A total of 120 patients who underwent Gd-EOB-DTPA-enhanced MR imaging using the same 3T scanner between April 2015 and August 2019 were included in the study. We excluded patients with poor breath holding or prior liver resection, and post-splenectomy patients. During the enrollment period, the software version of the scanner was changed from DV25 to DV25.1, and then to DV26. The subjects were comprised of the DV25, DV25.1, and DV26 groups, which included an initial consecutive 20 men and 20 women who were examined using the three corresponding software versions. The study population had a mean ± standard deviation (SD) age of 67.3 ± 9.6 years (68.3 ± 12.6 and 71.5 ± 10.6 years for men and women, respectively). Characteristics of subjects in each group are shown in Table 1.

Means ± SDs are presented for age, height, and weight. The numbers of subjects are shown for background liver.

MR imaging was performed using a 3T clinical scanner (Discovery 750w; GE Healthcare, Waukesha, WI, USA) with a 32-channel phased-array coil. For dynamic imaging, Gd-EOB-DTPA (0.025 mmol/kg; Bayer Yakuhin, Osaka, Japan) was administered intravenously, and hepatobiliary-phase images were acquired 20 min later using a T1-weighted three-dimensional gradient echo sequence (liver acquisition with volume acceleration [LAVA]). Typical imaging parameters are shown in Table 2. The array spatial sensitivity encoding technique (ASSET) was used with a reduction factor of 2.5. Calibration images were acquired with the body coil and surface coil using a three-dimensional fast spoiled gradient-recalled echo sequence, and the images acquired with the surface coil were used for ASSET reconstruction. The preset mode was used for radiofrequency transmission. Image non-uniformity correction was performed using PURE. Uncorrected and corrected images were stored in a picture archiving and communication system (PACS).

The uniformity of liver signal intensity was visually assessed on a PACS viewer (EV Insite; PSP Corp., Tokyo, Japan) by three board-certified diagnostic radiologists independently. Each image set was randomly displayed on the PACS viewer. The observers were given no information about non-uniformity correction. They reviewed all slices, adjusting the display grayscale according to their preference, and graded superficial hyperintensity, i.e., hyperintensity near the liver surface, as an indicator of non-uniformity, using a three-point scoring system: 1 (severe), 2 (mild), and 3 (negligible). The observers graded each image set using reference images provided to help them understand the scoring system (Figure 1). The visual score for each image set was defined as the mean value of the scores for the three observers.

The field of view and the number of slices were increased in large patients.

Histogram analysis of liver signal intensity was conducted in 20 patients for each group, i.e., the initial consecutive 10 men and 10 women without focal liver lesions > 3 cm. The analysis was performed using a previously described method [11] . A histogram of liver signal intensities for each image set was created using ImageJ software (ver. 1.53e; National Institutes of Health, Bethesda, MD, USA). Seven slices around the level of the porta hepatis were selected so that the axial coordinate of the selected slice was located 1 cm from that of the adjacent selected slice. Regions of interest (ROIs) were drawn manually to cover the entire liver on selected uncorrected image slices and were also applied to the corrected images. A histogram (frequency vs. signal intensity) for a given image set was generated using data from the seven slices, and a simple averaging of 15 consecutive frequency values was performed. The mode signal of the histogram was determined. The signal range including the mode signal and continuously showing frequencies of more than half of the frequency at the mode was determined and the width of this range was divided by the mode signal to define full-width at half-maximum (FWHM). Full-width at quarter-maximum (FWQM) was defined similarly. The FWHM and FWQM were regarded as quantitative indicators of image uniformity, with high values indicating large variations in liver signals and poor uniformity.

To evaluate the spatial distribution of the effect of correction, correction maps were created for the same patients included in the histogram analysis, by dividing the corrected image by the uncorrected image at the level of the porta hepatis using ImageJ software. Circular ROIs (1000 mm²) were placed on the uncorrected image at the center and four peripheral regions (anterior, posterior, right, and left). They were applied to the correction map, and a mean value was obtained for the ROI. The normalized correction factor (nCF) was defined as the mean value of each ROI divided by that of the posterior ROI. In addition, a nCF map was created by dividing the correction map by the mean value of the posterior ROI.

The liver-to-muscle ratio (LMR) and liver-to-spleen ratio (LSR) of signals were obtained as quantitative indices of liver parenchymal enhancement. ROIs were set on the corrected images and applied to the uncorrected images. For the liver, circular ROIs (100 mm²) were set in the anterior and posterior segments of the right hepatic lobe and medial segment of the left lobe, avoiding vessels, tumors, and artifacts (Figure 2). Liver signal intensity was defined as the average of the mean signal intensities of the three ROIs. For muscle, circular ROIs (100 mm²) were set in the right and left paravertebral muscles, while minimizing the inclusion of fat. Muscle signal intensity was defined as the average of the mean signal intensities in the right and left ROIs. A slice showing the right main branch of the portal vein was used to assess liver and muscle signals. Spleen signal intensity was defined as the mean signal intensity in a circular ROI (200 mm²) set at the splenic hilum level. The LMR and LSR were calculated for each image by dividing the liver signal by the muscle and spleen signals, respectively.

Values are expressed as means ± SD. Statistical analyses were conducted using R version 4.2.1 (R Foundation for Statistical Computing, Vienna, Austria). Comparisons among patient groups distinguished according to the software version used (independent groups) were performed by one-way analysis of variance (ANOVA) followed by the Tukey-Kramer test for parametric data, and the Kruskal–Wallis test followed by the Steel-Dwass test for nonparametric data. Data with and without non-uniformity correction were compared using the paired t-tests (for parametric data) and the Wilcoxon signed-rank test (nonparametric data). When comparing nCFs among the different ROIs (repeated measures), repeated-measures ANOVA followed by Scheffe’s multiple comparisons test was used for parametric data, and the Friedman test followed by the

Wilcoxon signed-rank test was applied for nonparametric data. In all analyses, p < 0.05 was taken to indicate statistical significance. For visual assessment, Fleiss’s kappa statistics were calculated to assess interobserver agreement.

Severe superficial hyperintensity was observed in most uncorrected image sets, and tended to be evident in the anterior and posterior parts of the liver. The mean visual scores were close to 1 in all three groups, and there were no significant differences among the groups (Figure 3). The visual scores were higher for the corrected than uncorrected images in all groups, indicating improved uniformity of liver signals (p < 0.05 for all groups). The score was highest for the DV26 group, followed by the DV25 and DV25.1 groups, and there were significant differences in all paired comparisons (p < 0.05 for all comparisons). Fleiss's kappa coefficient was 0.633, showing substantial agreement.

Neither the FWHM nor FWQM differed significantly among the three groups for the uncorrected images (Figure 4). The FWHM and FWQM were smaller for the corrected than uncorrected images in all groups, indicating improved image uniformity (p < 0.05 in all groups). The FWHM and FWQM were significantly larger for the DV25.1 group than for the other two groups (p < 0.05 in both group comparisons).

Examples of nCF maps are presented in Figure 5. The nCF were lower in the peripheral regions than the central region, indicating that PURE decreased peripheral pixel values relative to central values; this corresponds to depression of

superficial hyperintensity. The nCF maps in the DV25 and DV26 groups showed lower nCFs in the anterior and posterior regions than in the right and left regions, implying stronger depression of superficial hyperintensity in the anterior and posterior regions. The nCF in the DV25 and DV26 groups were significantly lower for the anterior region than for the right and left regions (p < 0.05 in both groups) (Figure 6).

The nCF maps demonstrated that the nCFs tended to be lower in the DV25.1 group throughout the image section than in the DV25 and DV26 groups, implying weaker correction effects (Figure 5). The nCFs for the right, left, and central regions were significantly lower in the DV25.1 group than the other two groups (Figure 6). For the anterior region, the nCF was significantly lower in the DV25.1 than DV26 group, but there was no significant difference between the DV25 and DV25.1 groups (Figure 6).

Neither LMR nor LSR differed significantly among the three groups for the uncorrected images (Figure 7). Compared with the uncorrected images, the LMR was higher for the corrected images (p < 0.05 in all groups). The LMR for the corrected images was significantly higher in the DV25 and DV26 groups than in the DV25.1 group (p < 0.05 in both group comparisons). Compared with the uncorrected images, the LSR was higher for the corrected images (p < 0.05 in all groups), but there were no significant differences among the three groups.