Validation and feasibility of fast knee MRI using a deep learning-assisted 3D iterative image enhancement system

Introduction

Knee pain is a common disease with various etiologies (1,2), among which osteoarthritis is particularly significant, affecting over 500 million individuals worldwide and involving the knee joint as the most commonly affected site (3). Knee pain can limit patients’ function and mobility, and eventually reduce the quality of life (4). Magnetic resonance imaging (MRI) is a non-invasive modality widely employed in the evaluation of knee injuries due to its high resolution, multi-planar capabilities, and multi-contrast imaging, which enable comprehensive assessment of menisci, ligaments, bone marrow, synovial fluid, fat, and other tissues (5,6).

Despite its advantages, the efficiency of conventional MRI of the knee remains suboptimal, with standard knee MRI protocols lasting more than 20 minutes (7). It poses significant challenges, particularly for patients with joint stiffness who may experience discomfort or pain during prolonged scanning, leading to motion artifacts, compromised image quality, or even exam failure (8). Hence, fast knee MRI is of great clinical significance for improving patient experience. Zhang et al. (9) reported that using deep learning (DL)-assisted compressed sensing yielded higher-quality images within the same scanning time, while Moon et al. (10) showed that ultrafast MRI could reduce scanning time by more than 80% and still provide acceptable results. These results demonstrated the possibility of substantially shortening scan times and enabling rapid or ultrafast imaging across multiple anatomical regions.

Reduced scanning parameters, combined with image-domain reconstruction algorithms, can also achieve high-quality, fast MRI. Currently, many studies have adopted convolutional neural networks (CNNs) for medical image super-resolution reconstruction, and these methods have been proven effective (11,12). For knee MRI, DL-based super-resolution techniques have also been widely applied. Vosshenrich et al. (13) showed that DL super-resolution can reduce the knee MRI scanning time to less than 5 minutes and demonstrated great performance in detecting internal derangements, while Walter et al. (14) compared conventional and DL-based knee MRI, showing that super-resolution knee MRI provided better image quality with similar diagnostic performance. DL-assisted three-dimensional iterative image enhancement (DL-3DIIE) provides another practical solution to address fast MRI. By integrating an image enhancement system between the MRI scanner and Picture Archiving and Communication Systems (PACS), DL-3DIIE reconstructs low-quality source images into high-quality images before storage, thereby improving image usability. Its feasibility has been preliminarily validated in spine (15) and prostate (16) imaging. In this study, DL-3DIIE was applied to optimize low-parameter knee MRI images (by reducing the number of excitations and in-plane resolution) acquired on scanners that do not support advanced acceleration technologies. Image quality was compared with conventional MRI and fast MRI, aiming to evaluate the feasibility of achieving high-resolution, rapid knee MRI under relatively limited clinical conditions.

Methods

Study design and participants

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This prospective study was approved by the Ethics Committee of Northern Jiangsu People’s Hospital (No. 2023js050), and informed consent was obtained from all participants. A total of 134 participants scheduled for knee MRI plain scans at the study hospital between September 2023 and January 2024 were prospectively included. The inclusion criteria were (I) participants without contraindications to MRI examination and (II) age >18 years. The exclusion criteria were (I) participants with previous knee surgery; (II) images with inadequate quality; or (III) incomplete scanning sequences.

MR acquisition

All images were acquired using a 3.0-T MR scanner (Signa HDx, GE, USA) equipped with a 12-channel knee coil. The participants were scanned using both high-resolution, high-scanning-resolution, long-acquisition-time conventional parameters to obtain high-resolution conventional images, and low-resolution, low-scanning-resolution, short-acquisition-time parameters to acquire low-resolution raw images. The low-resolution raw images were then optimized using the composite super-resolution reconstruction algorithm to generate high-resolution optimized reconstruction images. The detailed parameters and time for imaging sequences are shown in Table S1. This study included three imaging protocols: conventional MRI with standard parameters, fast MRI, in which acquisition time was shortened by reducing the number of excitations (NEX) and in-plane resolution, at the expense of signal-to-noise ratio (SNR), and DL-3DIIE MRI, in which fast images were optimized using a DL-3DIIE system to restore SNR and edge sharpness through iterative denoising and super-resolution. The fast protocol reduced acquisition time to 4 min 10 s, a 44.57% reduction compared with the conventional protocol (7 min 31 s).

Composite super-resolution reconstruction algorithm

The composite super-resolution reconstruction algorithm was based on the FDA-certified IQMR platform (https://medicvision.com/en/iqmr, Medic Vision Company, USA, Accessed 8 November 2021). It is a general 3D image enhancement system based on CNN-based image super-resolution reconstruction, aiming to improve image resolution while reducing noise. The algorithm is mainly divided into two modules: a super-resolution reconstruction module and a noise-reduction sharpener module, which post-process the original image. The model structure is shown in Figure 1. The basic CNN architecture for the super-resolution reconstruction module is a generative adversarial network (GAN), combined with the adjusted filter and loss function. The image resolution of the input data set is controlled by multiple parameters to match the processing parameter set of the iterative reconstruction algorithm and to obtain the best-resolution output image. The noise reduction and sharpening module aims to improve image sharpness and reduce noise. The CNN uses a fixed parameter set, and the algorithm employs a fixed nonlinear filter for data processing to enhance the visibility of small image details and improve image sharpness. Grouping is performed based on unique similarity measures by computing 3D features of the input images and performing a discrete Fourier transform in k-space. The similarity and noise statistical predictions are combined to separate the noise from the signal until the convergence criterion is reached (Figure 1). However, due to the proprietary nature of commercial software, the detailed model development processes cannot be provided.

Figure 1 Flow chart of the composite super-resolution reconstruction algorithm. IPPM, iterative process of image enhancement and parameter modulation; SRRM, super-resolution reconstruction module.

Imaging analysis

The primary measurement outcome of this study was the subjective image quality score based on clinicians’ judgement. The image quality score was calculated according to the method proposed by Shakoor et al. (17). Two experienced senior doctors (with 16- and 21-years’ experience, respectively) in musculoskeletal imaging diagnosis used a double-blind method to evaluate the image quality of each group according to the lesion, margin, artifact, and comprehensive diagnosis using a 5-point scale (5-point: excellent, 4-point: great, 3-point: acceptable, 2 points: hard, 1-point: unsuitable). At the same time, images were scored according to the comprehensive diagnostic score. When opinions were inconsistent, an agreement was reached through discussion.

The secondary outcome measurement was peak signal-to-noise ratio (PSNR), multi-scale structural similarity index (MS-SSIM), SNR, and contrast-to-noise ratio (CNR). To evaluate the similarity between images of the reconstructed and reference groups, two evaluation indices, namely PSNR and MS-SSIM (18), were calculated using MATLAB (R2023b, MathWorks, USA). Using the high-resolution conventional images as a reference, PSNR and MS-SSIM were calculated between the low-resolution raw reconstruction images and the conventional group, as well as between the high-resolution conventional and the optimized reconstruction group (Appendix 1). PSNR is an index based on the mean squared error that measures the difference between the image’s effective information and its noise. The larger the value, the smaller the image distortion. Generally, a PSNR greater than 40 dB represents excellent image quality. MS-SSIM assesses the similarity of images across luminance, contrast, and structure, with values closer to 1 indicating greater similarity between the two sets of images.

The SNR and CNR of bone marrow, cartilage, meniscus, ligament, muscle, fat, and joint fluid were measured by two senior technicians on a GE AW4.6 post-processing workstation, using a double-blind method to obtain objective image quality metrics. According to Altahawi et al. (19), 5–15 mm² circular regions of interest (ROIs) were selected from the air in the anterior aspect of the proximal level of the anterior cruciate ligament, the first complete level of the lateral femoral condyle, the patellar cartilage, the lateral main part of the meniscus, the distal part of the posterior cruciate ligament, the medial head of the gastrocnemius muscle, the infrapatellar fat pad, and the joint fluid of the lateral condyle in each sagittal sequence image. When selecting the ROIs, lesions and artefacts were avoided, and the position and area of each ROI in its corresponding part were kept consistent.

The SNR and CNR calculation formulas are as follows:

SNR=SItissue/SDairCNR=|(SItissue1−SItissue2)/SDair|[1]

Data collected included age, sex, height, body weight, body mass index (BMI), knee joint positions, and primary diagnoses.

Statistical analysis

All statistical analyses were performed using SPSS 22.0 (IBM Corp., Armonk, N.Y., USA). The Kolmogorov-Smirnov test was used to assess the normality of continuous variables. Continuous variables with normal distribution were presented as mean ± SD, and those with skewed distribution were presented as median (lower quartile, upper quartile). For SNR and CNR, one-way repeated-measures ANOVA was applied, followed by the Bonferroni-corrected paired t-test for post hoc comparisons. PSNR and MS-SSIM were compared using a paired-samples t-test. Subjective image quality scores were analyzed using the Friedman test, with Nemenyi post hoc pairwise comparisons. The intraclass correlation coefficient (ICC) was calculated to evaluate interobserver consistency of subjective image quality scores, with agreement interpreted as slight (<0.20), fair (≥0.20 and <0.40), moderate (≥0.40 and <0.60), substantial (≥0.60 and <0.80), and almost perfect (≥0.80). Categorical data were expressed as n (%) and compared using the chi-squared test. A two-sided P<0.05 was considered statistically significant.

Results

A total of 134 patients, with a mean age of 55.1±9.5 years and 46.3% male, were included in the final analysis. Of the 134 knees assessed, 59 (44.0%) were left knees. The mean BMI was 22.4±5.6 kg/m². The predominant diagnoses were degenerative disease (n=75, 56.0%) and trauma (n=57, 42.5%) (Table 1).

For the sagittal PDWI-FSE sequence, SNRs differed significantly among the three groups for all tissues (all P<0.001). Compared with conventional MRI, DL-3DIIE MRI showed significantly higher SNRs in bone marrow (15.96±3.41 vs. 31.25±5.61, P<0.05), cartilage (47.23±12.35 vs. 72.79±14.12, P<0.05), meniscus (12.56±4.68 vs. 20.38±6.95, P<0.05), ligament (7.85±2.21 vs. 13.12±4.98, P<0.05), muscle (34.11±7.22 vs. 55.54±12.92, P<0.05), fat (33.45±9.20 vs. 68.59±14.10, P<0.05), and joint fluid (147.56±24.75 vs. 233.84±38.85, P<0.05). DL-3DIIE MRI provided significantly higher SNRs than both conventional and fast MRI across all tissues (all P<0.05). For the sagittal T1WI-FSE sequence, similar results were observed (all P<0.001). DL-3DIIE MRI yielded the highest SNRs for all tissues compared with both conventional and fast MRI (all P<0.05) (Figure 2 and Table S2).

Figure 2 Comparison of SNR among conventional MRI, fast MRI, and DL-3DIIE MRI. DL-3DIIE, deep learning-assisted 3D iterative image enhancement system; MRI, magnetic resonance imaging; PDWI-FSE, proton density-weighted imaging-fast spin echo; SNR, signal-to-noise ratio; T1WI-FSE, T1-weighted imaging-fast spin echo.

For the sagittal PDWI-FSE sequence, CNRs differed significantly among the three groups for all tissue contrasts (all P<0.001). Compared with conventional MRI, DL-3DIIE MRI demonstrated significantly higher CNRs in cartilage-joint fluid (13.33±3.59 vs. 29.56±7.85, P<0.05), cartilage-bone marrow (4.24±1.79 vs. 7.51±2.95, P<0.05), meniscus-joint fluid (17.52±3.38 vs. 38.01±6.52, P<0.05), ligament-joint fluid (18.85±3.56 vs. 39.56±6.90, P<0.05), bone marrow-joint fluid (17.59±3.89 vs. 36.09±6.65, P<0.05), fat-joint fluid (15.59±3.50 vs. 29.59±7.62, P<0.05), and muscle-joint fluid (14.84±3.38 vs. 32.51±7.69, P<0.05). DL-3DIIE MRI also yielded significantly higher CNRs than fast MRI across all tissue contrasts (all P<0.05). For the sagittal T1WI-FSE sequence, similar results were observed (all P<0.001). Compared with conventional MRI, DL-3DIIE MRI achieved higher CNRs in cartilage-joint fluid (4.56±1.95 vs. 7.79±1.99, P<0.05), cartilage-bone marrow (25.18±7.95 vs. 32.07±7.93, P<0.05), meniscus-joint fluid (6.04±2.41 vs. 10.08±2.79, P<0.05), ligament-joint fluid (3.88±1.82 vs. 5.36±2.21, P<0.05), bone marrow-joint fluid (23.85±7.58 vs. 27.48±7.15, P<0.05), fat-joint fluid (4.98±2.85 vs. 7.48±2.91, P<0.05), and muscle-joint fluid (5.22±2.48 vs. 7.45±2.59, P<0.05). DL-3DIIE MRI consistently provided the highest CNRs compared with both conventional and fast MRI (all P<0.05) (Figure 3 and Table S3).

Figure 3 Comparison of CNR among conventional MRI, fast MRI, and DL-3DIIE MRI. CNR, contrast-to-noise ratio; DL-3DIIE, deep learning-assisted 3D iterative image enhancement system; MRI, magnetic resonance imaging; PDWI-FSE, proton density-weighted imaging-fast spin echo; T1WI-FSE, T1-weighted imaging-fast spin echo.

For quantitative image quality metrics, DL-3DIIE MRI achieved significantly higher PSNR and MS-SSIM values than fast MRI across all sequences (all P<0.001). Specifically, for PSNR, DL-3DIIE MRI outperformed fast MRI in the axial PDWI-FSE (70.75±1.56 vs. 65.84±1.52, P<0.001), sagittal PDWI-FSE (72.12±1.48 vs. 65.99±1.42, P<0.001), sagittal T1WI-FSE (69.73±1.49 vs. 67.12±1.31, P<0.001), and coronal PDWI-FSE (71.31±1.39 vs. 66.79±1.33, P<0.001) sequences. For MS-SSIM, DL-3DIIE MRI also yielded higher values than fast MRI in the axial PDWI-FSE (0.96±0.01 vs. 0.93±0.01, P<0.001), sagittal PDWI-FSE (0.96±0.02 vs. 0.92±0.01, P<0.001), sagittal T1WI-FSE (0.95±0.01 vs. 0.93±0.02, P<0.001), and coronal PDWI-FSE (0.96±0.01 vs. 0.94±0.01, P<0.001) sequences (Figure 4 and Table S4).

Figure 4 PSNR and MS-SSIM for fast MRI and DL-3DIIE MRI. DL-3DIIE, deep learning-assisted 3D iterative image enhancement system; MRI, magnetic resonance imaging; MS-SSIM, multi-scale structural similarity index; PDWI-FSE, proton density-weighted imaging-fast spin echo; PSNR, peak signal-to-noise ratio; T1WI-FSE, T1-weighted imaging-fast spin echo.

In subjective image quality assessment, the DL-3DIIE group demonstrated significantly higher scores for lesion conspicuity, margin definition, and overall diagnostic confidence than the conventional group (all P<0.05, both readers). In contrast, no significant differences were observed in motion artifact scores between the two groups (P=0.911 and 0.895 for the two readers, respectively) (Figure 5). The interobserver agreement for subjective image quality was good to excellent, with ICCs of 0.93, 0.88, 0.82, and 0.92 for lesion conspicuity, margin delineation, and motion artifact, respectively (Table S5).

Figure 5 Subjective image quality assessment. DL-3DIIE, deep learning-assisted 3D iterative image enhancement system.

Figure 6 presents the PD and T1 sequences of a 56-year-old female patient with limited right knee movement for 3 months.

Figure 6 A 56-year-old female patient has had limited movement of the right knee for 3 months. (A) At the bony attachment site of the lateral patellar retinaculum, small patchy hyperintensities are noted, which are suggestive of bone marrow edema. The imaging quality of DL-3DIIE MRI is significantly superior to that of fast MRI and slightly clearer than conventional MRI images. (B) A thin linear hyperintense signal on T2-weighted imaging is noted in the posterior horn of the medial meniscus, which is suggestive of a meniscal tear; small patchy hyperintensities are seen in the tibial plateau, suggestive of bone marrow edema. On DL-3DIIE MRI scans, the boundaries of bone marrow edema are the clearest; the linear hyperintense signals of meniscal tears are visualized most sharply. Its imaging quality is significantly superior to that of fast MRI and slightly better than that of conventional MRI. (C) Patchy hypointense signals on T1WI are noted in the subchondral bone of the patella, with ill-defined boundaries, which are suggestive of bone marrow edema. The imaging quality of DL-3DIIE MRI is significantly superior to that of fast MRI and slightly clearer than conventional MRI images. (D) On T2WI images, a small patchy hyperintense signal is seen in the body of the lateral meniscus, extending to the articular surface, suggesting a meniscal tear. On DL-3DIIE MRI scans, the edges of the meniscal tear are visualized most sharply, which is significantly superior to fast MRI images and slightly clearer than conventional MRI images. The red frames indicate locally enlarged regions. DL-3DIIE, deep learning-assisted 3D iterative image enhancement system; MRI, magnetic resonance imaging; T1WI, T1-weighted imaging; T2WI, T2-weighted imaging.

Discussion

The results of this study indicate that, in subjective evaluations, DL-based reconstructed images showed improved lesion conspicuity, clearer margin delineation, and higher diagnostic confidence than conventional MRI, while motion artifact levels were comparable between the two methods. In terms of quantitative metrics, the reconstructed images showed significantly higher PSNR and lower MS-SSIM values than the conventional images. Additionally, the reconstructed images showed higher SNR and CNR than those in the conventional group. These findings highlight the promising potential of this method for improving image quality and for clinical applications.

Super-resolution reconstruction algorithms for medical images can be roughly classified into interpolation, reconstruction, and DL algorithms. Although the interpolation method is easy to perform, it cannot provide additional information, which can easily distort the reconstructed image, producing the so-called “waxy” effect and further affecting the accurate detection of lesions (20). Unlike the interpolation method, the reconstruction method can reduce edge distortion and blur to a certain extent. However, it requires more computing resources for image registration (21,22). In the present study, a CNN-based reconstruction algorithm was adopted, combined with an adjusted filter block and a loss function. By fine-tuning the pre-trained model, the proposed model and parameters effectively restored a low-resolution image to a high-resolution, clearer image, reducing acquisition time by 44.57%.

Previous studies have shown that DL-based reconstruction is effective in improving SNR, resolution, and acquisition time (23,24), consistent with the present study. Indeed, the results revealed that the SNR and CNR of bone marrow, cartilage, meniscus, ligament, muscle, fat, and synovial fluid in both PD and T1 in the reconstruction group were significantly increased. The discrimination ability of the margin, motion artefact, and comprehensive diagnosis was significantly improved with simultaneous reconstruction, which could be partially explained by the increased sharpness of the reconstructed images (25). Apart from the effective algorithm, shortening the scanning time with low-resolution raw images also reduces unintentional movement caused by patient intolerance and improves the examination success rate (26). Although some studies have shown that DL algorithms can eliminate motion artefacts from the k-space level (27,28), this perspective needs to be further verified for this study. The underlying reason could be partially explained by the DL-3DIIE processing, which primarily focuses on enhancing SNR, CNR, and resolution while providing limited correction for existing motion artifacts. The PSNRs of the conventional and reconstruction groups were both greater than 40 dB, and the MS-SSIM was very close to 1, indicating that there were no apparent structural or signal changes in the post-processed images. In a previous study, Ma et al. (29) proposed a CNN-based reconstruction model for brain and knee MRI, achieving PSNR of 34.19 dB and MS-SSIM of 0.8994, which were significantly lower than those in the present study. The above indicators demonstrate that the composite super-resolution reconstruction algorithm can improve image resolution and reduce background noise without losing inherent image information or altering tissue structure across multiple dimensions.

From a clinical application perspective, patients expect shorter examination times and greater comfort, while radiologists need higher image quality to improve diagnostic accuracy, and medical institutions seek greater examination efficiency to reduce overall waiting time (30). The application of composite super-resolution reconstruction technology based on DL further balances the demands of patients, readers and medical institutions. In addition, a 15-year-old MRI machine was used in the present study, and its outdated hardware is associated with poor imaging quality. The incorporation of DL-based reconstruction technology significantly shortened the scanning time and produced high-quality images.

There are certain limitations in this study. Firstly, it was a single-centre study with a relatively small sample. The generalization and stability of the proposed algorithm should be verified with a larger sample size and multicenter joint cohorts. Secondly, the proposed model was built based on a commercial DL platform, and the technical details of the model are not available. Nevertheless, further optimization of the model structure could improve model performance. Thirdly, this study did not include healthy participants without knee lesions. As a result, the DL reconstruction algorithm could not be evaluated for identifying true negatives. Future studies should consider including a healthy control group to comprehensively assess the algorithm’s diagnostic performance, particularly its specificity. The inclusion of healthy controls should be considered in future validations.

Conclusions

In conclusion, the composite super-resolution reconstruction algorithm based on DL can improve image details and sharpness while preserving image integrity. Suitable adjustments to scanning parameters, combined with super-resolution reconstruction technology in the clinical setting, can significantly reduce scanning time, improve image quality, and facilitate precise diagnosis of knee MRI.

Acknowledgments

The authors would like to thank Jie Shi (Department of MR Research, GE Healthcare, Shanghai, China) for the valuable assistance in manuscript revision, editing, and project guidance.

Footnote

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-2099/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-2025-2099/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 Ethics Committee of Northern Jiangsu People’s Hospital (No. 2023js050), and informed consent was taken from all 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. Langworthy M, Dasa V, Spitzer AI. Knee osteoarthritis: disease burden, available treatments, and emerging options. Ther Adv Musculoskelet Dis 2024;16:1759720X241273009.
2. Ma Z, Liu Y, Zhang Z, Chen R, Fan H, Cao X, Ni L. Clinical applications of large language models in knee osteoarthritis: a systematic review. Front Med (Lausanne) 2025;12:1670824. [Crossref] [PubMed]
3. Geng R, Li J, Yu C, Zhang C, Chen F, Chen J, Ni H, Wang J, Kang K, Wei Z, Xu Y, Jin T. Knee osteoarthritis: Current status and research progress in treatment Exp Ther Med 2023;26:481. (Review). [Crossref] [PubMed]
4. Brophy RH, Fillingham YA. AAOS Clinical Practice Guideline Summary: Management of Osteoarthritis of the Knee (Nonarthroplasty), Third Edition. J Am Acad Orthop Surg 2022;30:e721-9.
5. Lemainque T, Huppertz MS, Yüksel C, Siepmann R, Kuhl C, Roemer F, Truhn D, Nebelung S. Current MR imaging of cartilage in the context of knee osteoarthritis (part 1): Principles and sequences. Radiologie (Heidelb) 2024;64:295-303. [Crossref] [PubMed]
6. Herrera D, Almhdie-Imjabbar A, Toumi H, Lespessailles E. Magnetic resonance imaging-based biomarkers for knee osteoarthritis outcomes: A narrative review of prediction but not association studies. Eur J Radiol 2024;181:111731. [Crossref] [PubMed]
7. Boutin RD, Eshed I, Kassarjian A, Vemuri NV. The Global Reading Room: Knee MRI Protocols. AJR Am J Roentgenol 2022;219:347-8. [Crossref] [PubMed]
8. Vidya Shankar R, Chang JC, Hu HH, Kodibagkar VD. Fast data acquisition techniques in magnetic resonance spectroscopic imaging. NMR Biomed 2019;32:e4046. [Crossref] [PubMed]
9. Zheng G, Fu J, Wang Z, Li W, Li A, Yu D. AI-assisted compressed sensing MRI improves imaging quality in rectal cancer: a comparative study with conventional acceleration techniques. Quant Imaging Med Surg 2025;15:2547-60. [Crossref] [PubMed]
10. Moon HE, Ha JY, Choi JW, Lee SH, Hwang JY, Choi YH, Cheon JE, Cho YJ. Ultrafast MRI for Pediatric Brain Assessment in Routine Clinical Practice. Korean J Radiol 2025;26:75-87. [Crossref] [PubMed]
11. Hahn S, Yi J, Lee HJ, Lee Y, Lim YJ, Bang JY, Kim H, Lee J. Image Quality and Diagnostic Performance of Accelerated Shoulder MRI With Deep Learning-Based Reconstruction. AJR Am J Roentgenol 2022;218:506-16. [Crossref] [PubMed]
12. Bai Z, Tao J. Super-Resolution Reconstruction of Brain MR Images Using Pseudo-3D Convolutional Network. Journal of Computer-Aided Design & Computer Graphics 2022;34:208-16.
13. Vosshenrich J, Breit HC, Donners R, Obmann MM, Walter SS, Serfaty A, Rodrigues TC, Recht M, Stern SE, Fritz J. Clinical Implementation of Sixfold-Accelerated Deep Learning Superresolution Knee MRI in Under 5 Minutes: Arthroscopy-Validated Diagnostic Performance. AJR Am J Roentgenol 2025;225:e2532878. [Crossref] [PubMed]
14. Walter SS, Vosshenrich J, Cantarelli Rodrigues T, Dalili D, Fritz B, Kijowski R, Park EH, Serfaty A, Stern SE, Brinkmann I, Koerzdoerfer G, Fritz J. Deep Learning Superresolution for Simultaneous Multislice Parallel Imaging-Accelerated Knee MRI Using Arthroscopy Validation. Radiology 2025;314:e241249. [Crossref] [PubMed]
15. Yao H, Jia B, Pan X, Sun J. Validation and Feasibility of Ultrafast Cervical Spine MRI Using a Deep Learning-Assisted 3D Iterative Image Enhancement System. J Multidiscip Healthc 2024;17:2499-509. [Crossref] [PubMed]
16. Zhao Y, Xie XL, Zhu X, Huang WN, Zhou CW, Ren KX, Zhai RY, Wang W, Wang JW. FOCUS-DWI improves prostate cancer detection through deep learning reconstruction with IQMR technology. Abdom Radiol (NY) 2026;51:1276-88. [Crossref]
17. Shakoor D, Kijowski R, Guermazi A, Fritz J, Roemer FW, Jalali-Farahani S, Eng J, Demehri S. Diagnosis of Knee Meniscal Injuries by Using Three-dimensional MRI: A Systematic Review and Meta-Analysis of Diagnostic Performance. Radiology 2019;290:435-45. [Crossref] [PubMed]
18. Horé A, Ziou D. Image Quality Metrics: PSNR vs. SSIM. 2010 20th International Conference on Pattern Recognition 2010;2010:23-26.
19. Altahawi FF, Blount KJ, Morley NP, Raithel E, Omar IM. Comparing an accelerated 3D fast spin-echo sequence (CS-SPACE) for knee 3-T magnetic resonance imaging with traditional 3D fast spin-echo (SPACE) and routine 2D sequences. Skeletal Radiol 2017;46:7-15. [Crossref] [PubMed]
20. Zhang H, Zhang L, Shen H. A Blind Super-Resolution Reconstruction Method Considering Image Registration Errors. Int J Fuzzy Syst 2015;17:353-64.
21. Du J, He Z, Wang L, Gholipour A, Zhou Z, Chen D, Jia Y. Super-resolution reconstruction of single anisotropic 3D MR images using residual convolutional neural network. Neurocomputing 2020;392:209-20.
22. Qiu D, Cheng Y, Wang X. Medical image super-resolution reconstruction algorithms based on deep learning: A survey. Comput Methods Programs Biomed 2023;238:107590. [Crossref] [PubMed]
23. Vosshenrich J, Fritz J. Accelerated musculoskeletal magnetic resonance imaging with deep learning-based image reconstruction at 0.55 T-3 T. Radiologie (Heidelb) 2024;64:758-65. [Crossref] [PubMed]
24. He W, Xu P, Zhang M, Xu R, Shen X, Mao R, Li XH, Sun CH, Zhang RN, Lin S. Deep learning-based reconstruction improves the image quality of low-dose CT enterography in patients with inflammatory bowel disease. Abdom Radiol (NY) 2025;50:3907-16. [Crossref] [PubMed]
25. Park C, Choo KS, Jung Y, Jeong HS, Hwang JY, Yun MS. CT iterative vs deep learning reconstruction: comparison of noise and sharpness. Eur Radiol 2021;31:3156-64. [Crossref] [PubMed]
26. Mirzai M, Nilsson J, Sund P, Norrlund RR, Diniz MO, Gottfridsson B, Häggström I, Johnsson ÅA, Båth M, Svalkvist A. Breathing motion compensation in chest tomosynthesis: evaluation of the effect on image quality and presence of artifacts. J Med Imaging (Bellingham) 2025;12:S13004. [Crossref] [PubMed]
27. Dabrowski O, Falcone JL, Klauser A, Songeon J, Kocher M, Chopard B, Lazeyras F, Courvoisier S. SISMIK for Brain MRI: Deep-Learning-Based Motion Estimation and Model-Based Motion Correction in k-Space. IEEE Trans Med Imaging 2025;44:396-408. [Crossref] [PubMed]
28. Ota H, Morita Y, Vucevic D, Higuchi S, Takagi H, Kutsuna H, Yamashita Y, Kim P, Miyazaki M. Motion robust coronary MR angiography using zigzag centric ky-kz trajectory and high-resolution deep learning reconstruction. MAGMA 2024;37:1105-17. [Crossref] [PubMed]
29. Ma Q, Lai Z, Wang Z, Qiu Y, Zhang H, Qu X. MRI reconstruction with enhanced self-similarity using graph convolutional network. BMC Med Imaging 2024;24:113. [Crossref] [PubMed]
30. Seghier ML. 7 T and beyond: toward a synergy between fMRI-based presurgical mapping at ultrahigh magnetic fields, AI, and robotic neurosurgery. Eur Radiol Exp 2024;8:73. [Crossref] [PubMed]