The institutional review board at our institution approved this retrospective study and waived the requirement of written informed consent. We searched the radiologic database of consecutive adult (≥ 18 years) patients who underwent gadoxetic acid–enhanced liver MRI using a 3-T scanner (MAGNETOM Skyra; Siemens Healthcare) between December 20, 2022, and January 20, 2023. Of 180 patients who were initially enrolled, 12 were excluded from the study owing to incomplete liver MR sequences for analysis. Therefore, the final cohort comprised 168 patients (104 men, 64 women; mean age ± standard deviation, 62.1 ± 12.9 years; range 19–88 years). Clinical indications for liver MRI were (a) HCC surveillance (n = 120); (b) metastasis surveillance (n = 30); and (c) focal liver lesion (FLL) characterization (n = 18).

To ensure a more focused review for solid liver lesions, we focused on patients with ≤ 5 solid FLLs to evaluate lesion conspicuity, excluding nonsolid FLLs, such as treated HCC lesions, benign cysts, typical hemangiomas, and arterioportal shunts. The MRI scans with no solid FLLs (n = 92) or more than 5 solid FLLs (n = 23) were excluded. Accordingly, 87 solid FLLs detected in 53 patients were included for lesion conspicuity evaluation. The lesion diagnoses were HCCs (n = 43), dysplastic nodules (n = 17), metastases (n = 9), focal nodular hyperplasia (FNH) (n = 5), hepatocellular adenomas (n = 4), FNH-like lesions (n = 2), benign lesions (n = 2), inflammatory lesions (n = 2), intrahepatic cholangiocarcinoma (n = 1), angiomyolipoma (n = 1), and sclerosing hemangioma (n = 1) (Fig. 1). HCCs were diagnosed based on pathologic examinations or imaging criteria for diagnosing HCC of Korean Liver Cancer Association-National Cancer Center guideline [22]. The diagnoses of metastases were established according to characteristic MRI findings, e.g., irregular or ill-defined margins, rim enhancement on MR dynamic images, hypoenhancement or targetoid appearance on hepatobiliary phase (HBP) images, and exhibiting greater than 20% interval growth on serial cross-sectional imaging in patients with underlying malignancy [23, 24]. Dysplastic nodules were diagnosed based on typical MRI features, like iso- or hyper-intensity on T1-weighted imaging, slight hypointensity on T2-weighted imaging, no arterial phase hyperenhancement (APHE), and isointensity or slight hypointensity on HBP, as well as stability on follow-up cross-sectional imaging [25, 26]. FNHs were diagnosed based on characteristic MRI findings, including homogeneous APHE, central scar, no “washout”, and iso- or hyper-enhancement on HBP, and stable findings on follow-up cross-sectional imaging [24, 27].

Study flowchart. CAIPIRINHA, controlled aliasing in parallel imaging results in a higher acceleration; FLLs, focal liver lesions; HBP, hepatobiliary phase; MRI, magnetic resonance imaging; VIBE, volume-interpolated breath-hold examination

MRI examinations were performed on a 3-T scanner (MAGNETOM Skyra; Siemens Healthcare). Routine liver MRI protocols involved T1-weighted dual-echo imaging, pre-contrast and gadoxetic acid–enhanced dynamic and HBP imaging, T2-weighted imaging, and diffusion-weighted imaging using three b values (50, 400, and 800 s/mm2). A standard dose of 0.025 mmol/kg of contrast agent (Primovist; Bayer Healthcare) was administered at a rate of 1.5 mL/s followed by 25 mL saline flush. For dynamic imaging, triple arterial phase, portal venous phase, transitional phase, and HBP were obtained using a spectrally fat-suppressed 3D VIBE after the injection. The timings for arterial phase (AP) imaging were determined by a real-time bolus-tracking technique with MR fluoroscopic monitoring. All patients underwent the following VIBE protocols: (a) a standard CAIPIRINHA-VIBE scanning and an additional CAIPIRINHA-VIBE with DL reconstruction scanning for pre-contrast images; and (b) a standard CAIPIRINHA-VIBE scanning and an additional CAIPIRINHA-VIBE with DL and high-resolution (HR) DL, respectively, reconstruction scanning for hepatobiliary phase (HBP) images. MRI acquisition parameters are detailed in Table 1. As the HR protocol was acquired in the HBP with more signal, a higher acceleration factor was chosen. The DL CAIPIRINHA acquisitions employed a more efficient sampling scheme for spectral fat suppression. While DL sequences were added to our protocol, they effectively replaced the “standard” CS sequences in practice.

The DL-based image reconstruction involved two sequential, independent processing steps (Fig. 2).

Schematic flow of DL-based reconstruction algorithm. Input and processing steps of the DL-based reconstruction algorithm in the upper row. The lower left diagram illustrates the underlying concept of the k-space to image reconstruction, which alternates between a conventional parallel imaging reconstruction followed by the estimation of a prior image using a neural network U. The conventional reconstruction corresponds to a linear optimization with elliptic hypersurfaces that is pursued from the current prior image with a stepsize λ. The lower right diagram depicts layers of the network architecture used in the super-resolution algorithm. DL, deep learning

In the first step, images were reconstructed from k-space data on the acquired resolution using a network architecture inspired by variational networks [28]. As input, the architecture received undersampled k-space as well as coil sensitivity maps estimated as a preposing step from separately acquired calibration scans. Images were then determined by 6 iterations consisting of a data consistency update in the form of a parallel imaging reconstruction followed by a neural network evaluation for image regularization. Limiting to conventional PAT sampling patterns had the advantage that estimated coil sensitivity maps can be optimized for a given acceleration. Furthermore, as the aliasing in image space was coherent, the training data can be cropped to smaller sizes and thereby allow for a supervised training with image regularization networks acting in all spatial dimensions. The network architecture was implemented in PyTorch [29] and a supervised training performed using about 5000 training pairs derived from about 500 fully sampled 3D datasets acquired from healthy volunteers on 1.5- and 3-T scanners (MAGNETOM scanners, Siemens Healthcare) in the head, abdomen, and pelvis. In alignment with data consistency principles, the network was tailored to enhance local image features, and as with clinically validated 2D methods, it was not expected to be sensitive to the content of the image [30]. A conventional 3D U-net [31] was used for the image regularization networks, and L1 was chosen as loss function and Adam [32] as optimizer. The obtained network was then exported in the ONNX format and integrated into the scanner reconstruction pipeline using the ONNX Runtime [33] as inference engine. Prospective execution time for this processing step was about 15 s for the employed 3D T1-weighted protocols utilizing the scanner integrated graphical processing units.

The second processing step interpolated the acquired images using a DL-based super-resolution algorithm as outlined in prior studies [18, 21]. The employed algorithm performed an initial upsampling by a factor of 2 in all spatial dimensions and was furthermore trained to perform a partial Fourier reconstruction in slice direction, consistent with the chosen acquisition protocol. The dataset utilized for supervised training comprised of high-resolution images, which served as ground-truth images. The input data used in the training were obtained by downsampling these images by a factor of 2 in all spatial dimensions.

Both processing steps were integrated into a research application for prospective use in the scanner reconstruction pipeline.

All de-identified MR images were independently reviewed by three abdominal radiologists (J.W.C., J.L., and S.K.J.) with 6, 6, and 10 years of experience in abdominal MRI, respectively, who were blinded to the MRI acquisition techniques. The readers underwent a short training session for interpretations and scores of all assessed parameters before initiating image analysis. All MR images of interest (i.e., pre-contrast and HBP images), with either the standard CAIPIRINHA-VIBE or DL CAIPIRINHA-VIBE, and HR-DL CAIPIRINHA-VIBE data sets, were randomly distributed to readers.

Image quality was qualitatively evaluated on pre-contrast and HBP axial images, in terms of liver edge sharpness, hepatic vessel conspicuity, bile duct conspicuity (only on HBP image), respiratory motion artifact, cardiac ghosting artifact, ringing artifact, perceived SNR, subjective noise level, synthetic appearance, overall artifact level, and overall image quality on a 4-point scale (Table 2) [13, 34,35,36]. A higher score implies sharper liver edge, better conspicuity of hepatic vessel and bile duct, less artifact, higher SNR, less noise, less synthetic appearance, and better image quality.

One researcher (H.W.) with 5 years of experience in abdominal MRI who did not participate in the review session recorded the information of FLLs (i.e., lesion number, size, location, and radiological diagnosis) by reviewing MRI reports and all available clinical information and radiological examinations. All this information was provided to readers for lesion localization. Lesion conspicuity was evaluated on pre-contrast and HBP axial images according to a 4-point scale (Table 2). To exclude nonsolid lesions and treated HCC lesions, matched T2-weighted images and AP images were provided to readers. A higher score indicates better lesion conspicuity.

For lesion detection analysis, lesions with conspicuity scores of 2–4 were defined as detected, while those with conspicuity scores of 1 (not visible) were defined as undetected [34]. Lesion detection rate was calculated by the number of detected solid FLLs divided by the number of total solid FLLs.

Wilcoxon signed-rank test was used for pairwise comparisons of the image quality scores, which were averaged across 3 readers. Interobserver agreement was assessed using the Gwet’s AC1 coefficient [37], as follows: 0.01–0.20, slight agreement; 0.21–0.40, fair agreement; 0.41–0.60, moderate agreement; 0.61–0.80, substantial agreement; and 0.81–1.00, almost perfect agreement. For statistical analyses, the conspicuity scores and number of lesions were considered the sum of observations of 3 readers. Based on the pooled data, lesion conspicuity and detection rate were evaluated by the generalized estimation equation method [38]. Statistical analyses were performed using the R software (version 4.3.1; The R Foundation for Statistical Computing), SAS software (version 9.4; SAS institute), and jackknife free-response receiver operating characteristic software (version 4.2.1). Two-tailed p ≤ 0.05 was indicated statistically significant.