We searched our institution’s databases for a clinical cohort under HCC surveillance and identified 2,576 patients who underwent MRI for further diagnostic workup between January 2016 and December 2017. The inclusion criteria were as follows: 1) age of 18 years or older; 2) patients at high risk for HCC with cirrhosis or chronic hepatitis B; 3) presence of at least one and up to five hepatic lesions (each ≥1 cm) on MRI; and 4) no previous treatment for hepatic lesions. The exclusion criteria were as follows: 1) patients with cirrhosis due to congenital hepatic fibrosis and vascular disorders such as hereditary hemorrhagic telangiectasia, Budd-Chiari syndrome, chronic portal vein occlusion, cardiac congestion, or diffuse nodular regenerative hyperplasia (n=17); and 2) patients with insufficient final diagnosis such as unknown final diagnosis of malignancy as a result of immediate locoregional therapy or insufficient follow-up (<2 years) for benign lesions to determine size stability (n=579). We did not include observations such as hepatic cysts, perfusion alteration, hepatic fat deposition or sparing, hypertrophic pseudomass, confluent fibrosis, and focal scar. However, we include hemangiomas in our study. A radiologist (S.K., with 9 years of experience in abdominal radiology) who was not involved in the image analysis for this study finalized the study population and selected hepatic lesions.

MRI was performed using 3.0-T systems (Magnetom Trio Tim, Siemens Healthineers, Erlangen, Germany; Intera Achieva, Ingenia, or Ingenia CX, Philips Healthcare, Best, the Netherlands; and Discovery MR 750w, GE Healthcare, Waukesha, WI, USA). The protocol included acquisition of dual-echo T1-weighted gradientecho images (in-phase and opposed-phase), T1-weighted 3-dimensional gradient-echo images with dynamic contrast enhancement, navigator-triggered single- or multi-shot T2-weighted images, and diffusion-weighted images at b-values of 0 or 50, 400, and 800 s/mm². Dynamic T1-weighted imaging was performed before and after administering one of the two extracellular contrast agents (ECAs) (gadoterate meglumine, Dotarem, Guerbet SA, Aulnay-sous-Bois, France; or gadobutrol, Gadovist, Bayer Pharma AG, Berlin, Germany) or a hepatobiliary agent (HBA) (gadoxetate disodium, Primovist, Bayer Pharma AG): a bolus injection of 0.2 mL/kg of gadoterate meglumine (n=67) at a rate of 2.0 mL/s or 0.1 mL/kg of gadobutrol (n=4) at a rate of 1.0 mL/s, or 0.1 mL/kg gadoxetate disodium (n=202) at a rate of 1.0 mL/s, followed by a 20-mL saline flush using a power injector. The choice of MRI contrast agents (ECA or HBA) was made at the discretion of the physicians. The physicians were fully informed by radiologists regarding the advantages and disadvantages of each contrast material at a multidisciplinary conference in our institution, such as the potential differences in terms of degree of arterial phase (AP) enhancement and relative frequency of artifacts between the two contrast materials, potential benefits of hepatobiliary phase (HBP) images, and cost of contrast materials. AP scanning was initiated using the test-bolus or bolus-tracking technique. Thereafter, portal venous phase (PVP) and delayed phase (DP) images were obtained by ECA-MRI, while PVP, transitional phase (TP), and 20-minute HBP images were obtained by HBA-MRI. The detailed parameters of the MRI sequences are presented in Supplementary Table 1.

A radiologist (S.K., with 9 years of experience in abdominal radiology) who was not involved in the image analysis bookmarked individual lesions, reported the lesion size and location, and provided a list for further review. The prior computed tomography or MRI examination was used to assess threshold or subthreshold growth, and ultrasound (US) was used to assess US visibility as a discrete nodule when available.

Two board-certified abdominal radiologists (M.J.K. and S.L. with 27 and 8 years of experience in liver imaging, respectively) who did not participate in the patient and lesion selection reviewed all images. They were blinded to the final diagnosis of each lesion but were informed that the study population consisted of patients at high risk for HCC. Both LI-RADS 2018 and KLCA-NCC 2018 criteria were independently applied to each lesion. After independent categorization, the inter-reader agreement was evaluated. Discrepancies between the readers were resolved by consensus. Consensus data were used for the main image analysis results.

According to LI-RADS 2018, the readers reviewed the presence or absence of major, ancillary, and LR-M imaging features (Supplementary Table 2) [6]. Each lesion was assigned to one of the following categories: LR-1 for definitely benign, LR-2 for probably benign, LR-3 for intermediate probability of malignancy, LR-4 for probably HCC, LR-5 for definitely HCC, LR-M for probably or definitely malignant but not specific for HCC, and LR-TIV for definite tumor in vein [6]. The LI-RADS 2018 category could be adjusted by applying ancillary features and then tie-breaking rules, if needed [6].

According to KLCA-NCC 2018, the readers assessed the presence or absence of major and ancillary imaging features and targetoid appearance (Supplementary Table 3) [10,11]. Each lesion was categorized as definite HCC, probable HCC, indeterminate, benign, and targetoid appearance. Definite HCC was defined as AP hyperenhancement with washout in the PVP, DP, or HBP [10,11]. In a lesion with some but not all of the major imaging features of HCC, probable HCC was assigned only when the lesion fulfilled at least one item from each of the following two categories of ancillary imaging features: favoring malignancy in general, and favoring HCC in particular [10,11]. The diagnostic criteria for HCC should be applied in a stepwise manner after excluding marked T2 hyperintensity or targetoid appearance [10,11]. When the imaging diagnosis was inconclusive, a lesion was defined as indeterminate [10,11]. Although the KLCA-NCC 2018 guidelines do not clearly define benign and other malignancy categories, we added benign lesions and targetoid appearance to the system in order to match the classification with the LI-RADS 2018. A lesion that fulfilled at least one of the ancillary features favoring benignity according to KLCA-NCC 2018 [10,11] was classified as a benign lesion. Targetoid appearance was evaluated on diffusion-weighted imaging (DWI) or contrast-enhanced sequences.

Definite HCC of KLCA-NCC 2018 corresponds to LR-5 of LI-RADS 2018, and probable HCC of KLCA-NCC 2018 corresponds to LR-4 of LI-RADS 2018. For the noninvasive diagnosis of HCC, definite HCC of KLCA-NCC 2018 was compared with LR-5 of LIR-ADS 2018, and definite or probable HCC (definite/probable HCC) of KLCA-NCC 2018 was compared with LR-5 or LR-4 (LR-5/4) of LI-RADS 2018.

The diagnoses of HCCs and non-HCC malignancies were based on pathology including surgical resection (n=285) or explant for transplantation (n=15). Benign diagnoses were obtained by pathology (n=14) or typical imaging features and stability at imaging for at least 2 years (n=38). The mean interval between the MRI and pathological diagnosis was 20.2 days (range, 0–90 days).

The baseline characteristics of patients and lesions were compared between the ECA-MRI and HBA-MRI groups using the Fisher’s exact test for categorical variables and the two-sample t test or Mann-Whitney U test for continuous variables. The per-lesion diagnostic performances of LI-RADS 2018 and KLCA-NCC 2018 were compared by using the generalized estimating equation method. The inter-reader agreement was evaluated using the Cohen κ coefficient. The κ values were interpreted as follows: poor, 0.00–0.20; fair, 0.21–0.40; moderate, 0.41–0.60; good, 0.61–0.80; and excellent, 0.81–1.00. Statistical analyses were performed using SAS software (version 9.4; SAS Inc., Cary, NC, USA). A P-value of <0.05 was considered statistically significant.

A total of 273 patients (mean age, 57.3 years; 188 men and 85 women) with 352 lesions were selected for the final statistical analysis (Fig. 1). The characteristics of the patients and lesions are presented in Table 1. The study consisted of 71 patients with 86 lesions who underwent ECA-MRI, and 202 patients with 266 lesions who underwent HBA-MRI. No statistically significant differences in sex, age, presence of liver cirrhosis, lesion size, proportions of categories according to LI-RADS 2018 and KLCA-NCC 2018, and final diagnosis were found between ECA-MRI and HBA-MRI groups. Hepatitis B (89.8%) was the predominant etiology of liver disease. Two hundred eighteen (79.9%) patients had one lesion; 37 (13.6%) had two lesions; and 18 (6.5%) had three or more lesions. The 352 lesions included 263 (74.7%) HCCs, 37 (10.5%) non-HCC malignancies, and 52 (14.8%) benign lesions.

The per-lesion diagnostic performances for LR-5 of LI-RADS 2018 and definite HCC of KLCA-NCC 2018 are demonstrated in Table 2.

On ECA-MRI, the sensitivity and specificity of LR-5 of LI-RADS 2018 and definite HCC of KLCA-NCC 2018 were not significantly different (75.8% vs. 69.4%, P=0.095 and 95.8% vs. 95.8%, P>0.999) for the entire cohort. In both subgroups according to the size of lesions (10–19 mm and ≥20 mm), there were no significantly differences in the sensitivities and specificities between LR-5 and definite HCC (all, P>0.05).

On HBA-MRI, the sensitivity was significantly higher for definite HCC of KLCA-NCC 2018 (79.1% vs. 68.2%, P<0.001) than for LR-5 of LI-RADS 2018 without a significant difference in specificity (93.9% vs. 95.4%, P=0.314) for all lesions. When lesions were stratified according to the size (10–19 mm and ≥20 mm), the sensitivities of definite HCC were significantly higher than those of LR-5 (10–19 mm: 75.8% vs. 51.6%, P<0.001; and ≥20 mm: 80.6% vs. 75.5%, P=0.017), but the specificities were not significantly different (all, P>0.05), similar to the results obtained for all lesions (Figs. 2, 3).

There were 27 lesions with threshold growth (24 HCCs and three non-HCC malignancies) according to LI-RADS 2018. By using only threshold growth (one of the major imaging features of LI-RADS 2018) with AP hyperenhancement, two additional HCCs were diagnosed as LR-5.

The per-lesion diagnostic performances of LR-5/4 of LI-RADS 2018 and definite/probable HCC of KLCA-NCC 2018 are summarized in Table 3.

On ECA-MRI, no significant differences in the sensitivity and specificity were observed between LR-5/4 of LI-RADS 2018 and definite/probable HCC of KLCA-NCC 2018 (87.1% vs. 83.9%, P=0.313 and 87.5% vs. 91.7%, P=0.307) for the entire cohort. When lesions were stratified according to the size (10–19 mm and ≥20 mm), the sensitivities and specificities of LR-5/4 and definite/ probable HCC were not significantly different for both subgroups (all, P>0.05).

On HBA-MRI, definite/probable HCC of KLCA-NCC 2018 showed significantly higher specificity (92.3% vs. 80.0%, P=0.003) than LR-5/4 of LI-RADS 2018 for diagnosing HCC without a statistically significant difference in sensitivity (85.6% vs. 88.1%, P=0.057) for all lesions. The specificity of definite/probable HCC for 10–19 mm-sized lesions was significantly higher than that of LR-5/4 (92.9% vs. 73.8%, P=0.002), but the sensitivities for 10–19 mm-sized lesions and ≥20 mm-sized lesions were not significantly different (all, P>0.05).

There were four false-positive diagnoses in LR-5 of LI-RADS 2018 and five false-positive diagnoses in definite HCC of KLCA-NCC 2018. All false-positive results were combined hepatocellular carcinoma-cholangiocarcinoma (cHCC-CCA) (Fig. 4, Supplementary Table 4).

In LR-5/4 of LI-RADS 2018, the false-positive results were eight non-HCC malignancies (six HCC-CCAs, one hepatoblastoma, and one metastasis) and eight benign lesions (two focal nodular hyperplasia-like nodules and six dysplastic nodules). In definite/probable HCC of KLCA-NCC 2018, the false-positive diagnoses were six cHCC-CCAs and one hepatoblastoma (Supplementary Table 5).

Supplementary Table 6 and 7 present the subgroup analysis results in patients with hepatitis B. The results of patients with hepatitis B were similar and in line with those of the entire cohort.

On ECA-MRI, the sensitivity and specificity of LI-RADS 2018 and KLCA-NCC 2018 were not significantly different (LR-5 vs. definite HCC: 75.0% vs. 67.9%, P=0.094 and 93.3% vs. 93.3%, P>0.999; LR-5/4 vs. definite/probable HCC: 87.5% vs. 83.9%, P=0.313 and 86.7% vs. 86.7%, P>0.999).

On HBA-MRI, definite HCC of KLCA-NCC 2018 showed significantly higher sensitivity (79.1% vs. 68.5%, P<0.001) than LR-5 of LI-RADS 2018 without a significant difference in specificity (94.9% vs. 96.6%, P=0.313). Definite/probable HCC of KLCA-NCC 2018 had higher specificity (93.2% vs. 81.4%, P=0.005) than LR-5/4 of LI-RADS 2018. The sensitivity was lower for definite/probable HCC than for LR-5/4 without statistical significance (85.0% vs. 87.7%, P=0.056).

The inter-reader agreement for the categorization of lesions according to LI-RADS 2018 and KLCA-NCC 2018 was excellent (κ=0.90–0.97). The detailed results of inter-reader agreement are provided in Supplementary Table 8.