be considered the preferred major feature diagnostic system over the original LI-RADS, particularly for early detection purpose.

Acknowledgements

None.