Abstract
Objective:
Routine clinical 3D magnetic resonance cholangiopancreatography (MRCP) is typically performed either as lower resolution breath-hold (BH) acquisition or higher resolution triggered navigator breathing (NAV) acquisition with a longer acquisition time, potentially limiting clinically applicability. Our aim was to evaluate the clinical feasibility of MRCP obtained using a new deep learning acceleration approach, Sonic DL™ 3D (SDL), as compared to the standard 3D MRCP..
Material and Methods:
Sixty-four and 32 patients underwent 1.5T and 3T MRCP scans, respectively. The standard 3D MRCP was obtained using a BH and NAV. Two SDL MRCP acquisitions for each magnetic field were performed with BH and scan times of 11–17s, utilizing the SDL based acquisition and reconstruction technique. Three radiologists visually evaluated all MRCP datasets for 4 different features: overall image quality, image noise, image sharpness and artifacts. Differences were compared using the Wilcoxon matched-pairs signed rank and chi-squared tests.
Results:
At both field strengths (1.5T and 3T) the proportion of overall image quality scores of 3 or higher (good, very good and excellent) and proportion of cases with reduced artifacts were better (p<0.05) for both SDL MRCP acquisitions compared to the standard with NAV or BH. However, SDL MRCP was not superior to the standard MRCP technique for all evaluated feature categories and scoring varied between readers.
Conclusion:
SDL MRCP demonstrated improved image quality consistency and scan time, however, variations between radiologists in quality scores were present, underlying the need for future development and validation.
Keywords: magnetic resonance cholangiopancreatography, deep learning reconstruction, accelerated MRI
INTRODUCTION
Magnetic resonance cholangiopancreatography (MRCP) is a non-invasive imaging approach to evaluate the hepatobiliary system utilizing the high T2 contrast between biliary tree and background soft tissues [1,2]. In most contemporary clinical practices, MRCP is performed using 3D imaging with a lower resolution breath-hold (BH) acquisition or a higher resolution triggered navigator breathing (NAV) acquisition requiring a longer acquisition time [3]. Ideally, a high resolution MRCP acquired with a short scan time (<15s) utilizing a BH acquisition is preferred as it supports consistent superior image quality across diverse patient populations. Towards accelerating MRCP data acquisition, various methods for under-sampling of k-space data have been applied, including parallel imaging [4] and compressed sensing [5]. Parallel imaging reduces scan time by utilizing correlations between multiple coil sensitivity profiles [4]. Compressed sensing benefits from the incoherence of under-sampling patterns, which leads to uncorrelated aliasing artifacts, allowing image reconstruction from highly under-sampled k-space, leading to shorted scan times and less artifacts [5]. While deep learning reconstruction (DLR) approaches [6,7] have become very popular and widely adopted in routine clinical practice for various MRI sequences, DLR of MRCP acquisitions aimed at acceleration and improving image quality consistency remain understudied [8].
The goal of this clinical feasibility study was to evaluate a deep learning-based acquisition and reconstruction technique, Sonic DL™ 3D (SDL, GE HealthCare, Waukesha, WI), aimed at acceleration of MRCP and improvements of image quality. The SDL version was compared to the standard 3D MRCP performed using a conventional parallel imaging (Array Spatial Sensitivity Encoding Technique or ASSET, GE HealthCare, Waukesha, WI).
MATERIAL AND METHODS
As a part of a prospective institutional review board approved single center study, 65 and 34 consecutive patients undergoing routine clinical MRCP at 1.5T and 3T, respectively, between September 2025 - November 2025 were included. The MRI examinations at lower field strength were performed using a 1.5T Signa™ Artist MRI scanner (GE HealthCare, Waukesha, WI) while all 3T examinations were performed using a 3T Signa™ Premier, (GE HealthCare, Waukesha, WI) clinical scanner. The standard 3D MRCP (Table 1), used as the reference standard, was obtained using a BH and NAV and accelerated with conventional parallel imaging. SDL MRCP (Table 1) was performed using a vendor supplied prototype version of an acquisition and reconstruction technique which combines a highly undersampled data acquisition with an unrolled deep learning-based reconstruction. Two different SDL MRCP acquisitions were performed, both using BH: SDL MRCP BH with 1.2×1.8×1.7 mm3 acquisition voxel size (scan time of 12 seconds) and SDL MRCP BH HighRes with 1.2×1.2×1.7 mm3 acquisition voxel size (scan time of 17 seconds), Table 1. The SDL reconstruction employs a physics-based signal model and enforces data consistency to promote accurate image reconstruction [9,10]. SDL utilizes under-sampling in the phase and slice dimensions with a pre-trained neural network employed to reconstruct the highly undersampled k-space [11]. In addition, a commercially available DLR algorithm (AIR™ Recon DL, GE HealthCare, Waukesha, WI) was applied to all acquisitions. The DLR algorithm is based on a Convolutional Neural Network (CNN) trained to remove imaging artifacts, reduce noise, and increase sharpness [12].
Table 1:
MRI acquisition parameters for the 1.5T acquisition
| Standard MRCP NAV | Standard MRCP BH | SDL MRCP BH | SDL MRCP BH HighRes | |
|---|---|---|---|---|
|
| ||||
| FOV | 300 × 240 mm2 | 300 × 195 mm2 | 300 × 195 mm2 | 300 × 195 mm2 |
| In-plane matrix | 256 × 128 | 256 × 104 | 256 × 104 | 256 × 166 |
| Slice Thickness | 1.7 mm | 1.7 mm | 1.7 mm | 1.7 mm |
| Slice Resolution | 50% | 50% | 50% | 50% |
| # of slices | 52 | 48 | 52 | 52 |
| TR/TE/ETL | 4000ms/900ms/220 | 2122ms/900ms/160 | 3124ms/400ms/400 | 2972ms/400ms/400 |
| Refocus Flip Angle | 126.6 | 99 | 160 | 160 |
| Parallel imaging acceleration factor * | 2 | 3 | 5 | 5 |
| Default Scan Time (sec) | 68 | 19 | 12 | 17 |
Array Spatial Sensitivity Encoding Technique (ASSET) was utilized as parallel imaging technique on standard acquisitions.
Three boards certified radiologists independently evaluated each MRCP acquisition (8 total: 4 at 1.5T and 4 at 3T) in an unblinded fashion using diagnostic quality monitors. A 5-point Likert scoring scale across four feature categories was used: overall image quality, signal to noise, image sharpness and artifacts (Supplementary Material Table S1).
Qualitative differences between the standard MRCP (Standard MRCP NAV and Standard MRCP BH) and SDL MRCP (SDL MRCP BH and SDL MRCP BH HighRes) were assessed using the Wilcoxon matched-pairs signed rank and chi-squared tests in GraphPad Prism v10.6.1 (Boston, MA). Interrater agreement assessments were estimated using Gwet’s AC2 statistic using ordinal weights in R 4.5.0 [13]. Using Altman's rules [14], benchmark values were assigned as follows: 0.8–1.0 = Very Good; 0.6–0.8 = Good; 0.4–0.6 = Moderate; 0.2–0.4 = Fair; 0.1–0.2 = Poor. A p-value of <0.05 was considered statistically significant.
RESULTS
One patient scanned at 1.5T and 2 patients scanned at 3T were excluded from final analyses since these patients were unable to tolerate the entire MRI examination duration. In total, 64 and 32 patients were included in the final analyses who were scanned at 1.5T and 3T, respectively. The mean (standard deviation) age for the 1.5T and 3T acquisitions were 65 (12) and 64 (15) years, respectively. Similarly, 54% and 72% of patients were female in the 1.5T and 3T acquisitions, respectively.
At both field strengths (1.5T and 3T), the proportion of overall image quality scores of 3 or higher (good, very good and excellent) were significantly higher for the SDL MRCP BH and SDL MRCP BH HighRes acquisitions compared to the MRCP with NAV or BH, with the exception of Reader 2 and Reader 3 for the comparison of Standard MRCP BH and SDL MRCP BH HighRes where the difference did not reach the level of statistical significance (p-values = 0.06 and 0.7), Table 3 and Table 4.
Table 3:
Qualitative Image Assessment at 1.5T
| Reader 1 | Reader 2 | Reader 3 | |||||||
|---|---|---|---|---|---|---|---|---|---|
| Median (IQR) | p-value | Median (IQR) | p-value | Median (IQR) | p-value | ||||
| Standard MRCP NAV | SDL MRCP BH | Standard MRCP NAV | SDL MRCP BH | Standard MRCP NAV | SDL MRCP BH | ||||
| IQ | 3 (2–3) | 4 (3–4) | <0.001 | 3 (2–3) | 3 (3–3) | 0.009 | 3 (2–3) | 3 (2–3) | 0.009 |
| % of scan with IQ>2 | 66% | 98% | <0.001 | 63% | 86% | <0.01 | 50% | 73% | <0.01 |
| Noise | 3 (3–4) | 4 (4–4) | <0.001 | 3 (3–4)* | 3 (2–3)* | <0.001* | 3 (3–4)* | 3 (2–3)* | <0.001* |
| Sharpness | 3 (2–3) | 4 (3–4) | <0.001 | 3 (2–4) | 3 (3–4) | <0.001 | 3 (2–3) | 3 (3–4) | 0.001 |
| Artifacts | 3 (2–3) | 4 (3–4) | <0.001 | 2 (2–3) | 3 (3–3) | <0.001 | 2 (2–3) | 3 (3–3) | <0.001 |
| Reader 1 | Reader 2 | Reader 3 | |||||||
| Median (IQR) | p-value | Median (IQR) | p-value | Median (IQR) | p-value | ||||
| Standard MRCP BH | SDL High Res | Standard MRCP BH | SDL High Res | Standard MRCP BH | SDL High Res | ||||
| IQ | 3 (3–4) | 4 (3–4) | <0.001 | 3 (2–3) | 3 (3–4) | 0.002 | 3 (2–3) | 3 (2–3) | 0.04 |
| % of scan with IQ>2 | 83% | 95% | <0.05 | 75% | 89% | 0.06 | 64% | 69% | 0.7 |
| Noise | 3 (3–4) | 4 (3–4) | <0.001 | 2 (2–3) | 3 (3–4) | <0.001 | 2 (2–3) | 3 (3–3) | <0.001 |
| Sharpness | 3 (2–4) | 4 (3–4) | <0.001 | 3 (3–4) | 4 (3–4) | 0.04 | 3 (3–4) | 3 (3–4) | 0.16 |
| Artifacts | 3 (2–4) | 4 (3–4) | <0.001 | 3 (3–3) | 3 (3–4) | 0.05 | 2 (2–3) | 3 (3–3) | <0.001 |
SDL MRCP performed worse than the standard MRCP
Table 4:
Qualitative Image Assessment at 1.5T
| 3T | Reader 1 | Reader 2 | Reader 3 | ||||||
|---|---|---|---|---|---|---|---|---|---|
| Median (IQR) | p-value | Median (IQR) | p-value | Median (IQR) | p-value | ||||
| Standard MRCP NAV | SDL MRCP BH | Standard MRCP NAV | SDL MRCP BH | Standard MRCP NAV | SDL MRCP BH | ||||
| IQ | 3 (2–3) | 4 (3–4) | <0.001 | 3 (2–3) | 3 (3–4) | 0.007 | 2 (1–3) | 3 (2–3) | 0.01 |
| % of scan with IQ>2 | 55% | 100% | <0.001 | 50% | 85% | <0.01 | 39% | 70% | 0.02 |
| Noise | 3 (2–3) | 4 (3–4) | <0.001 | 3 (2–3) | 3 (2–3) | 0.07 | 3 (2–3)* | 2 (2–3)* | <0.001* |
| Sharpness | 2 (2–3) | 3 (3–3) | <0.001 | 2 (2–4) | 4 (3–4) | 0.001 | 2 (2–3) | 3 (3–3) | 0.006 |
| Artifacts | 2 (2–3) | 4 (3–4) | <0.001 | 2 (2–3) | 3 (3–3) | <0.001 | 2 (1–3) | 3 (2–3) | 0.01 |
| Median (IQR) | p-value | Median (IQR) | p-value | Median (IQR) | p-value | ||||
| Standard MRCP BH | SDL MRCP BH HighRes | Standard MRCP BH | SDL MRCP BH HighRes | Standard MRCP BH | SDL MRCP BH HighRes | ||||
| IQ | 2 (1–3) | 4 (3–4) | <0.001 | 2 (1–3) | 3 (2–4) | <0.001 | 1 (1–2) | 3 (2–3) | <0.001 |
| % of scan with IQ>2 | 40% | 100% | <0.001 | 33% | 76% | <0.001 | 18% | 64% | <0.001 |
| Noise | 2 (2–3) | 4 (3–4) | <0.001 | 2 (1–2) | 3 (3–3) | <0.001 | 2 (1–2) | 3 (2–3) | 0.002 |
| Sharpness | 2 (2–3) | 4 (3–4) | <0.001 | 3 (1–3) | 4 (2–4) | <0.001 | 2 (2–3) | 3 (3–3) | <0.001 |
| Artifacts | 2 (1–2) | 4 (3–4) | <0.001 | 1 (1–3) | 3 (2–4) | <0.001 | 1 (1–2) | 3 (2–3) | <0.001 |
SDL MRCP performed worse than the standard MRCP
At 1.5T datasets, all 3 radiologists assigned significantly better scores (p-values p<0.001 to 0.009) for image quality, image sharpness, and image artifacts for SDL MRCP BH compared with standard MRCP performed with NAV, Table 2. In quality scores for signal to noise, two radiologists assigned significantly better (p<0.001) scores for standard MRCP performed with NAV. In comparison between Standard MRCP BH and SDL MRCP BH HighRes, all three radiologists assigned significantly better (p-values 0.04 to <0.001) scores for image quality, noise, and image artifacts (except for one radiologist) to SDL MRCP BH HighRes, Table 2. In quality scores for sharpness, all three radiologists assigned better for scores for MRCP SDL BH and MRCP BH HighRes, however, only for one radiologist the differences reach the statistical significance (p<0.001), Table 2.
Table 2:
MRI acquisition parameters for the 3T acquisition
| Standard MRCP NAV | Standard MRCP BH | SDL MRCP BH | SDL MRCP BH HighRes | |
|---|---|---|---|---|
|
| ||||
| FOV | 300 × 240 mm2 | 300 × 195 mm2 | 300 × 195 mm2 | 300 × 195 mm2 |
| In-plane matrix | 256 × 128 | 256 × 104 | 256 × 104 | 256 × 166 |
| Slice Thickness | 1.7 mm | 1.7 mm | 1.7 mm | 1.7 mm |
| Slice Resolution | 50% | 50% | 50% | 50% |
| # of slices | 48 | 48 | 48 | 48 |
| TR/TE/ETL | 4000ms/900ms/220 | 2067ms/900ms/160 | 4679ms/400ms/400 | 3825ms/400ms/400 |
| Refocus Flip Angle | 130 | 99 | 160 | 160 |
| Parallel imaging acceleration factor * | 3 | 3 | 6 | 6 |
| Default Scan Time (sec) | 26 | 14 | 11 | 17 |
Array Spatial Sensitivity Encoding Technique (ASSET) was utilized as parallel imaging technique on standard acquisitions.
For the 3T datasets, only one radiologist assigned significantly better image quality scores (p<0.001) for SDL MRCP BH compared with standard MRCP performed with NAV, Table 3. However, all three radiologists assigned significantly better image sharpness and artifacts scores (p<0.01) for SDL MRCP BH compared with standard MRCP performed with NAV, Table 3. In quality scores for signal to noise, two radiologists assigned better (p<0.001 and 0.07) scores for standard MRCP performed with NAV while the 3rd radiologist preferred SDL MRCP BH (p<0.001). In comparison between Standard MRCP BH and SDL MRCP BH HighRes, all three radiologists assigned significantly better (p-values 0.002 to <0.001) scores for image quality, signal to noise, image sharpness and image artifacts to SDL MRCP BH HighRes, Table 3.
Interrater agreements for quality score assessments (overall image quality, signal to noise, image sharpness, image artifacts) at 1.5T were good to very good (Gwet’s AC2 in the range of 0.613 to 0.895) except for moderate agreement for evaluation of signal to noise (Gwet’s AC2 of 0.504) for evaluation of noise of the Standard MRCP BH, Supporting Material Table S1. At 3T, interrater agreements for quality score assessments (overall image quality, signal to noise, image sharpness, image artifacts) were good to very good (Gwet’s AC2 in the range of 0.626 to 0.830) except for moderate agreement for evaluation of signal to noise ratio of SDL MRCP BH (Gwet’s AC2 of 0.430), and SDL MRCP BH HighRes (Gwet’s AC2 of 0.587) as well as evaluation of image sharpness of SDL MRCP BH (Gwet’s AC2 of 0.526), Supporting Material Table S2.
Representative images for 2 patients are shown in Figure 1 for 1.5T and Figure 2 for 3T.
Figure 1.
Representative images at 1.5T for two patients (A-D 1st patient, E-H 2nd patient). Red arrows point to the common bile duct while yellow arrow points to the main pancreatic duct.
Figure 2.
Representative images at 3T for two patients (A-D 1st patient, E-H 2nd patient). Red arrows point to the common bile duct while yellow arrows point to the main pancreatic duct.
DISCUSSION
A deep learning accelerated MRCP performed in BH (breath-hold duration of 11–17s) demonstrated better robustness than the standard MRCP performed either using NAV or BH as evident by the overall image quality scores of 3 or higher (good, very good and excellent), which were in significantly higher percentage of patients than the standard MRCP. SDL utilizes a deep-learning engine aimed at optimizing each individual reconstructed voxel by performing approximately 6.5 million operations on the raw data. Compared with the standard deep learning reconstruction algorithms and parallel imaging techniques, SDL is less sensitive to the limitations of coil geometry and g-factor [11] (a major limiting factor for parallel imaging techniques), resulting in fewer artifacts such as signal loss or aliasing. In the current study, all SDL acquisitions for all readers demonstrate significantly less artifacts compared to acceleration with parallel imaging techniques.
The current study has several limitations, especially potential for bias, since all radiologists were not blinded to the type of image acquisition. However, this limitation was partly mitigated by having independent review by three radiologists and image review was performed following strictly prospectively agreed criteria for image quality evaluation. The current study included 64 and 32 patients scanned at 1.5T and 3T, however, future large studies with more heterogeneous study populations are warranted, including pediatric population since only adult patients were included in this study. The present study also did not evaluate the impact of image quality on lesion detection, either in the biliary tree or pancreas. Future investigations will be necessary to test whether the improvements in image quality and robustness for SDL MRCP as described herein will translate to improved lesion detection and diagnostic performance. However, our comprehensive evaluation done by three board certified radiologists provides a robust multi-reader evaluation of image quality. Nevertheless, development and validation of quantitative methods for evaluation of image quality features such as signal to noise ratio or image sharpness is warranted but remains a challenge in the era of ever evolving and more sophisticated machine learning/artificial intelligence MRI acquisition and construction methods. Finally, SDL MRCP acquisition is a unique vendor specific acquisition, thus, multi-vendor validation is not possible now.
In conclusion, a deep learning accelerated MRCP performed in BH (breath-hold duration of 11–17s) demonstrated better robustness than the standard MRCP performed either using NAV or BH. SDL MRCP was not superior to the standard MRCP in all evaluated categories (overall image quality, signal to noise, image sharpness and artifacts) and variations between radiologists in quality scores were present, underlying the need for future development and validation.
Highlights:
- 3D MRCP at 1.5T and 3T is prone to imaging artifacts
- Breath-hold (BH) or navigator breathing (NAV) acquisition are commonly used
- Deep learning acceleration of MRCP with BH leads to improved image quality consistency.
Highlights:
3D MRCP at 1.5T and 3T is prone to imaging artifacts
Breath-hold (BH) or navigator breathing (NAV) acquisition are commonly used
Deep learning acceleration of MRCP with BH leads to improved image quality consistency.
Acknowledgement
NIH grant P41EB030006.
Footnotes
Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
REFERENCES
- [1].Barish MA, Yucel EK, Ferrucci JT, Magnetic resonance cholangiopancreatography, N Engl J Med 341 (1999) 258–264. 10.1056/NEJM199907223410407. [DOI] [PubMed] [Google Scholar]
- [2].Nandalur KR, Hussain HK, Weadock WJ, Wamsteker EJ, Johnson TD, Khan AS, D’Amico AR, Ford MK, Nandalur SR, Chenevert TL, Possible biliary disease: diagnostic performance of high-spatial-resolution isotropic 3D T2-weighted MRCP, Radiology 249 (2008) 883–890. 10.1148/radiol.2493080389. [DOI] [PubMed] [Google Scholar]
- [3].Yoon LS, Catalano OA, Fritz S, Ferrone CR, Hahn PF, Sahani DV, Another dimension in magnetic resonance cholangiopancreatography: comparison of 2- and 3-dimensional magnetic resonance cholangiopancreatography for the evaluation of intraductal papillary mucinous neoplasm of the pancreas, J Comput Assist Tomogr 33 (2009) 363–368. 10.1097/RCT.0b013e3181852193. [DOI] [PubMed] [Google Scholar]
- [4].Nagata S, Goshima S, Noda Y, Kawai N, Kajita K, Kawada H, Tanahashi Y, Matsuo M, Magnetic resonance cholangiopancreatography using optimized integrated combination with parallel imaging and compressed sensing technique, Abdom Radiol (NY) 44 (2019) 1766–1772. 10.1007/s00261-018-01886-0. [DOI] [PubMed] [Google Scholar]
- [5].Yoon JH, Lee SM, Kang H-J, Weiland E, Raithel E, Son Y, Kiefer B, Lee JM, Clinical Feasibility of 3-Dimensional Magnetic Resonance Cholangiopancreatography Using Compressed Sensing: Comparison of Image Quality and Diagnostic Performance, Invest Radiol 52 (2017) 612–619. 10.1097/RLI.0000000000000380. [DOI] [PubMed] [Google Scholar]
- [6].Hyun CM, Kim HP, Lee SM, Lee S, Seo JK, Deep learning for undersampled MRI reconstruction, Phys Med Biol 63 (2018) 135007. 10.1088/1361-6560/aac71a. [DOI] [PubMed] [Google Scholar]
- [7].Advancing MRI reconstruction: A systematic review of deep learning and compressed sensing integration, Biomedical Signal Processing and Control 111 (2026) 108291. 10.1016/j.bspc.2025.108291. [DOI] [Google Scholar]
- [8].Kim J, Nickel MD, Knoll F, Deep Learning-Based Accelerated MR Cholangiopancreatography Without Fully-Sampled Data, NMR Biomed 38 (2025) e70002. 10.1002/nbm.70002. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [9].Ahn S, Wollner U, McKinnon G, Jansen IH, Brada R, Rettmann D, Cashen TA, Huston J, DeMarco JK, Shih RY, Trzasko JD, Hardy CJ, Foo TKF, Deep learning-based reconstruction of highly accelerated 3D MRI, (2022). 10.48550/arXiv.2203.04674. [DOI] [Google Scholar]
- [10].Lebel RM, Performance characterization of a novel deep learning-based MR image reconstruction pipeline, (2020). 10.48550/arXiv.2008.06559. [DOI] [Google Scholar]
- [11].Sonic DL for volumetric imaging, (n.d.). https://signapulse.gehealthcare.com/autumn-2025/%20sonic-dl-for-volumetric-imaging (accessed February 22, 2026). [Google Scholar]
- [12].Lebel RM, Performance characterization of a novel deep learning-based MR image reconstruction pipeline, (2020). 10.48550/arXiv.2008.06559. [DOI] [Google Scholar]
- [13].Gwet KL, Handbook of inter-rater reliability: the definitive guide to measuring the extent of agreement among raters, Fourth edition, Advances Analytics, LLC, Gaithersburg, Md, 2014. [Google Scholar]
- [14].Altman DG Practical Statistics for Medical Research. London: Chapman and Hall, 1991. [Google Scholar]