MR Fingerprinting for Liver Tissue Characterization: A Histopathologic Correlation Study

Introduction

Diffuse liver disease, which affects approximately 1 billion people globally, is associated with clinically significant mortality (1,2). Its pathophysiologic changes include hepatic steatosis, inflammation, iron overload, and hepatic fibrosis. Although progressive fibrosis and cirrhosis can follow, these processes are dynamic and reversible through treatment (3). Therefore, assessing histologic changes is crucial for the diagnosis, prognosis, and therapy in patients with diffuse liver disease.

Biopsy is the standard to evaluate liver tissue in diffuse liver disease (4). Although histologic specimens provide detailed information at a microscopic level, liver biopsy is invasive and involves a risk of complications: Up to 84% of patients undergoing liver biopsy experience pain (5), and up to 5% experience active bleeding or infection (6), which makes repeated biopsies impractical. Moreover, the small size of tissue samples promotes sampling errors and interobserver variations (7). Therefore, a less invasive, quantitative, and reproducible method for assessing diffuse liver disease is needed.

MRI is a noninvasive approach in evaluating morphologic changes that allows for excellent quantification of liver tissue properties (8). Although CT and US are also useful for this, exposure to ionizing radiation is unavoidable in the former and relatively high intra- and interobserver variation is associated with the latter (9,10). MRI-derived T1, T2, T2* (1/R2*), and proton density fat fraction (PDFF) values are quantitative metrics reflecting the steatosis, inflammation, fibrosis, and siderosis status of liver tissue (8,11). Each metric assesses different liver tissue properties and could serve as imaging biomarkers for diffuse liver disease. For example, T1 and T2 values could be used to monitor inflammation and tissue fibrosis given their sensitivity to changes in extracellular water and extracellular matrix remodeling (12).

MR fingerprinting (MRF) is a rapid quantitative mapping technique that can simultaneously acquire multiple tissue properties in a single acquisition (13). Originally used for the brain, current development has enabled adoption to abdominal organs (14–17). Liver MRF enables simultaneous coregistered quantification of T1, T2, T2*, and PDFF maps in a single breath-hold acquisition, making it suitable for noninvasive liver tissue characterization (16). The bias and linearity of this technique have been investigated in a phantom in a previous study (16). However, past studies included mainly young healthy volunteers and only a few patients (13,14,16). Studies investigating the correlation of abdominal MRF techniques with liver histopathologic findings are lacking.

We hypothesized that the MRF-derived results would match separate reference quantitative mapping– and histologic analysis–derived observations. Our study aimed to compare liver MRF-derived metrics with separate reference quantitative mapping in participants with diffuse liver disease, evaluate test-retest repeatability of liver MRF, and validate MRF-derived measurements for histologic grading of liver biopsies.

Materials and Methods

MRI Scan Acquisition and Reconstruction

Measurements were performed on a 1.5-T scanner (Ingenia; Philips Healthcare) with a 32-channel body coil in transversal orientation.

Participants were examined in the supine position using two-dimensional single-section liver MRF in a single, approximately 18-second breath-hold scan at end-expiration (16,18). The sequence was composed of 18 acquisition blocks preceded by different preparation pulses (Fig 1). T1-weighted inversion preparation was applied before the first, fifth, and ninth blocks with delays of 15, 100, and 250 msec, respectively. Segments 11–14 and 15–18 used T2-weighted preparations of 35 and 60 msec, respectively (19). Blocks acquired 25 readouts with nine echoes each (T2* and PDFF) in 525 msec. A 400-msec recovery period between blocks was used to allow for signal recovery and improved signal-to-noise ratios (20,21). The acquisition parameters were as follows: first echo time msec, 1.9; change in echo time, 2.6; repetition time msec, 20.6; flip angle, 15°; field of view, 256 × 256 mm²; resolution, 2 × 2 mm²; 8-mm section thickness; golden radial trajectory (22) with 450 MRF time point images acquired over an 18-second breath hold. Data were reconstructed by using a sparse and locally low rank approach (23), followed by water and fat separation (24) to estimate MRF water images, MRF fat images, and PDFF, T2*, and B₀ maps. Finally, water-specific T1 and T2 values were estimated based on MRF water images by dictionary matching (13). The dictionary contained signal evolutions of approximately 102 000 combinations of T1 and T2 of interest (T1: 50–1500 msec with 5-msec intervals and 1500–2500 msec with 50-msec intervals; T2: 1–150 msec with 0.5-msec intervals and 150–300 msec with 5-msec intervals). Offline reconstruction was performed by a radiologist (S.F., with 7 years of experience) using software (Matlab2019a; MathWorks).

Figure 1:

Download as PowerPoint

Participants were also imaged using reference quantitative mapping acquired in three separate breath-hold acquisitions, composed of T1 modified Look-Locker inversion recovery (25), T2 gradient spin echo (26), and multiecho gradient echo (27) for T1, T2, and T2* PDFF mapping, respectively (Table E1 [online]). Liver MRF was also performed at the end of the same session (ie, no repositioning between the acquisitions) to evaluate intrasession repeatability.

Results

Participant Characteristics

The selection process is in Figure 2. A total of 56 participants (32 women; mean age, 59 years ± 15; age range, 28–87 years) were included to compare liver MRF with reference quantitative mapping. Clinical and laboratory demographics are shown in Table 1. A total of 23 participants (14 women; mean age, 52.7 years ± 12.7; age range, 29–74 years) underwent liver biopsy. The histopathologic characteristics of the biopsy samples are in Table 2. The average time between MRI and biopsy was 2.0 days ± 2.6.

Figure 2:

Download as PowerPoint

Table 1: Clinical and Laboratory Characteristics of Participants with Diffuse Liver Disease

Table 2: Clinical and Histopathologic Characteristics of the Liver Biopsy Group

Comparison of Liver MRF with Reference Standard Imaging and Scan-Rescan Repeatability of MRF Metrics

Representative MRF maps and reference T1, T2, T2*, and PDFF maps in a participant with diffuse liver disease are shown in Figure 3. Agreement with separate reference quantitative mapping measurements in participants with diffuse liver disease is shown in Figure 4A (R² = 0.86, 0.88, 0.54, and 0.78 for T1, T2, T2*, and PDFF values, respectively). In liver tissue, the agreement with reference standard measurements (R²) was 0.50, 0.25, 0.29, and 0.84 for T1, T2, T2*, and PDFF values, respectively. The T2* values of the spleen showed a relatively low correlation between MRF and reference quantitative mapping. Biases (MRF reference) and limits of agreement of the measurements with liver MRF in comparison to reference methods were 99 msec (−73 to 272 msec), −11 msec (−27 to 4 msec), 1.6 msec (−17 to 20 msec), and −0.6% (−9.6% to 8.5%) for T1, T2, T2*, and PDFF, respectively. The test-retest repeatability of each parameter acquired with MRF is shown in Figure 4B and 4C. The overall coefficients of variation for repeatability in the regions of interest of the liver were 3.2%, 5.5%, 7.1%, and 4.6% for T1, T2, T2*, and PDFF values, respectively. The interobserver agreement of MRF-based metrics is detailed in Appendix E2 (online).

Figure 3:

Download as PowerPoint

Figure 4:

Download as PowerPoint

Correlation with Histopathologic Findings

Representative maps and histologic images of participants with severe steatosis and fibrosis are shown in Figure 5 and Figure E1 (online), respectively. MRF-derived metrics correlated very strongly with semiquantitative histologic results of liver biopsies regarding steatosis and fairly with fibrosis, inflammation, and siderosis (Spearman correlation coefficient ρ for fibrosis, steatosis, inflammation, and siderosis: 0.27 [95% CI: 0.12, 0.40], 0.77 [95% CI: 0.70, 0.82], 0.36 [95% CI: 0.22, 0.48], and −0.46 [95% CI: −0.57, −0.33], respectively; P < .001 for all metrics) (Fig 6). After correcting for multiple comparisons, MRF showed higher PDFF in participants with higher steatosis grades (P < .001 for all combinations) and higher T1 values in participants with higher inflammation grades (grade 0 vs 1, P = .99; grade 0 vs 2, P = .06; grade 2 vs 3, P = .62; for all other combinations, P < .001). MRF showed lower T2* (higher R2*) values in participants with siderosis (grade 0 vs 1, P = .19; grade 1 vs 2, P = .73; grade 2 vs 3, P = .04; for all other combinations, P < .001). As shown in Figure E2 (online), MRF-derived metrics showed high diagnostic performance in the differentiation of moderate or severe changes from mild or no changes, except for fibrosis (area under the receiver operating characteristic curve for fibrosis, inflammation, steatosis, and siderosis: 0.62 [95% CI: 0.52, 0.61], 0.92 [95% CI: 0.88, 0.92], 0.98 [95% CI: 0.96, 0.97], and 0.74 [95% CI: 0.67, 0.74], respectively).

Figure 5:

Download as PowerPoint

Figure 6:

Download as PowerPoint

Discussion

Quantitative T1, T2, T2*, and proton density fat fraction (PDFF) are promising imaging biomarkers for assessing diffuse liver disease; however, these involve sequential acquisitions. Our prospective study compared liver MR fingerprinting (MRF)-derived T1, T2, T2*, and PDFF maps with separate reference quantitative mapping techniques and evaluated their repeatability and correlation with biopsy-based histologic results in participants with diffuse liver disease. We confirmed repeatability (coefficients of variation for T1, T2, T2*, and PDFF values: 3.2%, 5.5%, 7.1%, and 4.6%, respectively) with high agreement with reference quantitative mapping (R² = 0.86, 0.88, 0.54, and 0.99 for T1, T2, T2*, and PDFF mapping, respectively). Liver MRF metrics showed fair to very strong correlation with histologic evaluations and high diagnostic performance to differentiate moderate or severe changes from mild or no changes, except for fibrosis (area under the receiver operating characteristic curve for fibrosis, inflammation, steatosis, and siderosis: 0.62 [95% CI: 0.52, 0.62], 0.92 [95% CI: 0.88, 0.92], 0.97 [95% CI: 0.96, 0.97], and 0.74 [95% CI: 0.567, 0.74], respectively).

Whereas liver MRF showed high agreement with separate reference quantitative mapping (R² = 0.86, 0.88, 0.54, and 0.99 for T1, T2, T2*, and PDFF values, respectively), it produced slight overestimation of T1 (mean bias, 99 msec) compared with the modified Look-Locker inversion recovery and underestimation of T2 (mean bias, −11 msec) compared with the gradient spin echo, which is consistent with previous studies that used MRF (16,18,31). This could be partially explained by the tendency of modified Look-Locker inversion recovery and gradient spin echo to underestimate in vivo T1 (32,33) and overestimate in vivo T2 (34), respectively. The coefficients of determination between liver MRF and reference measurements were lower than those reported previously (16). This may be because of the inclusion of patients rather than healthy volunteers who are usually highly cooperative and have better breath-hold capabilities and less motion during acquisitions. The poor image quality of some reference quantitative mappings may also have reduced the coefficients of determination.

This study was performed in a clinical setting, thus including variances because of different MRI operators, participant movements, and poor breath holding. Nevertheless, liver MRF provided highly repeatable quantitative metrics and showed slightly higher and lower test-retest T1, T2*, and PDFF values, respectively, compared with shortened modified Look-Locker inversion recovery sequence, localized cardiac-triggered proton spectroscopy, and multigradient-echo acquisition (11,35). The inclusion of cooperative healthy volunteers in these studies, favoring general underestimations of the coefficients of variations and overestimations of PDFF coefficient of variation, may explain this disparity. The repeatability coefficient of MRF-based PDFF was calculated to be 4.8% (units in absolute PDFF percentages), which was comparable to those obtained with multiecho gradient-recalled-echo methods. A meta-analysis by Yokoo et al (36) reported the coefficient to be 3%–5%. It is noteworthy that our study performed test-retest in the same session (ie, the participants were not removed from the scanner between acquisitions), which may have overestimated the performance of the repeatability compared with those calculated using acquisitions performed with reposition or on different days.

Liver MRF acquires coregistered quantitative tissue maps in a single-breath hold acquisition, whereas reference quantitative mappings perform separate breath-hold acquisitions (scanning times: 11, 18, and 15 seconds for T1, T2, and T2* and PDFF mapping, respectively), reducing the acquisition time to approximately 40%. Because no additional time for multiple acquisition prescriptions or resting between breath holds are needed, participant fatigue and chances of poor-quality examinations are reduced. Moreover, the ability to acquire multiple tissue properties is an important aspect of liver MRF: Because all maps are obtained in a single acquisition during one breath hold, the maps are inherently aligned at the pixel level. This eliminates unavoidable misalignments from different breathing amplitudes during separate scans, enabling pixelwise multiparametric evaluation for downstream analysis in clinically relevant tasks, such as predicting treatment response and prognosis.

The MRF acquisition used in this study has some robustness to B₀ and B₁ errors because the parametric encoding for T1 and T2 uses preparation modules with adiabatic pulses (B₀/B₁ insensitive) and the readouts use spoiled gradient echoes with small flip angles (<15°), which have been demonstrated to be somewhat robust to B₁ variations in previous MRF studies (37,38). However, the sequence may be sensitive to other confounding factors, such as magnetization transfer (39), diffusion (40), or motion (41). The acquisition sequence can be improved to be insensitive to these parameters by fully encoding them as additional maps, as well as combining advanced model corrections for section profile, preparation pulses, and excitation B₁. For example, encoding T1 ρ could enhance sensitivity to fibrosis (17), improving overall liver tissue characterization. Comparison with MR elastography to assess fibrosis would be of interest for future studies. Most quantitative MRI methods, including MRF, focus on accurate estimation in a band of interest within the spectrum of each parameter, usually around the ranges of healthy tissue and relevant pathologic findings. Therefore, some parameter ranges (eg, very long T2 or T2*) may incur a bias, although further sequence modifications can be made for improved precision if needed. Encoding a large number of parameters may require further reconstruction improvements to maintain acquisition times within the breath-hold duration. However, liver MRF could also be performed freely breathing (possibly three-dimensionally to avoid through-plane motion) if participant motion is corrected (41).

Our study had several limitations. First, the current MRF framework was reconstructed offline, requiring up to 1 hour to produce coregistered T1, T2, T2*, and PDFF maps on an eight-core central processing unit workstation. Clinically feasible times can be achieved with a graphics processing unit–based implementation of the proposed approach or with deep learning approaches. Second, the liver MRF technique used in this study is a two-dimensional approach and does not cover the whole liver. Although a single section can provide clinically relevant information for diffuse liver disease, a three-dimensional alternative would be more valuable, especially for focal liver diseases (41,42). Third, the number of included participants undergoing liver biopsy was not large. Therefore, our findings merely provide initial proof of principle in a patient sample. In particular, because of the small sample size for liver iron deposition, which is relatively rare in Japanese compared with Western populations, the results regarding siderosis should be interpreted with caution. Fourth, we used a protocol that was preset by the manufactures for the reference quantitative mapping sequences used in our study. Small differences in parameters (eg, different echo spacing for T2* mapping) could potentially lead to some disagreement between techniques. Fifth, this study was conducted in a single institution using a single scanner. Future studies including larger samples from multiple centers are needed to further establish the clinical use of liver MRF.

In conclusion, liver MR fingerprinting provided repeatable T1, T2, T2*, and proton density fat fraction maps that were in high agreement with reference quantitative mapping and may correlate with histopathologic changes in liver tissue, especially hepatic steatosis, inflammation, and siderosis. This is a rapid and promising method for noninvasive liver tissue characterization and is suitable for the treatment of patients with diffuse liver disease. Future studies should include more biopsy samples and assess longitudinal changes to monitor treatment responses and disease progression.