Deep learning-based high-resolution united compressed sensing for gadoxetic acid-enhanced liver magnetic resonance imaging in the detection of colorectal liver metastases

Introduction

Liver metastases are a major determinant of patient prognosis, and their early detection is critical for optimizing treatment planning and improving survival outcomes (1). Among imaging modalities, gadolinium ethoxybenzyl diethylenetriamine pentaacetic acid-enhanced magnetic resonance imaging (Gd-EOB-DTPA-MRI), particularly during the hepatobiliary phase (HBP), has shown superior performance in the detection of small focal liver lesions due to hepatocyte-specific contrast uptake and the resulting high lesion-to-liver contrast (2,3). However, conventional magnetic resonance imaging (MRI) enhancement techniques remain limited by acquisition efficiency, image resolution, and noise control. The three-dimensional gradient echo (3D GRE) sequence, which serves as the fundamental sequence for contrast-enhanced liver MRI, plays a significant role in liver imaging. However, the routine slice thickness of 2.5–3 mm in the 3D GRE sequence limits the accurate assessment of sub-centimeter lesions. Thinner slice scans are helpful, but they are often accompanied by a reduction in the signal-to-noise ratio (SNR) and an increase in artifacts, which severely affect image quality.

Recent advancements in united compressed sensing (uCS) imaging have led to significant progress in MRI technology (4-6). uCS is a rapid imaging technique based on compressed sensing (CS) combined with half-Fourier and parallel imaging techniques. By using sparse sampling to improve acquisition efficiency and employing high-performance gradient systems, uCS significantly enhances the SNR and contrast-to-noise ratio (CNR), thus improving image clarity and detail (7). However, noise contamination during acquisition remains unavoidable, particularly when evaluating small lesions, which may hinder accurate lesion characterization and reduce diagnostic confidence.

These limitations highlight the clinical need for novel imaging strategies that can achieve both high spatial resolution and effective noise suppression to support the reliable early detection of liver lesions, including metastases. Colorectal liver metastases (CRLMs) are the most common type of malignant liver metastases and are of major clinical importance, as their detection directly influences surgical planning and patient outcomes. However, the accurate identification of CRLMs presents unique imaging challenges that demand high-resolution techniques.

CRLMs frequently present as sub-centimeter lesions with subtle, poorly defined margins. When evaluated using conventional slice thicknesses, severe partial volume effects can obscure these small lesions or blur their margins, making them difficult to distinguish from background liver parenchyma or benign hepatic nodules. Therefore, achieving higher spatial resolution is particularly critical for CRLMs, as it minimizes partial volume averaging, sharpens lesion boundaries, and significantly enhances the conspicuity of micro-metastases. Further, evaluating a relatively homogeneous lesion type provides methodological consistency, thereby allowing a more accurate assessment of imaging technologies. For these reasons, CRLM was chosen as the disease model in this study.

Deep learning-based reconstruction (DLR) is an innovative technology developed based on artificial intelligence (AI). It enables intelligent image denoising and super-resolution imaging, thereby optimizing image quality (8). Several studies have shown significant progress in the application of DLR technologies for brain tumor, head and neck, breast, and abdominal imaging (9-12). In liver imaging, recent reviews have further highlighted the growing role of AI in lesion detection, characterization, and quantitative assessment, underscoring the increasing maturity and clinical applicability of AI-based techniques in hepatic MRI (13-15). Compared with post-processing-based AI applications, DLR may offer more direct advantages for improving image quality and lesion conspicuity in Gd-EOB-DTPA-MRI of the liver. However, its application in Gd-EOB-DTPA-MRI of the liver remains relatively underexplored. Takayama et al. demonstrated that DLR significantly improved image quality and tumor detectability in HBP 3D T1-weighted imaging (16). Similarly, Castagnoli et al. showed that AIaugmented reconstructions improved image quality and reduced breath-hold duration in HBP acquisitions without compromising lesion detection (17).

In the present study, we specifically evaluated the performance of deep learning-based reconstruction united compressed sensing (DR-uCS) and deep learning-based reconstruction high-resolution united compressed sensing (DR-HR-uCS) in patients with CRLMs. Our objective was to determine whether these techniques can improve image quality and enhance the conspicuity of sub-centimeter liver lesions using 1.25‑mm slice thickness during the HBP. We present this article in accordance with the STARD-AI reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1-2540/rc).

Methods

Study participants

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This retrospective study was approved by the Medical Ethics Committee of Henan Cancer Hospital (No. 2023-KY-0089-001), which waived the requirement of informed consent because this study was retrospective in nature, and did not involve any additional interventions or procedures beyond routine clinical care.

This study included patients with CRLM because it is one of the most common malignant liver tumors, and the accurate detection of small CRLM lesions is crucial for clinical management and prognosis. Restricting the study cohort to a single, clinically relevant lesion type also ensured methodological consistency and reduced heterogeneity in the evaluation of the diagnostic performance of DR-uCS technology. A retrospective analysis was conducted on 86 patients with CRLM who underwent Gd-EOB-DTPA-MRI of the upper abdomen at Henan Cancer Hospital between June and December 2024. Among the patients, 50 were male and 36 were female, with an age range of 18 to 83 years (mean age, 47.1±16.1 years). The inclusion criteria were as follows: (I) a definitive pathological diagnosis of colorectal cancer; (II) suspicion of liver metastasis based on ultrasound or computed tomography (CT) findings; (III) imaging evaluation prior to surgery or local treatment; and (IV) use of a liver-specific contrast agent (Gd-EOB-DTPA). The exclusion criteria were as follows: (I) a history of or concurrent primary malignancies at other sites; (II) primary liver fibrosis or cirrhosis; and/or (III) significant respiratory motion artifacts impairing the diagnostic quality of the HBP images.

MRI acquisition

All examinations were performed using a consistent coil configuration comprising a 12-channel body coil combined with a 32-channel spine coil, which was applied uniformly across all patients during the study period. Patients were positioned supine with head-first entry. Prior to scanning, patients were instructed on breath-holding techniques.

A standardized gadoxetic acid-enhanced liver MRI protocol was applied to all patients. The contrast agent (gadoxetic acid disodium, Primovist; Bayer Healthcare, Berlin, Germany) was administered intravenously at a dose of 0.025 mmol/kg at a rate of 1.5 mL/s, followed by a 25 mL saline flush. Post-contrast imaging was performed at multiple time points, including during the arterial phase (~20–30 s), portal venous phase (~60–70 s), transitional phase (~3 min), and HBP (approximately 20 min after injection). The quantitative and qualitative analyses in this study were performed using HBP images only.

Two acquisition protocols were used for HBP imaging. A standard-resolution acquisition was used to generate the conventional uCS and DR-uCS image sets. These two reconstructions were derived from the same raw k-space dataset and acquired during the same breath-hold. In addition, a separate high-resolution HBP acquisition was performed during an additional breath-hold to generate the DR-HR-uCS image set. The standard-resolution acquisition used an in-plane spatial resolution of approximately 1.39 mm × 1.20 mm with a slice thickness of 2.5 mm, while the high-resolution acquisition used isotropic spatial resolution of approximately 1.25 mm × 1.25 mm × 1.25 mm. Detailed acquisition parameters are summarized in Table 1.

The reconstruction methods evaluated in this study included conventional uCS, DR-uCS, and DR-HR-uCS. The DR-uCS and DR-HR-uCS algorithms were implemented as vendor-provided commercial reconstruction tools integrated into the routine clinical workflow. No model training, fine-tuning, or parameter adjustment was performed using the study data. According to vendor technical documentation, the reconstruction algorithms are based on convolutional neural network architectures trained offline on large proprietary MRI datasets that include both phantom and human imaging data acquired across multiple anatomical regions and imaging conditions. The training data were independent of the present study cohort, and no overlap existed between the training and testing datasets. Because the algorithm is commercially proprietary, detailed information regarding the exact size and full composition of the training dataset is not publicly available.

For the reader study, all image sets were anonymized and presented in randomized order. The conventional uCS reconstruction was performed using vendor-default clinical parameters without additional manual optimization. Because the uCS reconstruction was implemented through a commercial vendor platform, the exact internal regularization parameters were proprietary and were not accessible or user-configurable through the clinical interface. No parameter tuning was performed for any of the reconstruction methods, including uCS, DR-uCS, and DR-HR-uCS, to reflect routine clinical practice and reduce potential bias related to parameter selection. All image sets were anonymized, randomized, and evaluated under identical reading conditions. However, as DR-HR-uCS was obtained from a separate high-resolution acquisition protocol, the comparison should be interpreted as a clinically relevant comparison under routine workflow conditions rather than as a fully controlled reconstruction-only comparison.

Image quality evaluation

The HBP images from the three groups were transferred to the United Imaging uWS-MR workstation (version R005).

Subjective evaluation

Two board-certified abdominal radiologists (Reader 1: Y.W., with 18 years of experience; Reader 2: D.S., with 10 years of experience) independently evaluated the anonymized image sets. The readers were blinded to the reconstruction techniques and to each other’s evaluations. To minimize interpretation bias, the image sets reconstructed by uCS, DR-uCS, and DR-HR-uCS were presented in randomized order. A washout period of 2 weeks was applied between reading sessions for different reconstruction techniques in the same patient to reduce recall bias. Image quality and artifact severity were independently assessed using standardized criteria, and the average scores from both readers were used for analysis. Qualitative image evaluation on HBP axial images included image quality, artifact severity, and liver edge clarity/liver vessel clarity. A higher score indicated clearer liver edges, better liver vessel clarity, fewer artifacts, and better image quality (18). Subjective image quality (overall image quality, artifact severity, and liver edge/vessel clarity), as well as lesion conspicuity, were evaluated using standardized Likert scales, as summarized in Table 2. Lesion conspicuity was assessed using a four-point Likert scale instead of the standard five-point scale to avoid a neutral midpoint and encourage more decisive interpretation, thereby improving inter-reader agreement (19,20). A formal intra-reader variability analysis was not performed, as the primary purpose of the reader study was to compare diagnostic performance among reconstruction methods rather than to assess reader reproducibility. Artifact severity was assessed based on the overall image series for each reconstruction method rather than on a single representative slice or figure panel.

Objective evaluation

Quantitative measurements were independently performed by two abdominal radiologists (C.X., with 18 years of experience, and J.Y., with 10 years of experience) using a standardized region-of-interest (ROI) protocol. Circular or oval ROIs of similar size were placed within the lesion and adjacent liver parenchyma while avoiding blood vessels, bile ducts, artifacts, lesion margins, and necrotic or non-enhancing components to reduce partial volume effects. For the SNR measurements, three ROIs were drawn within homogeneous areas of the liver parenchyma (segment VI or VII) and paraspinal muscle (erector spinae) while avoiding large vessels, bile ducts, and artifacts. For the CNR measurements, lesion ROIs were manually placed within the most conspicuous area of the metastases. Three repeated measurements were obtained for each dataset, and the mean value for each reader was recorded. Inter-reader reliability was assessed using intraclass correlation coefficients (ICCs), and the average of the two readers’ measurements was used for the final quantitative analysis. All ROIs were approximately 200 mm² in size (Figure 1). The following formulas were used:

• SNR for normal liver = signal intensity (SI) of normal liver/standard deviation (SD) of muscle;
• SNR for lesion = SI of lesion/SD of muscle;
• CNR = (SI of normal liver − SI of lesion)/SD of muscle.

Figure 1 Representative axial hepatobiliary phase liver MRI images reconstructed using three different techniques: (A) conventional uCS, (B) DR-uCS, and (C) DR-HR-uCS. The white arrow indicates the ROI in the adjacent liver parenchyma, the yellow arrow indicates the lesion ROI, and the blue arrow indicates the paraspinal muscle ROI. Signal intensities from the lesion and liver parenchyma ROIs were used to calculate lesion and liver signal, while the standard deviation of the paraspinal muscle ROI served as the noise reference for the SNR and CNR measurements. CNR, contrast-to-noise ratio; DR-HR-uCS, deep learning-based reconstruction high-resolution united compressed sensing; DR-uCS, deep learning-based reconstruction united compressed sensing; MRI, magnetic resonance imaging; ROI, region of interest; SNR, signal-to-noise ratio; uCS, united compressed sensing.

Diagnostic performance analysis: CRLM identification

The diagnosis of CRLM was primarily based on characteristic MRI features, including irregular or indistinct lesion margins, ring enhancement on dynamic imaging, and hypointensity or target-like changes on HBP images. Additionally, for patients with a known primary malignancy, lesion progression (defined as an increase in lesion size of more than 20% across two consecutive follow-up cross-sectional imaging studies) further supported the diagnosis. The number of liver metastases was recorded, and the diameter of metastatic lesions was measured across the three sets of HBP images.

The reference standard for CRLM was established using histopathology when available. For lesions without pathological confirmation, the final diagnosis was determined by multidisciplinary consensus on the basis of characteristic findings on gadoxetic acid-enhanced MRI, prior imaging examinations (including CT, MRI, and positron emission tomography-CT when available), serum tumor markers, clinical information, and longitudinal follow-up. Two experienced specialists (C.X., a radiologist with 18 years of experience in abdominal imaging, and J.Y., an oncologist with 9 years of experience in gastrointestinal malignancies) independently reviewed the available data and classified lesions as CRLM or non-CRLM. Inconsistencies were resolved by joint review and consensus. Non-CRLM lesions included benign hepatic lesions, such as cysts and hemangiomas, as determined by imaging characteristics and follow-up.

Histopathological confirmation was available for 68 lesions in 52 patients. For the remaining lesions, the diagnosis was established based on imaging and clinical follow-up, with a median follow-up duration of 9 months (range, 6–15 months). Lesions classified as non-CRLM according to the final reference standard served as negative controls for the specificity analysis.

Assessment of diagnostic time across reader experience levels

To evaluate the influence of clinical experience on the efficiency of CRLM detection, diagnostic time was measured using three radiologists representing different experience levels: junior (<5 years of experience), intermediate (5–10 years), and senior (>10 years). All readers independently reviewed anonymized imaging cases containing liver lesions on the same standardized workstation, and were blinded to clinical, pathological, and follow-up information. The image sets were presented in randomized order, and separate reading sessions with washout intervals were used to minimize recall bias across reconstruction methods.

Diagnostic time was defined as the interval from opening the image set to completion of lesion documentation, and was recorded in minutes using a screen-based timer. For each reader and each reconstruction method, the mean diagnostic time per case was calculated.

Diagnostic performance was assessed separately for lesions ≥1 cm and lesions <1 cm. Sensitivity and specificity were calculated on a per-lesion basis using the final reference standard as the ground truth. Lesions classified as non-CRLM according to the final reference standard served as negative controls for the specificity analysis.

Statistical analysis

The statistical analysis was performed using SPSS version 24.0 (IBM Corp.). Normality was assessed using the Shapiro-Wilk test. Data with a normal distribution were expressed as mean ± SD, while data with a non-normal distribution were presented as median [interquartile range (IQR)].

Subjective image assessments among the three reconstruction methods were compared using the Friedman test, followed by pairwise Wilcoxon signed-rank tests with Bonferroni correction. Quantitative measurements of SNR and CNR were compared using repeated-measures analysis of variance for normally distributed variables or the Friedman test for non-normally distributed variables. Diagnostic time was compared among reconstruction methods within each reader and among reader experience levels for each reconstruction method using appropriate parametric or nonparametric tests with Bonferroni-adjusted pairwise comparisons. Effect sizes (Cohen’s d values) were calculated for key comparisons, including SNR, CNR, and diagnostic time.

Comparisons of sensitivity and specificity among reconstruction methods were performed using the McNemar test for paired data. Inter-reader agreement for qualitative assessments (image quality, artifact severity, liver edge/vessel clarity, and lesion conspicuity) was evaluated using Cohen’s kappa statistics. For quantitative measurements (SNR and CNR), inter-reader reliability was assessed using ICCs. Agreement was interpreted as poor (0.00–0.20), fair (0.21–0.40), moderate (0.41–0.60), good (0.61–0.80), or excellent (0.81–1.00). A two-sided P value <0.05 was considered statistically significant.

Results

A total of 86 patients were included in the final analysis. Among them, 50 (58.1%) were male and 36 (41.9%) were female, with a mean age of 47.1±16.1 years (range, 18–83 years). All patients had pathologically confirmed colorectal cancer. A total of 116 lesions were classified as CRLMs, including 71 lesions ≥1 cm and 45 lesions <1 cm. Histopathological confirmation was available for 68 lesions in 52 patients. For the remaining lesions, the diagnosis was established by multidisciplinary consensus based on imaging findings, clinical information, and follow-up. In addition to the 116 CRLM lesions, 34 non-CRLM lesions were included as negative controls for the specificity analysis.

Overall image quality scores

The average image quality scores for uCS, DR-uCS, and DR-HR-uCS were 4 (IQR, 3, 4), 5 (IQR, 3, 5), and 5 (IQR, 4, 5), respectively. Both the DR-uCS and DR-HR-uCS groups had higher scores than the uCS group, and the difference was statistically significant (P<0.001). However, no statistically significant difference was observed between the DR-uCS and DR-HR-uCS groups (P=0.809) (Table 3 and Figure 2).

Figure 2 Hepatobiliary phase gadoxetic acid-enhanced MRI images reconstructed using conventional uCS (A), DR-uCS (B), and DR-HR-uCS (C) in a 45-year-old man with a focal liver lesion. DR-uCS and DR-HR-uCS demonstrated improved delineation of the liver margin (white arrow) and intrahepatic vessels (yellow arrow) compared with conventional uCS. DR-HR-uCS, deep learning-based reconstruction high-resolution united compressed sensing; DR-uCS, deep learning-based reconstruction united compressed sensing; MRI, magnetic resonance imaging; uCS, united compressed sensing.

Image artifact scores

The average artifact scores for uCS, DR-uCS, and DR-HR-uCS were 4 (IQR, 3, 4), 5 (IQR, 3, 5), and 4 (IQR, 4, 5), respectively. Both the DR-uCS and DR-HR-uCS groups exhibited higher scores than the uCS group, and the difference was statistically significant (P<0.001). No statistically significant difference was observed between the DR-uCS and DR-HR-uCS groups (P=0.776) (Table 3 and Figure 2).

Liver edge sharpness/hepatic vessel conspicuity

The average liver edge sharpness/hepatic vessel conspicuity scores for uCS, DR-uCS, and DR-HR-uCS were 4 (IQR, 3, 4), 5 (IQR, 3, 5), and 5 (IQR, 4, 5), respectively. Both DR-uCS and DR-HR-uCS showed significantly higher liver edge sharpness and hepatic vessel conspicuity scores than uCS, and the difference was statistically significant (P=0.03, 0.001, respectively). However, no statistically significant difference was observed between the DR-uCS and DR-HR-uCS groups (P=0.669) (Table 3 and Figure 2).

Quantitative analysis

The results of the SNR and CNR assessment of the three groups of HBP images are shown in Table 4. In terms of the liver SNR, the values were higher in the DR-uCS and DR-HR-uCS groups than in the conventional uCS group (P=0.001, 0.02, respectively), but comparisons between the DR-uCS group and the DR-HR-uCS group revealed no statistically significant differences (P>0.05). In terms of the lesion SNR, no statistically significant differences were observed in the pairwise comparisons between the uCS, DR-uCS, and DR-HR-uCS groups (P>0.05). In terms of the CNR, the DR-uCS and DR-HR-uCS groups had higher values than the conventional uCS group (P=0.002, Cohen’s d=1.45; P=0.04, Cohen’s d=1.12, respectively), but comparisons between the DR-uCS and DR-HR-uCS groups revealed no statistically significant differences (P>0.05) (Figure 3).

Figure 3 Hepatobiliary phase gadoxetic acid-enhanced MRI images reconstructed using conventional uCS (A), DR-uCS (B), and DR-HR-uCS (C) in a 58-year-old man with a focal liver lesion. DR-HR-uCS provided the sharpest depiction of the liver margin (white arrow) and intrahepatic vessels (yellow arrow), with fewer visible artifacts. DR-HR-uCS, deep learning-based reconstruction high-resolution united compressed sensing; DR-uCS, deep learning-based reconstruction united compressed sensing; MRI, magnetic resonance imaging; uCS, united compressed sensing.

Lesion conspicuity

For lesions ≥1 cm, no statistically significant differences in lesion conspicuity were observed among the uCS, DR-uCS, and DR-HR-uCS groups (P>0.05) (Figure 4). However, for lesions <1 cm, the lesion conspicuity score in the DR-HR-uCS group was significantly higher than that in both the uCS and DR-uCS groups, and the differences were statistically significant (P<0.001, 0.002, respectively) (Table 5 and Figure 5).

Figure 4 Hepatobiliary phase gadoxetic acid-enhanced MRI images reconstructed using conventional uCS (A), DR-uCS (B), and DR-HR-uCS (C) in a 55-year-old man with colorectal liver metastases (white arrows). For lesions ≥1 cm, no significant differences in lesion conspicuity were observed among the three reconstruction methods. DR-HR-uCS, deep learning-based reconstruction high-resolution united compressed sensing; DR-uCS, deep learning-based reconstruction united compressed sensing; MRI, magnetic resonance imaging; uCS, united compressed sensing.

Figure 5 Hepatobiliary phase gadoxetic acid-enhanced MRI images reconstructed using conventional uCS (A), DR-uCS (B), and DR-HR-uCS (C) in a 46-year-old female with liver metastasis from colorectal cancer (white arrows). For lesions <1 cm, the lesion conspicuity score in the DR-HR-uCS group was significantly higher than in both the uCS and DR-uCS groups. DR-HR-uCS, deep learning-based reconstruction high-resolution united compressed sensing; DR-uCS, deep learning-based reconstruction united compressed sensing; MRI, magnetic resonance imaging; uCS, united compressed sensing.

Diagnostic performance and reading time for lesions <1 cm

For lesions ≥1 cm, all three reconstruction methods showed high diagnostic performance. The sensitivity was 95.7% for uCS, 96.6% for DR-uCS, and 97.4% for DR-HR-uCS, with corresponding specificities of 94.1%, 94.8%, and 95.6%, respectively. No statistically significant differences were observed among the three methods (all P>0.05). For lesions <1 cm, DR-HR-uCS achieved the highest sensitivity (39/45. 87.1%), which was significantly greater than that of both uCS (72.4%) and DR-uCS (78.6%) (P=0.012 vs. uCS; P=0.036 vs. DR-uCS). The specificity of DR-HR-uCS (93.5%) was comparable to that of uCS (92.7%) and DR-uCS (93.1%) (all P>0.05). These findings indicate that DR-HR-uCS provides a clear advantage in detecting sub-centimeter lesions while maintaining diagnostic reliability.

Compared to uCS, the use of DR-HR-uCS significantly reduced the diagnostic time for CRLMs <1 cm across all reader experience levels (P<0.001) (Table S1). The effect sizes for senior versus junior comparisons were large to very large, with Cohen’s d values of 3.47 (uCS) and 6.20 (DR-uCS), respectively. Similarly, the senior versus intermediate comparisons yielded d values of 1.91 (uCS) and 2.68 (DR-uCS), suggesting a strong influence of reader experience on diagnostic time. Conversely, on the DR-HR-uCS images, the diagnostic times were consistently shorter across all experience levels, with no significant differences observed among the reader groups (P>0.05) (Figure S1).

Consistency analysis of subjective scores and objective measurements between the observer groups

The inter-observer agreement analysis for both subjective assessments and objective measurements demonstrated good to excellent consistency across all reconstruction methods (Table S2). For the qualitative parameters, Cohen’s kappa values ranged from 0.765 to 0.899, indicating substantial to almost perfect agreement. The highest agreement was observed in overall image quality under the uCS method (κ=0.899), while hepatic vessel conspicuity showed slightly lower agreement (κ=0.765) under DR-uCS. For the quantitative measurements, the ICCs ranged from 0.740 to 0.911, reflecting good to excellent inter-reader reliability. The highest ICC was observed for the liver SNR under uCS (ICC =0.911), while the CNR under DR-HR-uCS showed the lowest value (ICC =0.740).

Discussion

Gd-EOB-DTPA-MRI of the liver is one of the most sensitive imaging methods for detecting liver metastases from colorectal cancer. Gd-EOB-DTPA provides strong and sustained hepatocyte-specific enhancement during the HBP, increasing the contrast between the liver parenchyma and lesions, and thereby facilitating lesion detection (18). A previous study (21) optimized liver 3D GRE scanning sequences using uCS technology. However, as the acceleration factor increases, gains in acceleration efficiency become limited, and the typical slice thickness of 2.5–3 mm still carries the risk of missing sub-centimeter liver metastases. The results of this study show that DR-uCS images significantly outperform conventional uCS images in terms of SNR, CNR, overall image quality, severity, and lesion visibility.

The DLR technology evaluated in this study was implemented as a vendor-provided commercial tool integrated into the clinical workflow. This technology enables the transformation of low-resolution, low-SNR raw image data acquired using uCS into high-resolution, high-SNR images with richer details, without increasing the scan time. These findings demonstrate that DR-uCS technology significantly improves the quality of liver Gd-EOB-DTPA-enhanced HBP images. Additionally, 1.25-mm isotropic DR-HR-uCS imaging offers greater value in the diagnosis and differential diagnosis of sub-centimeter liver metastases in colorectal cancer.

The quality of liver images is influenced by complex anatomical structures, requiring high spatial resolution and SNR for clearer visualization of anatomical tissues and potential lesions. Conventional CS under high acceleration factors struggles to effectively denoise and recover fine details, thus compromising image quality and diagnostic performance (22,23). Conversely, DLR, trained on large-scale datasets, can cover a broader range of noise distributions, allowing for the adaptive selection of noise models, noise removal, and accurate signal recovery, leading to superior reconstruction results (24,25).

The results of this study demonstrated that DR-uCS significantly improves the SNR and CNR of 3D GRE sequence images compared to conventional uCS. Previous studies (24) have also reported improvements in SNR and CNR in 1.5T liver imaging using DLR technology, which may contribute to faster imaging time selection and improved image quality. However, in the results of this study, no statistically significant differences were observed in the lesion SNRs among the three groups, likely due to the lack of normal hepatocytes in liver metastases, which cannot take up the contrast agent and thus appear as low signal in the HBP.

Additionally, images from the DR-uCS and DR-HR-uCS groups outperformed conventional uCS images in terms of artifact severity scores. Although DR-HR-uCS employs higher spatial resolution with thinner slice thickness (1.25 mm), no statistically significant improvement in subjective sharpness was observed compared with DR-uCS. This finding may be explained by several factors. First, perceived sharpness is influenced not only by spatial resolution but also by noise characteristics and contrast. DLR emphasizes noise suppression and structural consistency, which may reduce edge enhancement compared to conventional reconstruction. Second, CS reconstruction involves a trade-off between noise suppression and preservation of high-frequency details, which may affect the visual appearance of sharpness depending on the parameter settings. Finally, the use of a Likert-based subjective scoring system may limit sensitivity in detecting subtle differences in edge definition. Therefore, higher spatial resolution does not necessarily translate into higher perceived sharpness in subjective evaluation.

Previous research has demonstrated the significant potential of uCS in reducing artifacts; however, Gibbs artifacts remain an inevitable phenomenon due to K-space acquisition and the inherent imaging principles (26). In traditional methods, Gibbs artifacts are typically processed through K-space filtering. In this study, the DLR, based on AI algorithms, does not discard the raw data but instead predicts and effectively supplements high-frequency K-space data, thereby improving the suppression of Gibbs-related artifacts. Although not systematically observed in this study, DLR may introduce subtle intensity variations or novel image features that differ from conventional reconstruction methods. These effects warrant further investigation.

Lesion conspicuity is attributed to improvements in image quality and partial volume reduction (27). In this study, 1.25-mm DR-HR-uCS scanning was employed, which enhances image quality by increasing spatial resolution and reducing noise. The uptake of Gd-EOB-DTPA by normal hepatocytes, particularly during the HBP, increases liver SI, thereby compensating reduced SNR. The findings of this study indicate that, through the use of thinner slice thickness and higher acceleration factors, DR-HR-uCS demonstrated superior performance in noise reduction and artifact suppression, while simultaneously reducing partial volume effects, making small lesions more conspicuous.

For lesions <1 cm in diameter, conventional imaging techniques often fail to provide sufficient information. In this study, DR-HR-uCS with 1.25-mm isotropic scanning used three-dimensional convolutional kernels during network training to learn information from three orthogonal directions, thereby supporting multiplanar reconstruction. This approach aids clinicians to more clearly identify and analyze small liver lesions from different angles, thereby improving early diagnostic accuracy. These findings are consistent with those reported by Zhao et al. (28), who demonstrated that adaptive CS-assisted 5T ethoxybenzyl (EOB)-MRI achieved excellent diagnostic performance for CRLMs, including lesions ≤10 mm, with high sensitivity and positive predictive value across various hepatic backgrounds.

Enhancing the conspicuity of small lesions holds significant clinical value, enabling more accurate detection of liver metastases and improving early diagnostic rates. Early diagnosis plays a critical role in the development of personalized treatment plans, and early intervention can significantly improve patient prognosis, survival and quality of life. The improved sensitivity observed with DR-HR-uCS, particularly for lesions <1 cm, is likely due to the combination of deep learning-based denoising and high-resolution reconstruction. These features enhance lesion conspicuity by improving edge definition, suppressing background noise, and preserving fine anatomical details. In small metastases where contrast against liver parenchyma is often limited, such enhancements may increase radiologists’ confidence in lesion identification, thereby increasing the true positive detection rate. This is especially relevant for clinical decision-making in patients undergoing surgical planning or systemic therapy. The uCS reconstruction in this study was performed using vendor-default clinical parameters without additional optimization, reflecting routine clinical practice. Conversely, DLR aims to balance denoising and structural preservation, which may result in comparable subjective sharpness despite the higher spatial resolution of DR-HR-uCS.

As anticipated, the diagnostic time of radiologists was significantly affected in the sub-centimeter CRLMs on uCS and DR-uCS images, with senior readers performing faster than junior and intermediate readers. However, with DR-HR-uCS imaging, the diagnostic times were consistently shorter across all experience levels and were no longer dependent on reader expertise. Notably, DR-HR-uCS also appeared to reduce the disparity in reading time between junior and senior readers, potentially due to improved image clarity and lesion conspicuity. This phenomenon may be attributed to the improved image uniformity, clarity, and lesion conspicuity provided by high-resolution DLR. The reduction in image noise and enhancement of lesion boundaries likely mitigated the interpretative challenges typically encountered by less experienced radiologists. As a result, the cognitive load and ambiguity during interpretation were substantially reduced, enabling more efficient lesion detection regardless of reader expertise. These findings suggest that DR-HR-uCS reconstruction not only improves diagnostic performance but may also help standardize interpretation, potentially reducing diagnostic variability in clinical practice. Future studies may further quantify cognitive load or diagnostic confidence to validate this hypothesis.

In this study, only HBP images were included for quantitative and qualitative analyses. This phase was selected because gadoxetic acid-enhanced HBP imaging provides the highest lesion-to-liver contrast for CRLMs, thereby maximizing lesion conspicuity and diagnostic confidence. Other dynamic phases (arterial and portal venous) were not systematically analyzed, as they are more susceptible to respiratory motion and generally provide lower contrast between metastatic lesions and background liver parenchyma. To minimize potential confounding factors related to phase-dependent image characteristics, we focused exclusively on HBP images. The performance of DR-HR-uCS in other contrast phases warrants further investigation in future studies.

The coil configuration used in this study consisted of a 12-channel body coil combined with a 32-channel spine coil, which is a commonly adopted setup in routine abdominal MRI. Although this configuration provides adequate signal coverage and supports advanced reconstruction techniques, variations in coil design, channel number, and scanner hardware across institutions may influence signal characteristics and reconstruction performance. Therefore, further multi-center studies are warranted to validate the generalizability of DR-HR-uCS across different hardware platforms and vendor environments.

Limitations

First, this was a single-center study with a limited sample size, focusing primarily on the preoperative assessment of patients with colorectal cancer. Second, due to the potential for Gd-EOB-DTPA to cause transient motion artifacts during the arterial phase, only HBP images were included in this study. The omission of arterial phase images may lead to an underestimation of lesion detection rates, as arterial phase imaging is particularly important for liver lesion assessment. Further research is needed to expand the dataset and comprehensively evaluate the application of DLR in Gd-EOB-DTPA-enhanced liver imaging. Further, potential biases in the reader study were primarily mitigated through randomization and blinding. Although a 2-week washout period was applied between reading sessions to reduce recall bias, intra-reader variability was not assessed. In addition, DR-HR-uCS differed from uCS and DR-uCS not only in the reconstruction method but also in the acquisition protocol. Specifically, DR-HR-uCS was obtained using a separate high-resolution acquisition with a different spatial resolution, a thinner slice thickness, a longer scan time, and an additional breath-hold compared with the standard-resolution acquisition used for uCS and DR-uCS. Therefore, the effect of DLR alone could not be isolated in DR-HR-uCS. The observed advantages of DR-HR-uCS should thus be interpreted as the combined effect of high-resolution acquisition and DLR, rather than as the effect of reconstruction alone. Moreover, because this study was performed on a single vendor platform using a fixed coil configuration, the generalizability of these findings to other scanners, coil setups, and vendor-specific implementations requires further validation.

Conclusions

The study findings suggest that the integration of DR-HR-uCS improves image quality by enhancing the SNR and reducing noise in Gd-EOB-DTPA-MRI of the liver. Moreover, DR-HR-uCS may improve the visibility and detection efficiency of sub-centimeter CRLMs, potentially facilitating early diagnosis. While these results are promising, further validation through larger, prospective, multi-center studies is warranted before broad clinical adoption.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the STARD-AI reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1-2540/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1-2540/dss

Funding: This study was supported by the Key Science and Technology Program of Henan Province (No. 242102311111).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1-2540/coif). The authors have no 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. The study was approved by the Medical Ethics Committee of Henan Cancer Hospital (No. 2023-KY-0089-001). The waiver of informed consent was granted because the study was retrospective in nature, and did not involve any additional interventions or procedures beyond routine clinical care.

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. Engstrand J, Nilsson H, Strömberg C, Jonas E, Freedman J. Colorectal cancer liver metastases - a population-based study on incidence, management and survival. BMC Cancer 2018;18:78. [Crossref] [PubMed]
2. Legou F, Chiaradia M, Baranes L, Pigneur F, Zegai B, Djabbari M, Calderaro J, Laurent A, Kobeiter H, Rahmouni A, Luciani A. Imaging strategies before beginning treatment of colorectal liver metastases. Diagn Interv Imaging 2014;95:505-12. [Crossref] [PubMed]
3. Tang W, Xiao Y, Kuang S, Rong D, He B, Grazioli L, Hussain SM, Wang J. Intraindividual crossover comparison of gadobenate dimeglumine-enhanced and gadoxetate disodium-enhanced MRI for characterizing focal liver lesions. Eur Radiol Exp 2025;9:23. [Crossref] [PubMed]
4. 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]
5. Hollingsworth KG. Reducing acquisition time in clinical MRI by data undersampling and compressed sensing reconstruction. Phys Med Biol 2015;60:R297-322. [Crossref] [PubMed]
6. Priyanka Kadavigere R, Nayak S S, Chandran M O, Shirlal A, Pires T, Pendem S. Impact of artificial intelligence assisted compressed sensing technique on scan time and image quality in musculoskeletal MRI - A systematic review. Radiography (Lond) 2024;30:1704-12. [Crossref] [PubMed]
7. Deng H, Dong X, Zhang Y, Zhou P, He Y, Yang L. Preliminary study on the feasibility of united compressed sensing with radial acquisition as a routine method for liver dynamic contrast-enhanced examination in elderly patients with malignancy. Insights Imaging 2025;16:65. [Crossref] [PubMed]
8. Li H, Alves VV, Pednekar A, Manhard MK, Greer J, Trout AT, He L, Dillman JR. Impact of Emerging Deep Learning-Based MR Image Reconstruction Algorithms on Abdominal MRI Radiomic Features. J Comput Assist Tomogr 2024;48:955-62. [Crossref] [PubMed]
9. Ottesen JA, Storas T, Vatnehol SAS, Løvland G, Vik-Mo EO, Schellhorn T, Skogen K, Larsson C, Bjørnerud A, Groote-Eindbaas IR, Caan MWA. Deep learning-based Intraoperative MRI reconstruction. Eur Radiol Exp 2025;9:29. [Crossref] [PubMed]
10. Kaniewska M, Zecca F, Obermüller C, Ensle F, Deininger-Czermak E, Lohezic M, Guggenberger R. Deep learning reconstruction of zero-echo time sequences to improve visualization of osseous structures and associated pathologies in MRI of cervical spine. Insights Imaging 2025;16:29. [Crossref] [PubMed]
11. Lee EJ, Chang YW, Lee EH, Cha JG, Kim SY, Choi N, Paek M, Darwish O. Image quality and diagnostic performance of deep learning reconstruction for diffusion- weighted imaging in 3 T breast MRI. Eur J Radiol 2025;185:111997. [Crossref] [PubMed]
12. Kubicka F, Tan Q, Meyer T, Nickel D, Weiland E, Wagner M, Marticorena Garcia SR. Deep-Learning-Based Reconstruction of Single-Breath-Hold 3 mm HASTE Improves Abdominal Image Quality and Reduces Acquisition Time: A Quantitative Analysis. Curr Oncol 2025;32:30. [Crossref] [PubMed]
13. Elhaie M, Koozari A, Alshammari QT. Machine and deep learning for MRI-based quantification of liver iron overload: a systematic review and meta-analysis. Radiologie (Heidelb) 2025;65:10-20. [Crossref] [PubMed]
14. Elhaie M, Koozari A, Koohi H, Alqurain QT. Machine learning and deep learning approaches in MRI for quantifying and staging fatty liver disease: A systematic review. J Med Imaging Radiat Sci 2025;56:102112. [Crossref] [PubMed]
15. Elhaie M, Koozari A, Arjmandi M, Najafizade N. Deep learning for hepatocellular carcinoma segmentation in MRI: A systematic review of models, performance, and challenges. Medicine (Baltimore) 2025;104:e47061. [Crossref] [PubMed]
16. Takayama Y, Sato K, Tanaka S, Murayama R, Jingu R, Yoshimitsu K. Effectiveness of deep learning-based reconstruction for improvement of image quality and liver tumor detectability in the hepatobiliary phase of gadoxetic acid-enhanced magnetic resonance imaging. Abdom Radiol (NY) 2024;49:3450-63. [Crossref] [PubMed]
17. Castagnoli F, Rata M, Shur J, Hopkinson G, Macdonald A, Stockton D, Nickel MD, Kannengiesser S, Messiou C, Koh DM, Winfield JM. AI-augmented reconstruction provides improved image quality and enables shorter breath-holds in contrast-enhanced liver MRI. Eur Radiol Exp 2025;9:46. [Crossref] [PubMed]
18. Takatsu Y, Harada S, Yamatani Y, Fukuzawa K, Nakamura M, Takano K, Miyati T. Characterization of the evaluation method for hepatobiliary phase image by liver magnetic resonance imaging using Gd-EOB-DTPA. Magn Reson Imaging 2025;120:110382. [Crossref] [PubMed]
19. Kankaraš M, Capecchi S. Neither agree nor disagree: use and misuse of the neutral response category in Likert-type scales. METRON 2025;83:111-40.
20. Westland JC. Information loss and bias in likert survey responses. PLoS One 2022;17:e0271949. [Crossref] [PubMed]
21. Sun W, Wang W, Zhu K, Chen CZ, Wen XX, Zeng MS, Rao SX. Feasibility of compressed sensing technique for isotropic dynamic contrast-enhanced liver magnetic resonance imaging. Eur J Radiol 2021;139:109729. [Crossref] [PubMed]
22. Jiang D, Dou W, Vosters L, Xu X, Sun Y, Tan T. Denoising of 3D magnetic resonance images with multi-channel residual learning of convolutional neural network. Jpn J Radiol 2018;36:566-74. [Crossref] [PubMed]
23. You X, Cao N, Lu H, Mao M, Wanga W. Denoising of MR images with Rician noise using a wider neural network and noise range division. Magn Reson Imaging 2019;64:154-9. [Crossref] [PubMed]
24. Afat S, Wessling D, Afat C, Nickel D, Arberet S, Herrmann J, Othman AE, Gassenmaier S. Analysis of a Deep Learning-Based Superresolution Algorithm Tailored to Partial Fourier Gradient Echo Sequences of the Abdomen at 1.5 T: Reduction of Breath-Hold Time and Improvement of Image Quality. Invest Radiol 2022;57:157-62. [Crossref] [PubMed]
25. Kidoh M, Shinoda K, Kitajima M, Isogawa K, Nambu M, Uetani H, Morita K, Nakaura T, Tateishi M, Yamashita Y, Yamashita Y. Deep Learning Based Noise Reduction for Brain MR Imaging: Tests on Phantoms and Healthy Volunteers. Magn Reson Med Sci 2020;19:195-206. [Crossref] [PubMed]
26. Plummer JW, Hussain R, Bdaiwi AS, Soderlund SA, Hoyos X, Lanier JM, Garrison WJ, Parra-Robles J, Willmering MM, Niedbalski PJ, Cleveland ZI, Walkup LL. A decay-modeled compressed sensing reconstruction approach for non-Cartesian hyperpolarized 129Xe MRI. Magn Reson Med 2024;92:1363-75. [Crossref] [PubMed]
27. Wei H, Yoon JH, Jeon SK, Choi JW, Lee J, Kim JH, Nickel MD, Song B, Duan T, Lee JM. Enhancing gadoxetic acid-enhanced liver MRI: a synergistic approach with deep learning CAIPIRINHA-VIBE and optimized fat suppression techniques. Eur Radiol 2024;34:6712-25. [Crossref] [PubMed]
28. Zhao X, Liu Y, Li S, Wang P, Chen H, Yin D, Wu X. AI-assisted compressed sensing and gadoxetic acid-enhanced MRI for evaluating colorectal liver metastases in complex hepatic backgrounds: a prospective 5T MRI study. Quant Imaging Med Surg 2026;16:41. [Crossref] [PubMed]