Application of AI-assisted compressed sensing in high resolution Gd-EOB-DTPA-enhanced MRI during hepatobiliary phase at 5.0 T

Introduction

For both primary liver cancer and metastases, imaging examination and accurate staging are the bases for making effective treatment plans. Contrast-enhanced magnetic resonance imaging (MRI) is a useful tool for the detection and diagnosis of liver tumors, and is extensively utilized due to its high resolution for soft tissues and absence of ionizing radiation (1). As a dual functional MRI contrast agent (2), liver cell-specific contrast agent [gadolinium ethoxybenzyl diethylenetriaminepentaacetic acid (Gd-EOB-DTPA)] has a dynamic enhancement effect similar to that of extracellular contrast agent [gadolinium diethylenetriaminepentaacetic acid (Gd-DTPA)], and can be specifically absorbed by liver cells for cholangiography 20 minutes after injection. Gd-EOB-DTPA can significantly improve the detection rate and diagnostic accuracy of small (with a diameter ≤20 mm) and subcentimeter (with a diameter ≤10 mm) focal lesions (3). The utilization of Gd-EOB-DTPA-enhanced MRI in patients with suspected liver metastases has been shown to be advantageous, particularly in detecting subcentimeter liver metastases, with both sensitivity and specificity exceeding 80% (4), which encourages a more objective and scientific staging and management of the tumor.

The acquisition time and spatial resolution of MRI have a trade-off relationship (5). In order to obtain high spatial resolution images to further improve the diagnostic performance, the scanning time will increase, whereas the signal-to-noise ratio (SNR) may decrease (6). Therefore, performing high-resolution liver MRI within a breath-holding time without the loss of image quality is crucial and challenging (7). Currently, techniques that are commonly used to shorten the scanning time include compressed sensing (CS), parallel imaging (PI), and half-Fourier (HF) acquisition. However, when using high acceleration factors, insufficient sampling of these methods will introduce various artifacts and amplify noise during image reconstruction, which can reduce image quality (8-10). To achieve a balance between the scanning speed and image quality, artificial intelligence (AI)-assisted compression sensing (ACS) has been developed (11,12). The ACS method combines CS, PI, and HF techniques with AI modules during the acquisition and reconstruction of the down-sampled image (13). Previous studies have shown that ACS significantly shortens the scanning time without sacrificing the image quality, thus improving the efficiency and reducing motion artifacts (11,12,14).

In clinical settings, magnetic field strengths used for abdominal imaging are mainly 1.5 and 3.0 T. Researchers have concluded that abdominal imaging at 3.0 T has benefits over images obtained at 1.5 T due to the scaled signal intensity (SI), which is proportional to the field strength (15). However, higher field strengths (>3.0 T) for abdominal imaging have proven challenging due to safety issues such as high specific absorption ratio (SAR) and technical problems such as low B0 and B1 homogeneity. Recently, the 5.0 T whole-body MRI scanner has been developed, which offers advantages to abdominal imaging and avoids the shortcomings of 7.0 T such as a high SAR (16-18). So far, the performance of high-resolution Gd-EOB-DTPA MRI during the hepatobiliary phase (HBP) has never been evaluated at 5.0 T. We assume that the application of ACS technology in Gd-EOB-DTPA MRI during the HBP could improve image quality at 5.0 T. This study aimed to assess the image quality and lesion detection performance of ACS accelerated Gd-EOB-DTPA MRI during the HBP in comparison of HBP imaging using PI at 5.0 T. We present this article in accordance with the STARD reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-264/rc).

Methods

Patients

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study was approved by the Biomedical Ethics Committee of The First Affiliated Hospital of University of Science and Technology of China (No. 2024-RE-15). The requirement for written informed consent was waived for this retrospective study. A total of 279 patients with suspected liver tumor lesions (including primary and secondary tumors) underwent Gd-EOB-DTPA-enhanced MRI from May 2023 to June 2024. The inclusion criteria were as follows: (I) complete clinical and imaging examinations; (II) liver lesions consistent with primary or secondary liver tumor manifestations; and (III) the lesion did not receive any treatment in the past such as systematic anti-tumor and local-region treatment, as this may affect the interpretation of image. The exclusion criteria were as follows: (I) poor image quality so that subjective evaluation and image analysis could not be carried out; (II) patients with a history of contrast agent allergy and/or poor renal function; and (III) excessive number of lesions (>20) which may interfere with image analysis.

MRI examination methods

All patients underwent abdominal scans on a 5.0 T MRI scanner (uMR Jupiter, United Imaging Healthcare, Shanghai, China) using a 24-channel anterior body coil combined with a 48-channel posterior spine coil placed in the patient table. Every patient received respiratory training before scanning to reduce potential respiratory artifacts. Scanning sequences included axial T1-weighted gradient echo (GRE) sequence, axial T2-weighted fast spin echo sequence, diffusion-weighted echo-planar imaging, as well as pre- and post-contrast three-dimensional (3D) fast spoiled GRE sequence (which is named quick-3D in United Imaging Healthcare MRI systems). The PI accelerated quick-3D sequence was applied for contrast-enhanced imaging for all phases and the HBP imaging was applied using two acceleration techniques: (I) axial and coronal quick-3D sequence with PI acceleration (PI-quick-3D); and (II) axial quick-3D sequence with ACS acceleration (ACS-quick-3D). Due to the fact that the respiratory motion is one of the most critical factors in the image quality of abdominal MRI, both the PI and ACS accelerated quick-3D scanning were performed within a clinically acceptable breath-holding duration for patients (17 s for PI and 18 s for ACS). As the scanning time was constrained, the spatial resolutions of ACS and PI quick-3D sequences were optimized respectively for best performance. The sequence parameters for HBP imaging are shown in Table 1.

The hepatocyte-specific contrast agent, Gd-EOB-DTPA (Primovist; Bayer Healthcare, Berlin, Germany), was injected via a high-pressure injector into the elbow vein at a rate of 1–2 mL/s, and with a dosage of 0.1 mL/kg, and then flushed with 20 mL of normal saline. Images of dual-arterial phase (AP) including the early and late phases, portal venous phase (PVP), transition phase (TP), and HBP were collected at 10 seconds (s), 60 s, 120 s, and 20 minutes after the injection of contrast agent.

ACS image reconstruction

ACS accelerates MRI, which includes techniques like partial Fourier transform, PI, CS, and convolutional neural networks (CNNs). It has been approved by US Food and Drug Administration (FDA), highlighting its safety and efficacy for clinical applications. The CNN architecture processes k-space data through multiscale sparsification. The central k-space region of the collected data is fully sampled, whereas peripheral regions undergo pseudo-random undersampling. Subsequently, undersampled k-space data are reconstructed into fully sampled images via the CNN (13,19). This architecture is predicated on a U-Net framework (14), incorporating residual blocks with skip connections to expedite convergence and a least-squares generative adversarial network methodology to optimize reconstruction fidelity. The underlying mathematical formulation assimilates principles from CS, partial Fourier, and PI. The AI module of ACS used in this study was trained by two million fully sampled images, constituted predominantly (98%) by human volunteer scans and supplemented by phantom data (2%). A pivotal innovation of ACS involves the incorporation of the AI module’s output as an additional constraint within the CS reconstruction paradigm (20). This integration enforces consistency between reconstructed images and fully sampled reference data. Mathematically, this is implemented by adding a regularization term to the objective function, in order to minimize the discrepancy between AI-predicted and actual reconstructed images. The objective function is given by:

Argmin‖Ex−y‖22+λ1‖ψ(x)‖1+λ2‖ϕ(xAI,x)‖1+λ3‖PPA‖2+λ4‖PF‖1[1]

where x denotes the image to be reconstructed, E represents the Fourier encoding matrix with a binary k-space sampling mask, y signifies the acquired k-space data, ψ indicates the sparse transform reflecting a total variation, xAI refers to the image created by the AI module, and λi represents the regularization parameters associated with each term (all regularization parameters are standardized to the same scale in the regularization computation, where λ1, λ2, λ3, and λ4 are determined by the algorithm automatically); additionally, and PF correspond to partially parallel acquisition and partial Fourier components, respectively. By integrating these components, the acceleration factor is established as the sole adjustable parameter for the ACS technique, which denotes the level of acceleration.

Image analysis

All contrast-enhanced images obtained during HBP were transferred to a post-processing workstation (uWS-MR, United Imaging Healthcare, Shanghai, China) and used for image analyses. For ACS-quick-3D, axial images were converted into 0.6 mm-thick coronal images using the multiplanar reformation method on the workstation.

Qualitative assessment

Two senior radiologists with expertise in abdominal imaging (reader 1 with 8 years and reader 2 with 13 years of experience, respectively) independently evaluated HBP images obtained at 5.0 T. They individually graded axial and coronal PI- and ACS-quick-3D HBP images without knowing the patient information, sequence, and final diagnosis results. The scoring criteria included image artifacts, lesion margin clarity, liver edge clarity, vessel clarity, bile duct clarity, and overall image quality. The biliary system includes the extrahepatic bile ducts (common bile duct and hepatic duct) and the intrahepatic bile ducts (first and second-degree intrahepatic bile duct branches). The image quality of the biliary system was assessed based on a 5-point grading scale proposed by Papanikolaou (21): 0, bile ducts not visible; 1, less than half of the bile ducts is visible, with moderate to severe artifacts; 2, more than half of the ducts is visible, with mild to moderate artifacts; 3, the entire length of the ducts is visible, with clear margins and mild artifacts; and 4, excellent visualization. The other aspects of the image quality ware graded based on a 5-point Likert scale (22): 1, severe artifacts, impaired image quality and unusable for diagnosis; 2, prominent artifacts, anatomical structures were blurry and interfering diagnosis; 3, moderate artifacts, image quality was adequate for diagnosis; 4, mild artifacts, good image quality; and 5, no artifacts, excellent quality. The average score from reader 1 and reader 2 was determined as the final score for each image set.

SNR, contrast-to-noise ratio (CNR), and contrast ratio (CR)

Four fixed-sized regions of interest (ROIs) with an area of 100 mm² were placed at the same anatomical level in axial PI and ACS sequence to measure the SI of the left liver lobe, right anterior lobe, and right posterior lobe parenchyma, avoiding blood vessels, bile ducts, and lesions. The SI of normal liver tissues was calculated as the average intensity of the four normal liver ROIs. The PI technique used in this study did not involve mask processing and the background intensity was not suppressed to zero or near-zero values. Therefore, four background ROIs (with an area of 500 mm²) were placed at the four corners of the image on the same slice where the ROIs for normal liver tissues were placed (Figure 1). Three slices were selected for drawing the lesion ROI: the slice that displays the largest cross-sectional area of the lesion, the upper adjacent slice, and the lower adjacent slice. On each selected slice, an ellipse-shaped ROI was drawn as large as possible to cover the lesion. The SI of the lesion was measured as the mean SI of the three lesion ROIs. If there were multiple lesions in the liver of a patient, the largest lesion was selected for ROI placement. The placement of all ROIs was determined by two radiologists who reached a consensus (readers 1 and 2). The background noise was calculated as the mean of standard deviations (SDs) of the SI in four background ROIs. The SNR of the liver parenchyma, CNR, and CR were calculated using the formulas:

SNR=LivertissueSI/BackgroundnoiseCNR=|LivertissueSI−LesionSI|/BackgroundnoiseCR=|LivertissueSI−LesionSI|/(LivertissueSI+LesionSI)[2]

Figure 1 The schematic diagram of ROIs. (A) T2-weighted image; (B) HBP image that displays the lesion with the largest cross-sectional area; (C) HBP image for drawing the ROIs of liver parenchyma. Using the T2-weighted image (A) as a reference, manually draw the largest ellipse-shaped ROI on the lesion on HBP image (B), avoiding large blood vessels, necrosis, and bleeding. The four smaller round-shaped ROIs of liver parenchymal and the four bigger round-shaped ROIs for the background are drawn on the HBP image (C). HBP, hepatobiliary phase; ROI, region of interest.

Image sharpness

The originally obtained axial HBP images were involved in the analysis of image sharpness. To measure the image sharpness, the state-of-art method for sharpness measurement based on maximal local variation (MLV) was applied (23). In this method, the MLV of each pixel was calculated as the maximum difference between the intensity of the pixel with respect to its eight neighboring pixels. By sliding this 3×3 window across an image I of size M×N pixels, the MLV map ϕ(I) is generated as:

ϕ(I)=(φ(I1,1)⋯φ(I1,N)⋮⋱⋮φ(IM,1)⋯φ(IM,N))[3]

where φ(Ii,j) is the MLV of a pixel which is located at (i,j) of the image I, subject to 1≤i≤M and 1≤j≤N. Human vision is more sensitive to regions with higher intensity variations, therefore, regions with higher MLV values work better for sharpness assessment. To enhance the MLV of high variation regions, nonlinear weighting was applied to amplify the tail end of MLV distribution. The weighted MLV map is given by:

ϕW(I)=(w1,1φ(I1,1)⋯w1,Nφ(I1,N)⋮⋱⋮wM,1φ(IM,1)⋯wW,Nφ(IM,N))[4]

where the weight wi,j is added to the MLV of the pixel at location (i,j) and wi,j=eηi,j. The maximum of MLV is normalized to 1 and ni,j is the ranking of φ(Ii,j) when sorted in ascending order.

In the last step of the method, the weighted MLV distribution is parameterized with generalized Gaussian distribution (24-26), given by:

f(ϕW(I);μ,γ,σ)=(γ2σΓ(1γ)Γ(1γ)Γ(3γ))e−(ϕW(I)−μσΓ(1γ)Γ(3γ))γ[5]

where μ, σ, and γ is the mean, the SD, and the shape parameter of the distribution, respectively, and Γ(.) is the gamma function. The SD of the generalized distribution of MLV, σ, is used as the sharpness measurement metric.

The calculation of MLV sharpness score was processed using MATLAB (MathWorks, Natick, MA, USA). For each patient, a single axial slice that displayed the largest cross-section of the tumor was selected to calculate the MLV sharpness score.

Diagnosis and lesion detection rate

The final diagnosis is based on the combination of clinical history, serum biomarkers [such as alpha-fetoprotein (AFP), carbohydrate antigen 199 (CA199), etc.], and MRI findings (27,28), or histological confirmation. The typical characteristic of hepatocellular carcinoma (HCC) is the “wash in” and “wash out” patterns on dynamic contrast-enhanced MRI, accompanied by high-risk factors for HCC and elevated AFP. Patients diagnosed with liver metastases all had a history of other primary tumors, and MRI features of liver lesions were similar to or consistent with the primary tumor.

Huang et al.’s research showed that subcentimeter HCC has better survival outcomes, thus early diagnostic criteria and immediate treatment for subcentimeter HCC may be warranted (29). Therefore, the lesions in this study were divided into two groups (with a diameter >10 and ≤10 mm). A radiologist with 30 years of experience in abdominal MRI established a standard reference for lesion detection, including the location and number of lesions by reviewing all MRI sequences as well as the clinical data. The other two radiologists with 8 and 13 years of respective experience in abdominal MRI independently assessed the number and location of lesions by reviewing HBP images acquired with PI and ACS.

Statistical analysis

Statistical analyses were performed using the software SPSS 26.0 (IBM Corp., Armonk, NY, USA). All the continuous variables in the results of the study were expressed as means ± SDs. The Kolmogorov-Smirnov test was used to examine whether the subjective and objective indices were normally distributed or not. To compare SNR, CNR, and CR of PI-quick-3D and ACS-quick-3D, non-parametric Mann-Whitney U test was used. In terms of the subjective assessment of the image quality and the lesion detection rate, chi-square test and Wilcoxon signed-rank test were used, respectively. The consistency of subjective scores and lesion detection results from the two radiologists was examined by the Cohen’s kappa. Inter-observer agreement κ value was explained as follows: 0–0.20, slight agreement; 0.21–0.40, fair agreement; 0.41–0.60, moderate agreement; 0.61–0.80, substantial agreement; and 0.81–0.99, almost perfect agreement. A P<0.05 was considered statistically significant.

Results

Diagnosis

According to the flowchart, a total of 133 patients with liver tumor lesions were included in the study, including 104 males and 29 females, age range 29–82 years, with an average age of 60.5±10.1 years (Figure 2). The final diagnosis of these lesions was confirmed through biopsy or surgical pathology examination (n=55) or a combination of typical imaging findings and clinical follow-up (n=78). Participants included 47 cases of HCC (20 cases of postoperative recurrence), 3 cases of intrahepatic cholangiocarcinoma (ICC, 2 cases of postoperative recurrence), and 28 cases of metastasis (Table 2). There were more males in the enrolled patients, as HCC has obvious male-dominant characteristics, which may be related to male risk aggregation and hormone level differences (30).

Figure 2 Flowchart of the study population. CLM, colorectal liver metastasis; Gd-EOB-DTPA, gadolinium ethoxybenzyl diethylenetriaminepentaacetic acid; HCC, hepatocellular carcinoma; ICC, intrahepatic cholangiocarcinoma; MRI, magnetic resonance imaging.

Image quality scores

In the subjective evaluation of image quality in 133 cases, except for similar image artifact scores (for both axial and coronal image sets, P=0.850, and 0.054, respectively), ACS-quick-3D images (including axial and coronal images) had significantly higher scores in lesion margin clarity, liver edge clarity, vessels clarity, bile ducts clarity, and overall image quality (all P<0.001) than PI-quick-3D images (Table 3, and Figures 3,4).

Figure 3 Colorectal liver metastasis of a 63-year-old man. The margin of the lesions, vessels, as well as liver edge are more visible on the axial HBP image acquired by ACS (B) than the HBP image acquired by PI (A). ACS, artificial intelligence-assisted compressed sensing; HBP, hepatobiliary phase; PI, parallel imaging.

Figure 4 MRC evaluation based on HBP images. MIP images showed that both the intrahepatic and extrahepatic bile ducts, as well as the cystic duct, are more clearly displayed on the reconstructed coronal HBP image obtained by ACS (A) than both coronal (B) and reconstructed coronal (C) HBP images obtained by PI. ACS, artificial intelligence-assisted compressed sensing; HBP, hepatobiliary phase; MIP, maximum intensity projection; MRC, magnetic resonance cholangiography; PI, parallel imaging.

Consistency between the two readers

The κ values of subjective scores of the axial and coronal HBP images acquired by PI-quick-3D and ACS-quick-3D from the two readers were all greater than 0.6, indicating good consistency (Tables 4,5).

Objective evaluation of images

In the objective evaluation of image quality in 133 cases, MLV sharpness score, SNR, CNR, and CR of the HBP images obtained by ACS-quick-3D were significantly higher than those of PI-quick-3D (3.74×10⁻²±1.21×10⁻², 341.22±112.17, 145.05±71.94, and 0.266±0.094 vs. 2.46×10⁻²±0.68×10⁻², 226.00±74.39, 88.79±45.42, and 0.242±0.098, P<0.001) (Table 6). This indicates that the ACS-quick-3D images have higher clarity, which is beneficial for displaying anatomical structures, lesions, and the detection of small lesions.

Comparison of lesion detection capability

According to the reference standard, a total of 535 lesions were detected (233 >10 mm, 302 ≤10 mm). The two readers showed good consistency in lesion detection results between ACS and PI sequences (κ>0.8) (Table 7). Therefore, only the results of the reader with 13 years of experience were chosen for analysis. Among the 233 focal liver lesions with a diameter >10 mm, 231 lesions (99.14%) and 230 (98.71%) lesions were detected by ACS-quick-3D and PI-quick-3D, respectively. Among the 302 focal liver lesions with a diameter ≤10 mm, 297 lesions (98.34%) and 287 (95.03%) lesions were detected by ACS-quick-3D and PI-quick-3D, respectively. The lesion detection rate of ACS-quick-3D was similar to that of PI-quick-3D, and there was no statistically significant difference (P=0.614, and 0.279, respectively) (Table 8).

Discussion

This is the first study to compare the image quality of contrast-enhanced imaging with PI and ACS during HBP at 5.0 T. The results showed that with a similar scanning time (17 s for PI vs. 18 s for ACS), the HBP images acquired by ACS were superior to PI-quick-3D in subjective evaluations of lesion margin clarity, liver edge clarity, vessel clarity, display of biliary system, and overall image quality, as well as objective evaluations including SNR, CNR, CR, and image sharpness. In terms of the lesion detection rate, the lesion detection rate for ACS-quick-3D was similar to that of PI-quick-3D, especially for lesions smaller than 10 mm located under the liver capsule.

Studies have shown that compared to contrast-enhanced computed tomography, GD-EOB-GTPA-enhanced MRI can improve the detection and diagnosis of primary liver tumors and metastases (1,31,32). However, MRI examinations are challenged by the long scanning time, which is particularly important for abdominal imaging where respiratory-triggered or breath-hold sequences are primarily applied and motion artifacts are inevitable (33). Another challenge of MRI is that high spatial resolution imaging is limited by decreased SNR. ACS combines CS, HF, and PI, and innovatively integrates a deep learning neural network as an AI module into the reconstruction process (34). The AI module reduces noise and artifacts, complements the defects of CS, HF, and PI when they are used individually, and provides reliable MRI images for clinical diagnosis in a shorter time (35,36). Li et al. applied ACS technology to single-breath-hold T2-weighted imaging and showed promising performance as ACS provided significantly better abdominal image quality and lesion detectability with a considerable decrease in scanning time as compared with the conventional respiratory-triggered PI sequence (11). Another study in nasopharyngeal carcinoma demonstrated that ACS could not only shorten the scanning time but also improve image quality compared to the PI technique (37). This study applied ACS technology to HBP imaging and obtained high-resolution thin-slice images with scanning time similar to that of the conventional PI sequence (18 vs. 17 s). Under the constraint that the scanning time of HBP imaging should be acquired within one breath-hold, ACS and PI sequence parameters were optimized separately to meet their clinical limits. The study by Wang et al. (38) demonstrated that the visualization time of intrahepatic bile ducts after Gd-EOB-DTPA injection was 10.00±1.67, 11.50±3.32, and 14.50±4.22 minutes in the non-cirrhosis group, Child-Pugh A group, and Child-Pugh B group, respectively. The adequate hepatocyte phase could be acquired 5 minutes after the initial visualization of intrahepatic bile ducts. Therefore, in this study, HBP scanning time for all patients was uniformly delayed to 20 minutes to minimize the impact of temporal variations. Additionally, the total scanning time for the three sequences (including axial and coronal PI-quick-3D, and axial ACS-quick-3D) did not exceed 1 minute, and there was no fixed scanning order.

Early detection of tumor lesions is crucial for early tumor treatment and improving survivability, especially for HCC (29,39). When the selected slice is thick or the lesion is smaller than the image thickness, partial volume effects would be caused and lead to missed diagnosis or misdiagnosis of small liver lesions (40,41). Thin-slice scanning or imaging the body in different anatomical planes help to address such missed diagnosis or misdiagnosis. Although decreasing slice thickness can reduce partial volume effects, SNR would decrease and scanning time would increase (42). The ACS acceleration allowed thin-slice imaging without increasing the scanning time. Due to the high-resolution imaging of ACS-quick-3D in this study, the CR between lesions and liver parenchyma is amplified. Thin-slice acquisition could effectively reduce partial volume effects, which allowed the lesions located under the capsule that were not shown in PI-quick-3D images to be clearly displayed in ACS-quick-3D images (Figures 5,6), although there was no statistically significant difference. These results suggest that ACS can improve image quality while ensuring high lesion detection rates, especially for subcentimeter lesions under the liver capsule.

Figure 5 A 76-year-old man with intrahepatic cholangiocarcinoma. A lesion (white arrow) is detected under the capsule at the top of the liver in axial (A), coronal (B) and reconstructed sagittal (C) HBP images using PI, whereas the axial HBP image using ACS (D), combined with its coronal (E) and sagittal (F) reconstructions showed that this low signal intensity was a false-positive lesion caused by partial volume effects. ACS, artificial intelligence-assisted compressed sensing; HBP, hepatobiliary phase; PI, parallel imaging.

Figure 6 A 51-year-old man with recurrent hepatocellular carcinoma after surgery. A subcentimeter lesion (white, thin arrow) can be clearly detected under the capsule at S1 segment by the axial (C) and reconstructed coronal (D) HBP images using ACS, whereas the axial (A) and coronal (B) HBP images using PI could not determine whether it is a real lesion or artifacts caused by partial volume effects. The clarity of the liver (white, thick arrow) and lesion edges (black, thin arrow) in ACS image were better than that in PI image. ACS, artificial intelligence-assisted compressed sensing; HBP, hepatobiliary phase; PI, parallel imaging.

Due to the efficiency of ACS acceleration, high-resolution and thin-slice images can be acquired once in axial plane within a single breath-hold and used to reconstruct corresponding coronal images. For both axial and coronal planes, ACS-accelerated high-resolution images showed significantly higher qualitative evaluation scores and similar and even better lesion detection rate than the PI sequence. Previous research has shown that contrast-enhanced magnetic resonance cholangiography (MRC) is a practical technique that complements the T2-weighted MRC (43). In order to achieve a higher image resolution, researchers have proposed the use of navigated 3D contrast-enhanced T1-weighted MRC, which is mainly used for biliary imaging and has been shown to be feasible and robust (44,45). However, navigator gating requires a longer scanning time and would be affected by irregular breathing, leading to motion artifacts. With the help of ACS, the axial high-resolution HBP images with a thickness of 0.65 mm were obtained within only one breath-hold, and reconstructed sagittal and coronal images were able to display the bile duct with high level of visibility and image quality. This indicates that the application of ACS-quick-3D in HBP imaging helped to visualize the anatomy of the bile ducts, and investigate its relationship with intrahepatic lesions and the function of bile ducts.

Previous studies have shown the feasibility and improvement of abdominal MRI at 5.0 T as compared to lower field strengths. For structural imaging, images at 5.0T were shown to have higher SNR and better visualization of small vessels and lesions, thereby improving the lesion detection rate and diagnostic accuracy (17,18,46-49). Besides, quantitative MRI parameters such as apparent diffusion coefficient derived from diffusion-weighted imaging, and proton density fat fraction measured by multi-echo Dixon imaging at 5.0 T exhibited good repeatability and consistency with traditional field strengths (16-18,50,51). However, 5.0 T MRI faces challenges intrinsically caused by the ultra-high field strength. Compared to 1.5 and 3.0 T, the chemical shift increases at 5.0 T and impairs B0 field homogeneity. For B0 shimming of the 5.0 T system, a pre-scan low spatial resolution dual-echo GRE sequence was applied for B0 mapping. The phase difference of the two echoes was calculated to estimate the B0 field variations across the field of view (FOV). Subsequently, active shim coils modulated electrical currents to generate a supplemental magnetic field, which was superimposed to an uncorrected magnetic field to calibrate a B0 field. Meanwhile, the wavelength of the B1 field is shortened due to the fact that Larmor frequency increases at 5.0 T. This results in standing wave effects and causes B1 inhomogeneity across the body. For the 5.0 T scanner in this study, an eight-channel volume transmission coil was mounted in the bore of the magnet. These channels were independently connected to the radio frequency power amplifier and evenly distributed around the body. They transmitted and adjusted radiofrequency in parallel, ensuring a homogeneous RF excitation across a large volume.

There are several limitations to this study. First, among the 133 participants, only 55 (41.35%) cases had been pathologically confirmed. According to current research and guidelines (52,53), systemic therapy is recommended for liver metastases, whereas surgical resection combined with chemotherapy is advised for patients with resectable colorectal cancer liver metastases. Among the patients enrolled in this study, 50 cases of liver cancer (including HCC and ICC) were deemed unsuitable for surgery due to advanced tumor stages or had surgical interventions declined by their families. Additionally, 28 patients with multiple metastases also opted out of surgical treatment. Meanwhile, the main purpose of this study was to evaluate the image quality and the detection rate rather than the characterization of the lesions. Second, all images were acquired at 5.0 T yet comparisons with images obtained at other magnetic field strengths were not involved. However, the aim of this research was to evaluate the image quality of ACS in relation to PI, rather than making comparisons across various field strengths. Also, the ACS used in this study was specific to 5.0 T and was trained by fully-sampled images exclusively obtained at 5.0 T. Besides, previous studies have demonstrated the superiority and feasibility of abdominal MRI at 5.0 T over both 1.5 T and 3.0 T (16-18,51). Additionally, our parallel head-to-head investigation comparing 3.0 and 5.0 T contrast-enhanced scans in patients with suspected hepatic tumors is currently underway. The subjective and objective image quality, as well as imaging features of dynamic MRI were compared. Third, we did not evaluate the diagnostic performance of contrasted-enhanced MRC for biliary diseases or anatomical variations. This will be carried on in depth in our future studies, where it would be compared with the traditional T2 magnetic resonance cholangiopancreatography. Fourth, the PI noise is non-uniform across the FOV, which would influence the estimation of the noise level. However, the objective of this study was to compare the image quality of ACS and PI. The same ROI locations were used for both techniques, and the SNR at the same location was compared. In future studies, other methods for measuring PI noise such as acquiring a pure noise image might be introduced. Finally, no significant improvement was demonstrated in lesion detection rates of ACS sequence. However, the radiologists in this study reported that ACS-quick-3D displayed subcapsular lesions better than did PI-quick-3D. Subcapsular lesions would be exclusively involved to assess the detection ability of the sequence in future studies.

Conclusions

Compared to the traditional PI method, ACS-quick-3D achieved higher subjective and objective image scores, providing better image quality for displaying lesions, bile ducts, and vessels. Meanwhile, with a similar acquisition time to traditional PI sequence, ACS enabled high-resolution thin-slice imaging and showed comparable performance in detecting small liver lesions. This study demonstrated a promising application of ACS in high-resolution liver MRI at 5.0 T.

Acknowledgments

We would like to thank all the participants who contributed to our research.

Footnote

Reporting Checklist: The authors have completed the STARD reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-264/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-264/dss

Funding: The study was supported by the National Key R&D Program of China (No. 2019YFA0709300).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-264/coif). All authors report the funding from the National Key R&D Program of China (No. 2019YFA0709300). R.T. and X.S. are employed by Shanghai United Imaging Healthcare Co., Ltd. The authors have no other conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments, and was approved by the Biomedical Ethics Committee of the First Affiliated Hospital of University of Science and Technology of China (No. 2024-RE-15). The requirement for written informed consent was waived for this retrospective study.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Fowler KJ, Linehan DC, Menias CO. Colorectal liver metastases: state of the art imaging. Ann Surg Oncol 2013;20:1185-93. [Crossref] [PubMed]
2. Murakami T, Sofue K, Hori M. Diagnosis of Hepatocellular Carcinoma Using Gd-EOB-DTPA MR Imaging. Magn Reson Med Sci 2022;21:168-81. [Crossref] [PubMed]
3. Raman SS, Leary C, Bluemke DA, Amendola M, Sahani D, McTavish JD, Brody J, Outwater E, Mitchell D, Sheafor DH, Fidler J, Francis IR, Semelka RC, Shamsi K, Gschwend S, Feldman DR, Breuer JUnited States EOB Study Group. Improved characterization of focal liver lesions with liver-specific gadoxetic acid disodium-enhanced magnetic resonance imaging: a multicenter phase 3 clinical trial. J Comput Assist Tomogr 2010;34:163-72. [Crossref] [PubMed]
4. Ichikawa S, Goshima S. Gadoxetic Acid-Enhanced Liver MRI: Everything You Need to Know. Invest Radiol 2024;59:53-68. [Crossref] [PubMed]
5. Clarke SE, Saranathan M, Rettmann DW, Hargreaves BA, Vasanawala SS. High resolution multi-arterial phase MRI improves lesion contrast in chronic liver disease. Clin Invest Med 2015;38:E90-9. [Crossref] [PubMed]
6. Ihara K, Onoda H, Tanabe M, Iida E, Ueda T, Kobayashi T, Higashi M, Nickel MD, Imai H, Ito K. Breath-hold High-resolution T1-weighted Gradient Echo Liver MR Imaging with Compressed Sensing Obtained during the Gadoxetic Acid-enhanced Hepatobiliary Phase: Image Quality and Lesion Visibility Compared with a Standard T1-weighted Sequence. Magn Reson Med Sci 2024;23:146-52. [Crossref] [PubMed]
7. Yang AC, Kretzler M, Sudarski S, Gulani V, Seiberlich N. Sparse Reconstruction Techniques in Magnetic Resonance Imaging: Methods, Applications, and Challenges to Clinical Adoption. Invest Radiol 2016;51:349-64. [Crossref] [PubMed]
8. Bydder M, Ali F, Ghodrati V, Hu P, Yao J, Ellingson BM. Minimizing echo and repetition times in magnetic resonance imaging using a double half-echo k-space acquisition and low-rank reconstruction. NMR Biomed 2021;34:e4458. [Crossref] [PubMed]
9. Schulz J. P Marques J, Ter Telgte A, van Dorst A, de Leeuw FE, Meijer FJA, Norris DG. Clinical application of Half Fourier Acquisition Single Shot Turbo Spin Echo (HASTE) imaging accelerated by simultaneous multi-slice acquisition. Eur J Radiol 2018;98:200-6. [Crossref] [PubMed]
10. Feng L, Benkert T, Block KT, Sodickson DK, Otazo R, Chandarana H. Compressed sensing for body MRI. J Magn Reson Imaging 2017;45:966-87. [Crossref] [PubMed]
11. Li H, Hu C, Yang Y, Zhao Y, Lin C, Li Z, Liu Q. Single-breath-hold T2WI MRI with artificial intelligence-assisted technique in liver imaging: As compared with conventional respiratory-triggered T2WI. Magn Reson Imaging 2022;93:175-80. [Crossref] [PubMed]
12. Zhao Y, Peng C, Wang S, Liang X, Meng X. The feasibility investigation of AI -assisted compressed sensing in kidney MR imaging: an ultra-fast T2WI imaging technology. BMC Med Imaging 2022;22:119. [Crossref] [PubMed]
13. Wang S, Cao G, Wang Y, Liao S, Wang Q, Shi J, Li C, Shen D. Review and Prospect: Artificial Intelligence in Advanced Medical Imaging. Front Radiol 2021;1:781868. [Crossref] [PubMed]
14. Sheng RF, Zheng LY, Jin KP, Sun W, Liao S, Zeng MS, Dai YM. Single-breath-hold T2WI liver MRI with deep learning-based reconstruction: A clinical feasibility study in comparison to conventional multi-breath-hold T2WI liver MRI. Magn Reson Imaging 2021;81:75-81. [Crossref] [PubMed]
15. Radbruch A, Paech D, Gassenmaier S, Luetkens J, Isaak A, Herrmann J, Othman A, Schäfer J, Nikolaou K. 1.5 vs 3 Tesla Magnetic Resonance Imaging: A Review of Favorite Clinical Applications for Both Field Strengths-Part 2. Invest Radiol 2021;56:692-704. [Crossref] [PubMed]
16. Zhang Y, Yang C, Liang L, Shi Z, Zhu S, Chen C, Dai Y, Zeng M. Preliminary Experience of 5.0 T Higher Field Abdominal Diffusion-Weighted MRI: Agreement of Apparent Diffusion Coefficient With 3.0 T Imaging. J Magn Reson Imaging 2022;56:1009-17. [Crossref] [PubMed]
17. Zheng L, Yang C, Sheng R, Dai Y, Zeng M. Renal imaging at 5 T versus 3 T: a comparison study. Insights Imaging 2022;13:155. [Crossref] [PubMed]
18. Zheng L, Yang C, Liang L, Rao S, Dai Y, Zeng M. T2-weighted MRI and reduced-FOV diffusion-weighted imaging of the human pancreas at 5 T: A comparison study with 3 T. Med Phys 2023;50:344-53. [Crossref] [PubMed]
19. Liu K, Xi B, Sun H, Wang J, Chen C, Wen X, Zhang Y, Zeng M. The clinical feasibility of artificial intelligence-assisted compressed sensing single-shot fluid-attenuated inversion recovery (ACS-SS-FLAIR) for evaluation of uncooperative patients with brain diseases: comparison with the conventional T2-FLAIR with parallel imaging. Acta Radiol 2023;64:1943-9. [Crossref] [PubMed]
20. Zhai R, Huang X, Zhao Y, Ji M, Lv X, Qian M, Liao S, Li G. Intelligent Incorporation of AI with Model Constraints for MRI Acceleration. Proceedings of the 29th Annual Meeting of ISMRM, 2021.
21. Papanikolaou N, Prassopoulos P, Eracleous E, Maris T, Gogas C, Gourtsoyiannis N. Contrast-enhanced magnetic resonance cholangiography versus heavily T2-weighted magnetic resonance cholangiography. Invest Radiol 2001;36:682-6. [Crossref] [PubMed]
22. Choi MH, Kim B, Han D, Lee YJ. Compressed sensing for breath-hold high-resolution hepatobiliary phase imaging: image noise, artifact, biliary anatomy evaluation, and focal lesion detection in comparison with parallel imaging. Abdom Radiol (NY) 2022;47:133-42. [Crossref] [PubMed]
23. Bahrami K, Kot AC. A Fast Approach for No-Reference Image Sharpness Assessment Based on Maximum Local Variation. IEEE Signal Processing Letters 2014;21:751-5.
24. Mittal A, Moorthy AK, Bovik AC. No-reference image quality assessment in the spatial domain. IEEE Trans Image Process 2012;21:4695-708. [Crossref] [PubMed]
25. Moorthy AK, Bovik AC. Blind image quality assessment: from natural scene statistics to perceptual quality. IEEE Trans Image Process 2011;20:3350-64. [Crossref] [PubMed]
26. Saad MA, Bovik AC, Charrier C. Blind image quality assessment: a natural scene statistics approach in the DCT domain. IEEE Trans Image Process 2012;21:3339-52. [Crossref] [PubMed]
27. American College of Radiology. Liver Imaging Reporting and Data System. Available online: https://www.acr.org/Clinical-Resources/Clinical-Tools-and-Reference/Reporting-and-Data-Systems/LI-RADS
28. Görgec B, Hansen I, Kemmerich G, Syversveen T, Abu Hilal M, Belt EJT, et al. Clinical added value of MRI to CT in patients scheduled for local therapy of colorectal liver metastases (CAMINO): study protocol for an international multicentre prospective diagnostic accuracy study. BMC Cancer 2021;21:1116. [Crossref] [PubMed]
29. Huang P, Shi Q, Ni X, Liu L, Zhou C, Wu F, Wang C, Xiao Y, Miao G, Yang C, Zeng M. Subcentimeter hepatocellular carcinoma (HCC) on gadoxetic-acid-enhanced MRI: less frequent typical imaging features compared to 1-2 cm HCC but better prognosis after surgical resection. Abdom Radiol (NY) 2023;48:3391-400. [Crossref] [PubMed]
30. Cao M, Ding C, Xia C, Li H, Sun D, He S, Chen W. Attributable deaths of liver cancer in China. Chin J Cancer Res 2021;33:480-9. [Crossref] [PubMed]
31. Guo J, Seo Y, Ren S, Hong S, Lee D, Kim S, Jiang Y. Diagnostic performance of contrast-enhanced multidetector computed tomography and gadoxetic acid disodium-enhanced magnetic resonance imaging in detecting hepatocellular carcinoma: direct comparison and a meta-analysis. Abdom Radiol (NY) 2016;41:1960-72. [Crossref] [PubMed]
32. Choi JY, Lee JM, Sirlin CB. CT and MR imaging diagnosis and staging of hepatocellular carcinoma: part II. Extracellular agents, hepatobiliary agents, and ancillary imaging features. Radiology 2014;273:30-50. [Crossref] [PubMed]
33. Graves MJ, Mitchell DG. Body MRI artifacts in clinical practice: a physicist's and radiologist's perspective. J Magn Reson Imaging 2013;38:269-87. [Crossref] [PubMed]
34. Sui H, Gong Y, Liu L, Lv Z, Zhang Y, Dai Y, Mo Z. Comparison of Artificial Intelligence-Assisted Compressed Sensing (ACS) and Routine Two-Dimensional Sequences on Lumbar Spine Imaging. J Pain Res 2023;16:257-67. [Crossref] [PubMed]
35. Zhao Q, Xu J, Yang YX, Yu D, Zhao Y, Wang Q, Yuan H. AI-assisted accelerated MRI of the ankle: clinical practice assessment. Eur Radiol Exp 2023;7:62. [Crossref] [PubMed]
36. Wang Q, Zhao W, Xing X, Wang Y, Xin P, Chen Y, Zhu Y, Xu J, Zhao Q, Yuan H, Lang N. Feasibility of AI-assisted compressed sensing protocols in knee MR imaging: a prospective multi-reader study. Eur Radiol 2023;33:8585-96. [Crossref] [PubMed]
37. Liu H, Deng D, Zeng W, Huang Y, Zheng C, Li X, Li H, Xie C, He H, Xu G. AI-assisted compressed sensing and parallel imaging sequences for MRI of patients with nasopharyngeal carcinoma: comparison of their capabilities in terms of examination time and image quality. Eur Radiol 2023;33:7686-96. [Crossref] [PubMed]
38. Wang C, Sun WR, Wu N, Zhang Z, Zhang LX, Yi WQ, Yuan XD. A convenient and reproducible protocol for acquisition of the hepatocyte phase for liver function-impaired patients in gadoxetic acid disodium-enhanced magnetic resonance imaging. Quant Imaging Med Surg 2024;14:1904-15. [Crossref] [PubMed]
39. Hirata K, Nakaura T, Okuaki T, Tsuda N, Taguchi N, Oda S, Utsunomiya D, Yamashita Y. 3D hybrid profile order technique in a single breath-hold 3D T2-weighted fast spin-echo sequence: Usefulness in diagnosis of small liver lesions. Eur J Radiol 2018;98:113-7. [Crossref] [PubMed]
40. Stadler A, Schima W, Ba-Ssalamah A, Kettenbach J, Eisenhuber E. Artifacts in body MR imaging: their appearance and how to eliminate them. Eur Radiol 2007;17:1242-55. [Crossref] [PubMed]
41. Krupa K, Bekiesińska-Figatowska M. Artifacts in magnetic resonance imaging. Pol J Radiol 2015;80:93-106. [Crossref] [PubMed]
42. Dietrich O, Reiser MF, Schoenberg SO. Artifacts in 3-T MRI: physical background and reduction strategies. Eur J Radiol 2008;65:29-35. [Crossref] [PubMed]
43. Choi SH, Kim KW, Kwon HJ, Kim SY, Kwon JH, Song GW, Lee SG. Clinical usefulness of gadoxetic acid-enhanced MRI for evaluating biliary anatomy in living donor liver transplantation. Eur Radiol 2019;29:6508-18. [Crossref] [PubMed]
44. Lee JH, Kim B, Kim HJ, Kim BW, Wang HJ, Kim JK. High spatial resolution navigated 3D T1-weighted hepatobiliary MR cholangiography using Gd-EOB-DTPA for evaluation of biliary anatomy in living liver donors. Abdom Radiol (NY) 2018;43:1703-12. [Crossref] [PubMed]
45. Lee ES, Lee JM, Yu MH, Shin CI, Woo HS, Joo I, Stemmer A, Han JK, Choi BI. High spatial resolution, respiratory-gated, t1-weighted magnetic resonance imaging of the liver and the biliary tract during the hepatobiliary phase of gadoxetic Acid-enhanced magnetic resonance imaging. J Comput Assist Tomogr 2014;38:360-6. [Crossref] [PubMed]
46. Zhang Y, Sheng R, Yang C, Dai Y, Zeng M. Higher field reduced FOV diffusion-weighted imaging for abdominal imaging at 5.0 Tesla: image quality evaluation compared with 3.0 Tesla. Insights Imaging 2023;14:171. [Crossref] [PubMed]
47. Xu Q, Zhao H, Gao R, Wang X, Xu J, Sun G, Xue K, Yang Y, Li E, Zhu L, Wu W, Feng F. Insulinoma detection and surgery planning: a comparative study of 5.0T MRI versus 3.0T MRI and MDCT. Abdom Radiol (NY) 2024; Epub ahead of print. [Crossref]
48. Zhao H, Xu Q, Gao R, Yin B, Sun G, Xue K, Yang Y, Li E, Zhu L, Feng F, Wu W. Clinical Feasibility of 5.0 T MRI/MRCP in Characterizing Pancreatic Cystic Lesions: Comparison with 3.0 T and MDCT. Diagnostics (Basel) 2024.
49. Yin L, Li Z, Shang M, Li Z, Tang B, Yu D, Gan J. Magnetic resonance cholangiopancreatography at 5.0 T: quantitative and qualitative comparison with 3.0 T. BMC Med Imaging 2024;24:331. [Crossref] [PubMed]
50. Jiang Z, Sun W, Xu D, Yu H, Mei H, Song X, Xu H. Stability and repeatability of diffusion-weighted imaging (DWI) of normal pancreas on 5.0 Tesla magnetic resonance imaging (MRI). Sci Rep 2023;13:11954. [Crossref] [PubMed]
51. Liu J, Wang Z, Yu D, Yang Y, Li Z, Wang X, Yang Y, Cheng C, Zou C, Gan J. Comparative analysis of hepatic fat quantification across 5 T, 3 T and 1.5 T: A study on consistency and feasibility. Eur J Radiol 2024;180:111709. [Crossref] [PubMed]
52. Liu LL, Lin YK, Xiang ZL. The Value of Primary Tumor Resection in Patients with Liver Metastases: A 10-Year Outcome. Ann Surg Oncol 2025;32:1083-92. [Crossref] [PubMed]
53. Benson AB, Venook AP, Al-Hawary MM, Arain MA, Chen YJ, Ciombor KK, et al. Colon Cancer, Version 2.2021, NCCN Clinical Practice Guidelines in Oncology. J Natl Compr Canc Netw 2021;19:329-59. [Crossref] [PubMed]