Impact of deep-learning image reconstruction on multiplexed sensitivity encoding diffusion-weighted imaging in the female pelvis

Introduction

Diffusion-weighted imaging (DWI) is a standard sequence in the imaging protocol for female pelvis as recommended by many of the European Society of Urogenital Radiology (ESUR) guidelines (1-4). DWI is sensitive to tissue microstructure and has been used in disease detection, tissue characterisation and therapy monitoring (5). Implementation of DWI in the female pelvis is not without its challenges, in the presence of bowel peristalsis, physiologic uterine movement and gas in the rectum and sigmoid colon. Although conventional DWI based on single-shot echo-planar imaging (EPI) is advantageous in its fast acquisition that minimizes movement artefacts, geometric distortion is a major shortcoming when encountering boundaries with different magnetic susceptibility (6). Furthermore, EPI-DWI suffers from signal-intensity dropout and image blurring due to the relatively long traversal through the k-space along the phase-encoding direction, resulting in low-resolution images (7).

With the advent of DWI using multiplexed sensitivity encoding (MUSE), it has been able to deliver high-resolution imaging, overcoming shot-to-shot phase variations in multi-shot EPI, thus enabling reduction in geometric distortion and large field-of-view imaging (8) by increasing the number of excitations.

There has been considerable attention to applying deep-learning image reconstruction (DLRecon) to improve image quality and efficiency. Many algorithms have been proposed for magnetic resonance imaging (MRI) reconstruction, from under-sampled k-space data (9), low-quality images (10), to accelerated MRI reconstruction (11). Most of these works have been performed outside of the pelvis with promising results (12-14). In bladder carcinoma, DLRecon was able to reduce acquisition time, while improving image quality, signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) (15). In prostate cancer, DLRecon also improved SNR and CNR with better image quality (16).

Herein, we proposed to evaluate the impact of DLRecon on MUSE DWI in the female pelvis by assessing the image quality both qualitatively, based on a 5-point scale and quantitatively using SNR, CNR and apparent diffusion coefficient (ADC) values with or without DLRecon. We present this article in accordance with the STARD reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0005/rc).

Methods

Subjects

Consecutive female subjects who were scheduled for MRI of the pelvis were prospectively recruited between November 2022 and March 2025 from one centre. These recruited subjects were referred for both benign and malignant gynaecological conditions. MRI examinations without MUSE DWI were excluded. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the ethics committee of The University of Hong Kong/Hospital Authority Hong Kong West Cluster (No. UW25-148) and informed consent was obtained from all individual participants.

MUSE MRI

All MRI examinations were performed on a 3T scanner (Signa Premier, GE Healthcare, Waukesha, WI, USA) using a 30-channel phased-array coil for signal reception. Patients were required to fast for at least 6 hours before MRI examinations and 20 mg of intravenous hyoscine butylbromide (Buscopan, Boehringer Ingelheim) was given before MRI examinations to reduce peristaltic movements. Standard 2-shot MUSE was acquired in an axial plane with three b values (0, 400 and 800 s/mm²) (Table 1). An additional 4-shot MUSE was acquired in a randomly selected subset of patients for qualitative assessment. Other conventional sequences were acquired as part of the clinical evaluation.

Prototype vendor-specific DLRecon was used to optimize MUSE image reconstruction through offline processing. In contrast to the conventional MUSE reconstruction approach (8), the vendor-supplied deep learning-enabled MUSE framework (DLRecon MUSE) incorporated two fundamental advancements. Initially, a robust complex signal averaging process was facilitated by a pretrained neural network (DL phase correction), which yielded high-fidelity phase maps (17). This model was structured as a 3-layer U-net residual network, featuring approximately 4.4 million parameters across 10,000 kernels. Notably, it excluded bias terms and utilized ReLU activation functions to ensure generalizability. The network was developed via supervised learning on a dataset of over 10,000 internal images encompassing diverse SNR profiles. By subtracting the predicted residual from the input, a denoised image was produced to derive the phase estimate. Subsequently, the averaged data underwent refinement through a highly optimized convolutional neural network (DLRecon) designed to mitigate noise and truncation artefacts (18). Operating within the complex-valued domain, this residual network leveraged over 100,000 unique pattern recognition kernels. Like the phase correction model, it maintained generalizability by omitting bias terms and employing ReLU activations, allowing the system to adapt across varying signal intensities and noise levels. The denoising level in this study was set at high (75%) DL strength to reconstruct MRI images from raw K-space data.

Image analysis

Two board-certified radiologists (over 15 years’ experience in pelvic imaging) were blinded to the anonymized and randomly distributed non-DLRecon and DLRecon MUSE images and reviewed these on ITK-SNAP (Version 4.2.2; http://www.itksnap.org). The two reviewers qualitatively scored the high b-value images (b=800 s/mm²) on overall image quality, artefacts, lesion conspicuity and sharpness using a 5-point scale as previously described (19) (Appendix 1). The same task was repeated at least 4 weeks apart by radiologist 1. Intra-observer and inter-observer agreements were evaluated on qualitative scoring.

Analysis of quantitative parameters was performed using the dedicated workstation (version 4.7, GE Healthcare, Chicago, USA) and Mango software (version 4.1) (20-22) on 2-shot MUSE. Three spherical regions of interest (ROIs) with radius of 6 mm were placed on the gluteus muscle, uterus and fat by one experienced radiographer and verified by radiologist 1 (Figure 1).

Figure 1 Exemplar showing placement of region of interest in the uterus, fat and muscle. ROI, region of interest.

The SNR of each reconstructed image was calculated using the following equations:

SNRmuscle/uterus=MeansignalintensityintheROIsofthemuscle/uterusStandarddeviationofthefatROI[1]

The CNR of each reconstructed image was calculated using the following equation:

CNR=AbsolutedifferenceofsignalintensitybetweentheROIsoftheuterusandmuscleStandarddeviationofROIthefat[2]

The ADC values of the gluteus muscle and the uterus were compared. In addition, ADCs were computed by delineating volumes of interest (VOIs) around the lesions on the ADC map, if present. The VOIs were subsequently verified by radiologist 1.

Statistical analysis

Inter-observer and intra-observer reliability were evaluated by Cohen’s kappa: κ values of 0.01–0.20 are defined as no agreement, 0.21–0.40 as fair, 0.41–0.60 as moderate, 0.61–0.80 as substantial and 0.81–1.00 as excellent agreement (23). The qualitative assessment of image quality was compared using paired-sample Wilcoxon signed-rank test. The quantitative data from the measured ROIs (SNR, CNR, and ADC) were analyzed using paired-sample t-test and Wilcoxon signed-rank test depending upon the distribution of the data. The statistical significance was defined as P<0.050 and P values were adjusted using Bonferroni correction for multiple comparisons. No contradictions were found across any comparison after applying Bonferroni correction. Bland-Altman analysis was performed to compare the lesion ADCs derived from the non-DLRecon and DLRecon reconstructed images.

Results

During the study period, there were 69 female patients who underwent pelvic MRI. Four patients were excluded: unrelated to gynaecological condition (N=1), incomplete MUSE DWI (N=3). A total of 65 female patients (mean age 55.0±13.6 years) were included in the final study cohort, providing a study power of 80%. All patients underwent standard 2-shot MUSE. Among them, 42 patients were randomly selected to undergo additional 4-shot MUSE. There were 59 lesions identified, malignant lesions n=27, benign lesions n=32 (Figure 2). Among the malignant lesions, there were endometrial carcinoma (n=15), cervical carcinoma (n=5), ovarian carcinoma (n=4), vaginal carcinoma (n=1) and nodal metastases (n=2). Malignant lesions were confirmed on histology and discussed at multi-disciplinary meetings. For benign lesions, there were endometrial polyp/hyperplasia (n=6), uterine leiomyomata (n=17), teratoma (n=2), endometrioma (n=1) and benign cystic ovarian masses (n=6). These were subsequently confirmed on histology or follow-up imaging.

Figure 2 Flowchart of patient recruitment. DWI, diffusion-weighted imaging; MRI, magnetic resonance imaging; MUSE, multiplexed sensitivity encoding.

Both the inter-observer and intra-observer agreements were substantial to excellent in all categories. For inter-observer agreements: overall image quality κ=0.772, artefacts κ=0.725, lesions conspicuity κ=0.664, sharpness κ=0.684 and overall agreement κ=0.715. For intra-observer agreements: overall image quality κ=0.815, artefacts κ=0.753, lesions conspicuity κ=0.698, sharpness κ=0.682 and overall agreement κ=0.744.

Qualitative analysis showed that DLRecon improved the image quality (Table 2, Figure 3). It not only eliminated the noise level but also increased lesion conspicuity (Figure 4). Overall image quality was significantly improved after DLRecon. When comparing the paired 2-shot and 4-shot MUSE, the image quality of 2-shot MUSE DLRecon was not statistically different from conventional 4-shot MUSE (Table 2, Figure 4). Nevertheless, we noted DLRecon had less impact on improving image quality when susceptibility artefact was severe (Figure 5).

Figure 3 Bar graphs showing the median and range of qualitative scores for different components of image quality between 2-shot MUSE with and without DLRecon. DLRecon, deep-learning image reconstruction; DW, diffusion-weighted; MUSE, multiplexed sensitivity encoding.

Figure 4 Exemplar showing the improvement of image quality by DLRecon. (A) 2-shot MUSE without DLRecon; (B) 2-shot MUSE with DLRecon; (C) 4-shot MUSE without DLRecon. DLRecon increased edge sharpness of the uterus outline (yellow arrows) and endometrial lining (red arrows) (A,B) and 2-shot MUSE with DLRecon had comparable image quality to 4-shot MUSE without DLRecon (B,C). DLRecon, deep-learning image reconstruction; MUSE, multiplexed sensitivity encoding.

Figure 5 No significant improvement in image quality with DLRecon when susceptibility artefact (orange arrows) was severe. (A) 2-shot MUSE without DLRecon; (B) 2-shot MUSE with DLRecon. DLRecon, deep-learning image reconstruction; MUSE, multiplexed sensitivity encoding.

Quantitatively, on 2-shot MUSE, the SNR and CNR both increased on images with DLRecon. These were significantly different throughout the different b values for SNR and CNR (all P<0.050), except for CNR at b=400 s/mm² (Figure 6).

Figure 6 SNR and CNR comparison on 2-shot MUSE MRI with and without DLRecon. CNR, contrast-to-noise ratio; DLRecon, deep-learning image reconstruction; MRI, magnetic resonance imaging; MUSE, multiplexed sensitivity encoding; ROI, region of interest; SNR, signal-to-noise ratio.

ADCs were not statistically different in the gluteus muscle and uterus on both 2-shot MUSE with DLRecon and without DLRecon (P=0.194 and P=0.127, respectively, Figure 7). Similarly, in the lesions, there were no significant differences in the ADC quantified from the ADC maps generated with or without DLRecon on 2-shot MUSE, with the differences between the two measurements staying within the 95% limits of agreement (P=0.414, Figure 8).

Figure 7 ADC comparison on 2-shot MUSE MRI with and without DLRecon. ADC, apparent diffusion coefficient; DLRecon, deep-learning image reconstruction; MRI, magnetic resonance imaging; MUSE, multiplexed sensitivity encoding; ROI, region of interest.

Figure 8 Bland-Altman plot of ADC measurements of lesions from 2-shot MUSE with and without DLRecon. ADC, apparent diffusion coefficient; DLRecon, deep-learning image reconstruction; MUSE, multiplexed sensitivity encoding; SD, standard deviation.

Discussion

Our study showed that DLRecon improved the image quality of 2-shot MUSE and there was no statistical difference when compared with 4-shot MUSE without DLRecon, therefore gaining the advantage of high image fidelity. DLRecon increased SNR and CNR, while ADC quantification was preserved, avoiding the introduction of unwanted variability during the process of disease monitoring.

Application of DWI in the female pelvis comes with various constraints with the movements of the pelvic organs, bowel activity and susceptibility artefacts from tissue-air interface resulting in signal loss and geometric distortion. MUSE overcomes some of these challenges (8) and there is an ongoing effort in optimizing it for clinical integration (24,25). In the current study, the application of DLRecon improved image quality, specifically in increasing lesion conspicuity and image sharpness. Integration of multi-shot MUSE with DLRecon may further enhance the image quality, specifically in reducing susceptibility artefact, within a clinically acceptable acquisition timeframe (25). Nevertheless, the improvement in image quality may be negligible with DLRecon when susceptibility artefact was severe, evident in some of the cases in this study cohort.

The improvement of image quality was coupled with increased SNR and CNR, while maintaining the quantitative values of ADC with DLRecon, both on normal structures and lesions, in agreement with previous findings on urinary bladder, prostate and rectal MRI (15,16,26). The stability of ADC on DLRecon images is important in ensuring that the introduction of DLRecon would not significantly affect quantification of ADC values, especially when serial examinations are performed to monitor the response to treatment. To evaluate the effect of DLRecon on quantification, MUSE DWI was acquired on the same machine using the same protocol to negate the effect of variations in field strength and imaging protocols (27). It was also noted that the interquartile ranges were larger on the DLRecon images, but were slowly closing the gaps as the b values increased, suggesting DLRecon amplified the signal of normal tissues of high water content on b=0 images without discrimination, but only focusing on tissues with restricted water diffusion on b=800 images.

Integration of DLRecon could potentially save acquisition time as shown in other studies, initially with T2W imaging (28,29) and subsequently with pelvic DWI (15,30). MUSE has many advantages in advancing the quality of DWI in the female pelvis, but is penalized with long acquisition time compared to conventional DWI based on EPI. In this study, we demonstrated that the image quality of DLRecon images of 2-shot MUSE was not significantly different from that acquired by 4-shot MUSE. Higher shot MUSE is computationally more demanding and may not be necessary for all patients. However, DLRecon allows MUSE to be retrospectively reconstructed to produce image quality that will not be statistically dissimilar from 4-shot MUSE, which may aid in increasing diagnostic certainty and confidence in selected challenging cases. Hence, if higher shot MUSE is desirable, DLRecon could achieve that by using lower shot MUSE, attaining high image fidelity without the need for additional acquisition.

This study has several limitations. First, the findings from this study might not be generalizable due to the small sample size and single-center study design. Future studies with larger sample size and multi-center participation should be encouraged to validate these findings. Second, although the inter-observer agreement was substantial, both observers were experienced radiologists. Future studies should include radiologists with different levels of experience to evaluate the interplay between rater’s experience and DLRecon to further refine its clinical impact. Third, the current DLRecon was a prototype provided by vendor and only offered offline analysis, limiting its clinical utility. Subsequent in-line integration would help clinical adoption and acceptance. Finally, the impact of DLRecon on clinical diagnosis was not evaluated and future evaluation concurrent with other imaging sequences is encouraged to elucidate its impact on clinical diagnosis.

Conclusions

In conclusion, DLRecon improved image quality of MUSE in the female pelvis by reducing the impact of artefacts, improving lesion conspicuity and image sharpness, while increasing SNR and CNR but preserving the stability of ADC quantification.

Acknowledgments

Part of the data was presented as abstract at ISMRM and ISMRT Annual Meeting and Exhibition 2023 (https://archive.ismrm.org/2023/1473.html).

Footnote

Reporting Checklist: The authors have completed the STARD reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0005/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0005/dss

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0005/coif). C.W.L., C.Y.L., P.L., and X.W. are employees of GE Healthcare. The other 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. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the ethics committee of The University of Hong Kong/Hospital Authority Hong Kong West Cluster (No. UW25-148), and informed consent was obtained from all individual participants.

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. Forstner R, Thomassin-Naggara I, Cunha TM, et al. Erratum to: ESUR recommendations for MR imaging of the sonographically indeterminate adnexal mass: an update. Eur Radiol 2017;27:2258. [Crossref] [PubMed]
2. Kubik-Huch RA, Weston M, Nougaret S, et al. European Society of Urogenital Radiology (ESUR) Guidelines: MR Imaging of Leiomyomas. Eur Radiol 2018;28:3125-37. [Crossref] [PubMed]
3. Nougaret S, Horta M, Sala E, et al. Endometrial Cancer MRI staging: Updated Guidelines of the European Society of Urogenital Radiology. Eur Radiol 2019;29:792-805. [Crossref] [PubMed]
4. Manganaro L, Lakhman Y, Bharwani N, et al. Staging, recurrence and follow-up of uterine cervical cancer using MRI: Updated Guidelines of the European Society of Urogenital Radiology after revised FIGO staging 2018. Eur Radiol 2021;31:7802-16. [Crossref] [PubMed]
5. Koh DM, Collins DJ. Diffusion-weighted MRI in the body: applications and challenges in oncology. AJR Am J Roentgenol 2007;188:1622-35. [Crossref] [PubMed]
6. Zhang Z, Huang F, Ma X, et al. Self-feeding MUSE: a robust method for high resolution diffusion imaging using interleaved EPI. Neuroimage 2015;105:552-60. [Crossref] [PubMed]
7. Porter DA, Heidemann RM. High resolution diffusion-weighted imaging using readout-segmented echo-planar imaging, parallel imaging and a two-dimensional navigator-based reacquisition. Magn Reson Med 2009;62:468-75. [Crossref] [PubMed]
8. Chen NK, Guidon A, Chang HC, et al. A robust multi-shot scan strategy for high-resolution diffusion weighted MRI enabled by multiplexed sensitivity-encoding (MUSE). Neuroimage 2013;72:41-7. [Crossref] [PubMed]
9. Hyun CM, Kim HP, Lee SM, Lee S, Seo JK. Deep learning for undersampled MRI reconstruction. Phys Med Biol. 2018;63:135007. [Crossref] [PubMed]
10. Ye X, Wang P, Li S, et al. Simultaneous superresolution reconstruction and distortion correction for single-shot EPI DWI using deep learning. Magn Reson Med 2023;89:2456-70. [Crossref] [PubMed]
11. Kofler A, Haltmeier M, Schaeffter T, et al. An end-to-end-trainable iterative network architecture for accelerated radial multi-coil 2D cine MR image reconstruction. Med Phys 2021;48:2412-25. [Crossref] [PubMed]
12. Afat S, Herrmann J, Almansour H, et al. Acquisition time reduction of diffusion-weighted liver imaging using deep learning image reconstruction. Diagn Interv Imaging 2023;104:178-84. [Crossref] [PubMed]
13. Wessling D, Gassenmaier S, Olthof SC, et al. Novel deep-learning-based diffusion weighted imaging sequence in 1.5 T breast MRI. Eur J Radiol 2023;166:110948. [Crossref] [PubMed]
14. Wilpert C, Neubauer C, Rau A, et al. Accelerated Diffusion-Weighted Imaging in 3 T Breast MRI Using a Deep Learning Reconstruction Algorithm With Superresolution Processing: A Prospective Comparative Study. Invest Radiol 2023;58:842-52. [Crossref] [PubMed]
15. Zhang X, Wang Y, Xu X, et al. Bladder MRI with deep learning-based reconstruction: a prospective evaluation of muscle invasiveness using VI-RADS. Abdom Radiol (NY) 2024;49:1615-25. [Crossref] [PubMed]
16. Ueda T, Ohno Y, Yamamoto K, et al. Deep Learning Reconstruction of Diffusion-weighted MRI Improves Image Quality for Prostatic Imaging. Radiology 2022;303:373-81. [Crossref] [PubMed]
17. Wang X, Litwiller D, Guidon A, Lan P, Sprenger T. Robust Complex Signal Averaging for Diffusion Weighted Imaging. In: Proceedings of the 23rd Annual Meeting of ISMRM; 2023 Jun 3-8; Toronto, Canada. Abstract 3963.
18. Lebel RM. Performance characterization of a novel deep learning-based MR image reconstruction pipeline. arXiv 2020. arXiv:2008.06559.
19. An H, Ma X, Pan Z, et al. Qualitative and quantitative comparison of image quality between single-shot echo-planar and interleaved multi-shot echo-planar diffusion-weighted imaging in female pelvis. Eur Radiol 2020;30:1876-84. [Crossref] [PubMed]
20. Lancaster JL, Cykowski MD, McKay DR, et al. Anatomical global spatial normalization. Neuroinformatics 2010;8:171-82. [Crossref] [PubMed]
21. Lancaster JL, Laird AR, Eickhoff SB, et al. Automated regional behavioral analysis for human brain images. Front Neuroinform 2012;6:23. [Crossref] [PubMed]
22. Lancaster JL, McKay DR, Cykowski MD, et al. Automated analysis of fundamental features of brain structures. Neuroinformatics 2011;9:371-80. [Crossref] [PubMed]
23. McHugh ML. Interrater reliability: the kappa statistic. Biochem Med (Zagreb) 2012;22:276-82.
24. Daimiel Naranjo I, Lo Gullo R, Morris EA, et al. High-Spatial-Resolution Multishot Multiplexed Sensitivity-encoding Diffusion-weighted Imaging for Improved Quality of Breast Images and Differentiation of Breast Lesions: A Feasibility Study. Radiol Imaging Cancer 2020;2:e190076. [Crossref] [PubMed]
25. Chang HC, Chen G, Chung HW, et al. Multi-shot Diffusion-Weighted MRI With Multiplexed Sensitivity Encoding (MUSE) in the Assessment of Active Inflammation in Crohn's Disease. J Magn Reson Imaging 2022;55:126-37. [Crossref] [PubMed]
26. Feng W, Zhu L, Xia Y, et al. Deep learning-based reconstruction: a reliability assessment in preoperative magnetic resonance imaging for primary rectal cancer. Quant Imaging Med Surg 2024;14:8927-41. [Crossref] [PubMed]
27. Wennmann M, Thierjung H, Bauer F, et al. Repeatability and Reproducibility of ADC Measurements and MRI Signal Intensity Measurements of Bone Marrow in Monoclonal Plasma Cell Disorders: A Prospective Bi-institutional Multiscanner, Multiprotocol Study. Invest Radiol 2022;57:272-81. [Crossref] [PubMed]
28. Herrmann J, Wessling D, Nickel D, et al. Comprehensive Clinical Evaluation of a Deep Learning-Accelerated, Single-Breath-Hold Abdominal HASTE at 1.5 T and 3 T. Acad Radiol 2023;30:93-102. [Crossref] [PubMed]
29. Ren J, Li Y, Liu FS, et al. Comparison of a deep learning-accelerated T2-weighted turbo spin echo sequence and its conventional counterpart for female pelvic MRI: reduced acquisition times and improved image quality. Insights Imaging 2022;13:193. [Crossref] [PubMed]
30. Herrmann J, Benkert T, Brendlin A, et al. Shortening Acquisition Time and Improving Image Quality for Pelvic MRI Using Deep Learning Reconstruction for Diffusion-Weighted Imaging at 1.5 T. Acad Radiol 2024;31:921-8. [Crossref] [PubMed]