Prospective Evaluation of Accelerated Brain MRI Using Deep Learning-Based Reconstruction: Simultaneous Application to 2D Spin-Echo and 3D Gradient-Echo Sequences

Go to:

INTRODUCTION

Magnetic resonance imaging (MRI) is indispensable for functional and anatomical brain evaluation, continually benefiting from advancements in spatial resolution and signal-to-noise ratio (SNR). The aging global population has led to an increase in the incidence of stroke and vascular dementia [1]. Consequently, the demand for brain MRI increased by one-third from 2009 to 2019, with approximately 2.5 million exams in the United States [2]. However, challenges such as limited MRI availability and lengthy scan times have resulted in waiting periods of up to 18 weeks, with 26% of patients experiencing waiting times of over 6 weeks [3].

To address the issue of prolonged acquisition times, techniques such as parallel imaging and undersampling, including compressed sensing, have been employed, although they can reduce the SNR and spatial resolution [4, 5]. Recent advancements in deep learning-based reconstruction (DLR) algorithms have reduced the scan time without compromising image quality in various organs, including the brain [5], knee [6], and spine [7, 8]. These algorithms enable detailed postoperative imaging of small organs such as the pituitary gland by denoising and reducing artifacts [9].

Bash et al. [5] reported acceptable image quality and volumetric results for accelerated 3D T1-weighted brain MR images reconstructed with DL in a prospective, multicenter, multi-reader study. Although routine brain MRI typically includes 2D spin-echo (SE) T2-weighted imaging (T2WI), T2 fluid-attenuated inversion recovery (T2-FLAIR), and T1-weighted imaging (T1WI), the image quality of accelerated DL-based reconstruction (Accel-DL) of these sequences has not yet been evaluated.

Initially, DLR referred to the reconstruction of images from undersampled k-space data for accelerated MRI [10]. However, over the last decade, DLR has been expanded to include denoising [11] and super-resolution [12]. DLR now encompasses convolutional neural network-based models with supervised learning, regardless of whether processing occurs in the image domain, k-space domain, or through direct mapping. Despite these advancements, additional evidence is required to fully integrate DLR into clinical settings, particularly to understand how it may alter imaging characteristics and potentially obscure small lesions [13].

In this prospective, multi-sequence, multi-reader, multi-vendor study, we evaluated the effect of Accel-DL on improving the quality of brain MRI and reducing the scan time compared to that of conventional MRI, which uses 2D SE and 3D gradient-echo (GRE) sequences in clinical settings.

Go to:

MATERIALS AND METHODS

Study Population

This prospective study was approved by the Institutional Review Board of Seoul National University Hospital (IRB No. 2205-116-1327). Patients who visited the neurology clinic between August 2022 and February were consecutively enrolled. The inclusion criteria were as follows: a) age of 18 years or older, b) non-acute symptoms such as dizziness, headaches, or subtle sensory changes (i.e., without acute focal neurologic deficits), and c) requirement of brain MRI for further evaluation. The exclusion criteria were as follows: a) contraindications for MRI (e.g., pacemakers, cochlear implants, or claustrophobia), b) lack of consent to participate, or c) impaired capacity to consent.

MRI Acquisition

MRI examinations were performed using three scanners from three different vendors: Siemens (Magnetom Skyra 3T; Erlangen, Germany), Philips (Ingenia CX 3T; Amsterdam, Netherlands), and GE (Discovery 750w 3T; Chicago, IL, USA). Patients were randomly assigned to the different MRI scanners. The patients underwent both conventional and Accel-DL brain MRI protocols, including 2D axial T1WI, T2WI, T2-FLAIR, and 3D T1WI. Scanning parameters are listed in Supplementary Table 1.

Deep-Learning Based Reconstruction

Initially, accelerated MRI acquisition was achieved by employing specifically designed MRI scan parameters that reduced the scan time. The level of acceleration achievable without significant artifact deterioration varied depending on the conventional scan parameters specific to each sequence. Subsequently, the acquired digital imaging and communications in medicine (DICOM) images were reconstructed using a commercially available convolutional neural network-based MR image reconstruction software (SwiftMR, v2.0.1.0; AIRS Medical, Seoul, Korea) within the DICOM domain. A DICOM-based U-Net architecture was developed in a supervised manner and trained on numerous image pairs to transform low-quality images from undersampled k-space data into high-quality images [14]. This approach helped to distinguish noise from details under various imaging conditions, including different contrasts, sequences, and field strengths (detailed in Supplementary Fig. 1).

Qualitative Analysis

Four experienced neuroradiologists (Y.H.J., with 2 years of experience; K.H.L., with 6 years of experience; K.S.C., with 7 years of experience; and J.Y.L., with 8 years of experience) assessed the image quality at standard workstations. A typical randomized crossover design with a 4-week washout period was used [15]. Specifically, randomized datasets were presented for rating, with reconstruction methods and clinical information blinded by removing and anonymizing the DICOM headers.

For each reader, reconstruction (i.e., Accel-DL vs. conventional MRI), and participant (n = 150), 54 items (i.e., axial T1WI = 11, axial T2WI = 15, axial T2-FLAIR = 17, and 3D T1WI = 11) were included across four categories: overall image quality; structure delineation; artifacts; and intracranial lesions, including lesion conspicuity. Overall image quality and artifacts were evaluated using a 5-point scale, whereas structural delineation and lesion conspicuity were assessed using a 3-point scale [8, 16]. For axial T1WI and 3D T1WI, the focus areas included gray matter-white matter (GM-WM) differentiation and the basal ganglia. The midbrain and cerebellum were primarily evaluated using axial T2-FLAIR to best visualize these areas. Artifacts on axial T1WI and 3D T1WI included motion and flow, with axial T2-FLAIR assessing sulcal hyperintensity, specifically cerebrospinal fluid artifacts, in the prepontine cistern. Finally, intracranial lesion conspicuity, particularly WM hyperintensities (WMHs), and the Fazekas scale for periventricular WM on axial T2-FLAIR were assessed [17], along with the enlarged perivascular space (PVS) count on axial T2WI. Further details are provided in the Supplement.

Quantitative Analysis

Axial T1WI, T2WI, T2-FLAIR, SNR, and contrast-to-noise ratio (CNR) were calculated by one neuroradiologist (K.S.C., with 7 years of experience), using a manually drawn spherical region of interest (ROI). ROIs measurements were performed on the largest axial T2-FLAIR WMH lesion (Slesion), encompassing the lesion while ensuring that it did not extend beyond its boundaries to mitigate measurement error. Next, an ROI of identical size and level as the lesion was placed in the contralateral cerebral WM (Sparenchyma) and airspace outside the cranium (Snoise) [18]. Finally, the SNR and CNR were calculated for both 3D T1WI and axial T2-FLAIR as follows: 1) SNR = Slesion/Snoise or Sparenchyma/Snoise when no lesions were detected and 2) CNR = (Slesion − Sparenchyma)/Snoise.

For volumetric analysis, brain segmentation and parcellation of the cortical and subcortical structures were performed using the fully automated NeuroQuant (CorTechs Lab, San Diego, CA, USA) and 3D T1WI [19]. For segmenting WMH on axial T2-FLAIR images, LesionQuant (CorTechs Lab, San Diego, CA, USA) categorized the lesion into four distinct groups based on their location: leukocortical, periventricular, infratentorial, and deep WM (detailed in the Supplement).

Statistical Analysis

For qualitative analysis, the scores from four readers were averaged, and Wilcoxon signed-rank test was used to compare the average image quality scores between Accel-DL and conventional MRI. Fleiss’ kappa was used to assess inter-reader reliability [20]. For further analysis, we used linear mixed models (LMMs) to assess the impact of multiple readers, reconstruction, and multiple vendors. Specifically, we modelled vendors and reconstruction as fixed effects and readers as a random effect, with and without interaction between the reconstruction and vendors [8]. The model fitting of LMMs with and without interaction was then compared using analysis of variance. For quantitative analysis, paired t-test was used to compare quantitative measures, including SNR and CNR, between Accel-DL and conventional MRI. For agreement in volumetric analysis, Pearson’s correlation coefficient was calculated. Bland-Altman analysis was performed to assess inter-method reliability of the volumetric measurement. All statistical analyses were conducted using Python packages (statsmodels 0.14.0 and SciPy 1.11.4) and R packages (lme4 and dplyr) with P < 0.05 set as the threshold for statistical significance.

Go to:

RESULTS

Patients

A total of 150 patients were included in the final analysis in the current study: mean age, 57.3 ± 16.2 years; 51 (34.0%) male (Fig. 1). Each group of 50 participants was scanned using one of three 3T scanners from three different vendors. The clinical characteristics of the patients are summarized in Table 1.

Fig. 1

Study participants. MRI = magnetic resonance imaging, Accel-DL = accelerated deep learning-based reconstruction

* Click for larger image

* Download as PowerPoint slide

Table 1

Characteristics of the study participants for each scanner

* Click for larger image

* Click for full table

* Download as Excel file

MR Acquisition Time

For all vendors, the mean scan time reduced from 175.8 ± 50.3 to 106.6 s ± 30.5 s (mean = 39.4%; median = 41.0%; range = 24.2%–51.3%): GE, 174.0 ± 54.2 to 109.2 s ± 34.6 s (37.2%); Philips, 175.2 ± 25.4 to 107.5 s ± 18.5 s (38.7%); and Siemens, 178.2 ± 75.3 to 103.0 s ± 42.9 s (42.2%) according to the alphabetical order. When comparing SE and GRE sequences, the mean scan times had reduced from 153.33 ± 27.04 to 96.89 s ± 24.35 s (37.23%) and 243.33 ± 43.13 to 135.67 s ± 32.33 s (44.58%), respectively (P = 0.23) (Table 2). Bar plots comparing the scan times across different vendors are provided in Supplementary Figure 2.

Table 2

Magnetic resonance acquisition time

* Click for larger image

* Click for full table

* Download as Excel file

Qualitative Evaluation

Accel-DL showed significant improvements over conventional MRI in overall image quality (3.78 ± 0.71 vs. 3.36 ± 0.61, P < 0.001), structure delineation (2.47 ± 0.61 vs. 2.35 ± 0.62, P < 0.001), and artifacts (3.73 ± 0.72 vs. 3.71 ± 0.69, P = 0.016). However, no significant differences (Accel-DL vs. conventional) were observed in lesion conspicuity (2.09 ± 0.79 vs. 2.01 ± 0.80, P = 0.06), enlarged PVS grading (1.23 ± 0.42 vs. 1.24 ± 0.43, P = 0.44), and the Fazekas scale (1.08 ± 0.65 vs. 1.04 ± 0.66, P = 0.12). To evaluate lesion conspicuity, WMH was observed in 129 out of 150 patients (86.0%), showing no difference among vendors (Table 1). The details of each item are presented in Table 3, and more comprehensive data are available in Supplementary Table 2 and Supplementary Figure 3. A comparison of conventional MRI and Accel-DL for both 2D axial SE and 3D GRE sequences in representative cases is shown in Figure 2. Approximately 2% (3 out of 150) of the cases exhibited issues with enhanced artifacts, potentially mimicking pontine lacunar infarctions (Fig. 2C). When combining all sequences, a fair to substantial agreement was observed among the readers for overall image quality (κ = 0.34, fair), structure delineation (κ = 0.50, moderate), lesion delineation (κ = 0.63, substantial), and artifacts (κ = 0.35, fair).

Fig. 2

Representative comparison of conventional and DL-based reconstruction in brain MRI. A: On 3D T1-weighted images, Accel-DL showed improved structural delineation of the right globus pallidus and corticomedullary differentiations (arrowheads), along with a reduction in fine-granular noise. However, they also display more pronounced subtle herringbone pattern artifacts across the image when compared to conventional MRI (arrows). B: On T2-FLAIR images, the lesion conspicuity of the enlarged perivascular space in the right basal ganglia is enhanced, with more distinct corticomedullary differentiation in Accel-DL compared to conventional MRI (arrowheads). C: On T2 FLAIR, an exaggerated pulsation artifact may resemble a chronic lacunar infarction in the right pons (arrowheads). D: Both Accel-DL and conventional images were assigned a Fazekas scale rating of 3, yet the delineation of white matter hyperintensity margins appears sharper with Accel-DL (arrowheads). E: T2-weighted images processed with Accel-DL demonstrate more clearly defined enlarged perivascular spaces (arrowheads) compared with conventional images, as well as vascular structures within the choroid plexus (arrows). DL = deep learning, MRI = magnetic resonance imaging, Accel-DL = accelerated brain MRI with DL-based reconstruction, FLAIR = fluid-attenuated inversion recovery, GRE = gradient-recalled echo, SE = spin-echo

* Click for larger image

* Download as PowerPoint slide

Table 3

Qualitative image comparison between conventional and DL-based accelerated MRI across different MRI sequences

* Click for larger image

* Click for full table

* Download as Excel file

In the LMM analysis, vendor-reconstruction interaction resulted in better model fit than that of the model without interaction (P = 0.02). Accel-DL significantly improved overall image quality and delineation on 3D T1WI (β = 0.82, P < 0.001; β = 0.52, P < 0.001) and T2-FLAIR (β = 0.52, P < 0.001; β = 0.24, P < 0.001), with notable artifact reduction on 3D T1WI (β = 0.26, P < 0.001). Considering vendor-reconstruction interaction in vendor comparison, GE showed greater benefits than Siemens and Philips in overall image quality on T2-FLAIR (β = −0.27, P < 0.001; β = −0.37, P < 0.001) and better delineation on T1WI and 3D T1WI than Philips (β = −0.20; P < 0.001; β = −0.19; P = 0.001) when applying Accel-DL (Supplementary Fig. 4). Detailed results are provided in Supplementary Tables 3, 4.

Quantitative Evaluation

Accel-DL showed enhanced SNR and CNR in axial T2-FLAIR imaging, achieving an SNR of 82.0 ± 23.1 compared to 31.4 ± 10.8 with conventional MRI (P = 0.02); however, in 3D T1WI, Accel-DL and conventional MRI showed comparable SNR values of 22.3 ± 13.2 and 18.3 ± 9.2, respectively (P = 0.76).

In the volumetric analysis, no significant differences were observed in 1253 out of 1276 (98.2%) regional volumes (r = 0.92 ± 0.10, range = 0.38–1.00), except for the entorhinal cortex, fusiform gyrus, frontal and temporal poles, and thalamus (i.e., the deep GM including the thalamus) using NeuroQuant. In the LesionQuant analysis, five of the six (83.3%) lesion categories (i.e., deep white lesion count, total lesion count, leukocortical lesion count, periventricular lesion count, infratentorial lesion count, and total lesion burden) showed no significant differences, except for leukocortical lesion counts (r = 0.64 ± 0.29, range = 0.12–0.92). For NeuroQuant and LesionQuant, comparisons between conventional MRI and Accel-DL for 3D T1WI in representative cases are shown in Figure 3A and 3B, respectively. For hippocampal, lateral ventricular, total GM (i.e., cortical and subcortical), and cerebral WM volumes, the comparison between conventional MRI and Accel-DL using Bland-Altman plots revealed no significant differences in the volumetric measurement of all relevant regions, indicating high inter-method reliability (Fig. 4).

Fig. 3

Comparison of conventional MRI (upper row) and Accel-DL (lower row): (A) Regional volumetric segmentation on 3D T1-weighted gradient echo images (T1WI) using NeuroQuant; (B, C) WM hyperintensity segmentation using LesionQuant, overlaid on (B) T2 fluid-attenuated inversion recovery and (C) 3D T1WI. In (B), the segmentation mask (blue, arrowheads) from LesionQuant in Accel-DL more accurately aligns with the underlying WM hyperintensity compared with conventional reconstruction, with the lesion notably shifting from a deep (green) to a periventricular (blue) location (arrows). In (C), on 3D T1WI overlaid with gray matter segmentation, noise is reduced in Accel-DL, particularly in the right cerebral WM, compared with that in conventional reconstruction, resulting in clearer margins for the overlaid WM hyperintensity segmentation masks (purple, arrowheads). MRI = magnetic resonance imaging, Accel-DL = accelerated brain MRI with deep learning-based reconstruction, T1WI = T1-weighted image, WM = white matter

* Click for larger image

* Download as PowerPoint slide

Fig. 4

Bland-Altman plot for comparison of volumetric results (A–F) using 3D T1-weighted images (NeuroQuant) for hippocampal, lateral ventricular, total gray matter (i.e. cortical and subcortical), and cerebral white matter volumes: plots show the mean of “between-reconstruction” volumes and the difference between these volumes, where the sky-blue line indicates the mean and red lines represent ±2 SDs. Blue error bars represent the 95% confidence intervals for both upper and lower limits of agreement. The mean is close to the zero line, and the majority of the differences fall within red lines. SD = standard deviation

* Click for larger image

* Download as PowerPoint slide

Go to:

DISCUSSION

In this multi-vendor, multi-sequence, multi-reader, prospective study, Accel-DL improved image quality, reduced artifacts, and enhanced structure delineation. This increased the SNR while maintaining lesion conspicuity for up to half of the conventional acquisition times. The enlarged PVS grades and Fazekas scales were comparable. Brain parcellation and lesion segmentation results were similar for both conventional MRI and Accel-DL, suggesting that Accel-DL can enhance diagnostic imaging and reduce costs. Previous studies focused on 2D [4, 9] or 3D models [5, 21]; however, this is the first study to evaluate a model that applies both on a large sample (n = 150).

In a qualitative analysis involving 259200 assessments by four readers across four sequences and 54 evaluation items for 150 participants, Accel-DL significantly enhanced MRI image quality. It improved the overall quality and structural delineation in axial T2WI, T2-FLAIR, and 3D T1WI, but less so in axial T1WI. Artifacts reduced in 3D T1WI and T2-FLAIR images. Our study cohort included neurology clinic outpatients, with WMH observed in 129 of the 150 patients (86.0%). On T2-FLAIR, lesion conspicuity and the Fazekas test showed no significant differences between DLR and conventional imaging. On T2WI, the enlarged PVS grading showed comparable results. Interobserver agreement ranged from fair (κ = 0.34–0.35 for image quality and artifacts) to substantial (κ = 0.50–0.63 for structure delineation and lesion conspicuity), indicating generally acceptable results [4, 21].

Vendor interaction with the reconstruction method (conventional vs. Accel-DL) significantly affected all sequences (P < 0.001). This indicates that the reconstruction effects on image quality vary by vendor, complicating generalization and possibly requiring vendor-specific optimization. In summary, GE, as the legacy MR scanner, demonstrated greater benefits than Siemens and Philips in terms of overall image quality on T2-FLAIR and delineation on T1WI, T2-FLAIR, and 3D T1WI with Accel-DL (Supplementary Fig. 4). However, Philips and Siemens outperformed GE in terms of overall image quality, artifact reduction, and delineation of various T1- and T2-weighted sequences using Accel-DL. Accel-DL can upgrade image quality in outdated MR systems, offering an economical path for resource-constrained settings; further studies are needed to confirm these findings. However, the application of Accel-DL under certain conditions, depending on the equipment, can result in varying performance enhancements across different evaluation items. The scan-time reduction averaged 41% across various vendors, reaching up to 51%, nearly doubling efficiency. This was especially pronounced in 3D T1WI GRE sequences compared to that in 2D SE sequences (44.6% vs. 37.2%, P = 0.23). Accelerated 3D MRI with Accel-DL’s denoising has significant implications for neurodegenerative disease research, particularly Alzheimer’s disease. Volumetric analyses, crucial for classifying atrophy patterns and predicting disease progression, are also beneficial. For instance, hippocampal volume, a key indicator of cognitive decline [22], can be rapidly assessed to facilitate the inclusion of advanced MR sequences in routine clinical practice for neurodegenerative diseases.

Accel-DL increased the SNR and CNR in axial T2-FLAIR images, with no notable differences between 3D T1WI and conventional MRI. A high SNR via denoising raises concerns about potential blurring and detail loss in fine structures. Data loss in k-space-based imaging, a known issue [23], has been previously addressed using compressed sensing techniques. Despite these concerns, a recent study [9] has shown that Accel-DL enhances the visualization of fine structures such as the pituitary gland. Our present study also observed improved delineation of structures, such as corticomedullary differentiation, indicating the effectiveness of Accel-DL in detailed imaging.

However, DLR images, including those in our study, are often considered to be smoothed [24]. This effect is partly due to the squared-term pixel-wise loss, such as L2 loss, used in U-Net models for training stability. Generative adversarial networks generate sharper images than variational autoencoder using squared-term loss [25]. Moreover, DLR algorithms tend to produce “over-smoothed” images when attempting greater acceleration, similar to the effect seen with iterative reconstruction (IR) in CT when reducing a great radiation dose [26, 27]. Despite this, IR is widely accepted in clinical practice, as it alters the interpretation of images by radiologists. Future improvements can be achieved by engineering and incorporating various loss functions.

Recent studies have highlighted that while DLR reduces Gaussian noise in MR images, it also makes image artifacts more prominent [13]. In our study, enhanced artifacts mimicking pontine lacunar infarctions (Fig. 2C) were observed in approximately 2% (3 out of 150) of the cases (the others are provided in Supplementary Fig. 5). When reviewed or compared with conventional images, these cases were identified as artifacts rather than true lesions. These artifacts are enhanced rather than newly introduced by DLR but might be mistaken for pontine infarction without conventional images as a reference. Previously known artifacts such as “sulcal hyperintensities” [28] highlight the need for improvements in DLR for clinical practice.

In brain volumetry using NeuroQuant, only 23 of the 1276 regional volumes (1.8%) differed between conventional MRI and Accel-DL, showing a high correlation (r = 0.92 ± 0.10) and suggesting possible interchangeability for follow-up MRI (Fig. 4). Discrepancies mainly occurred in the temporal lobe, including the entorhinal cortex, fusiform gyrus, temporal pole, and thalamus. These variations may be due to motion artifacts or field inhomogeneities. Among the lesion counts, only leukocortical lesions showed a significant difference (P < 0.001), likely because of clearer GM-WM differentiation. Although beneficial, these marginal improvements in SNR and structure delineation by Accel-DL may complicate the interchangeability with conventional MRI in some cases, underscoring the need for more research, especially on multiple sclerosis, to ascertain the most accurate method.

The present study has several limitations. First, this prospective evaluation did not include patients with large lesions or structural distortions, such as acute hemorrhage, infarction, or tumors, highlighting the need for further research. Future studies are needed to investigate whether the diagnostic performance improves in pathological conditions, particularly in patients with acute stroke who require rapid imaging. Second, 3D T1WI and T2-FLAIR images were not acquired according to the scan parameters recommended by NeuroQuant and LesionQuant, which may have led to inaccurate results. However, the primary objective of this study was to evaluate the impact of Accel-DL on improving the quality of brain MRI and reducing the scan time compared to that of conventional MRI. As this limitation applies equally to both Accel-DL and conventional MRI, it does not have a significant impact on the purpose of the study. Finally, although this was a multi-reader study with more than 250000 assessment items, the inter-reader agreement for overall image quality and artifacts was fair; however, similar recent studies [4, 21] have shown that such agreements are not uncommon.

In conclusion, Accel-DL substantially reduced the scan time and significantly improved the quality of brain MRI in both SE and GRE sequences without compromising volumetry, including lesion quantification.

Go to:

Supplement

The Supplement is available with this article at https://doi.org/10.3348/kjr.2024.0653.

Click here to view.(3M, pdf)

References

1. 1. Finger CE, Moreno-Gonzalez I, Gutierrez A, Moruno-Manchon JF, McCullough LD. Age-related immune alterations and cerebrovascular inflammation. Mol Psychiatry 2022;27:803–818. * PubMed * CrossRef 2. 1. OECD. Health at a glance 2021: OECD indicators. Paris: OECD Publishing; 2021. 3. 1. NHS England. Monthly diagnostics data 2022–23. [accessed on December 2, 2023]. Available at: https://www.england.nhs.uk/statistics/statistical-work-areas/diagnostics-waiting-times-and-activity/monthly-diagnostics-waiting-times-and-activity/monthly-diagnostics-data-2022- 23. 4. 1. Kim SH, Choi YH, Lee JS, Lee SB, Cho YJ, Lee SH, et al. Deep learning reconstruction in pediatric brain MRI: comparison of image quality with conventional T2-weighted MRI. Neuroradiology 2023;65:207–214. * PubMed * CrossRef 5. 1. Bash S, Wang L, Airriess C, Zaharchuk G, Gong E, Shankaranarayanan A, et al. Deep learning enables 60% accelerated volumetric brain MRI while preserving quantitative performance: a prospective, multicenter, multireader trial. AJNR Am J Neuroradiol 2021;42:2130–2137. * PubMed * CrossRef 6. 1. Johnson PM, Lin DJ, Zbontar J, Zitnick CL, Sriram A, Muckley M, et al. Deep learning reconstruction enables prospectively accelerated clinical knee MRI. Radiology 2023;307:e220425 * PubMed * CrossRef 7. 1. Yoo H, Yoo RE, Choi SH, Hwang I, Lee JY, Seo JY, et al. Deep learning-based reconstruction for acceleration of lumbar spine MRI: a prospective comparison with standard MRI. Eur Radiol 2023;33:8656–8668. * PubMed * CrossRef 8. 1. Lai YC, Ray KM, Lee AY, Hayward JH, Freimanis RI, Lobach IV, et al. Microcalcifications detected at screening mammography: synthetic mammography and digital breast tomosynthesis versus digital mammography. Radiology 2018;289:630–638. * PubMed * CrossRef 9. 1. Kim M, Kim HS, Kim HJ, Park JE, Park SY, Kim YH, et al. Thin-slice pituitary MRI with deep learning-based reconstruction: diagnostic performance in a postoperative setting. Radiology 2021;298:114–122. * PubMed * CrossRef 10. 1. Wang S, Su Z, Ying L, Peng X, Zhu S, Liang F, et al. Accelerating magnetic resonance imaging via deep learning. [accessed on December 2, 2023]. Available at: https://doi.org/10.1109/ISBI.2016.7493320. 11. 1. Lee D, Yoo J, Ye JC. Deep artifact learning for compressed sensing and parallel MRI. arXiv [Preprint]; 2017 [accessed on October 22, 2024]. Available at: https://doi.org/10.48550/arXiv.1703.01120. 12. 1. Chen Y, Xie Y, Zhou Z, Shi F, Christodoulou AG, Li D. Brain MRI super resolution using 3D deep densely connected neural networks. [accessed on December 2, 2023]. Available at: https://doi.org/10.1109/ISBI.2018.8363679. 13. 1. Kiryu S, Akai H, Yasaka K, Tajima T, Kunimatsu A, Yoshioka N, et al. Clinical impact of deep learning reconstruction in MRI. Radiographics 2023;43:e220133 * PubMed * CrossRef 14. 1. Jeong G, Kim H, Yang J, Jang K, Kim J. All-in-one deep learning framework for MR image reconstruction. arXiv [Preprint]; 2024 [accessed on October 22, 2024]. Available at: https://doi.org/10.48550/arXiv.2405.03684. 15. 1. Park SH, Han K, Jang HY, Park JE, Lee JG, Kim DW, et al. Methods for clinical evaluation of artificial intelligence algorithms for medical diagnosis. Radiology 2023;306:20–31. * PubMed * CrossRef 16. 1. Vranic JE, Cross NM, Wang Y, Hippe DS, de Weerdt E, Mossa-Basha M. Compressed sensing-sensitivity encoding (CS-SENSE) accelerated brain imaging: reduced scan time without reduced image quality. AJNR Am J Neuroradiol 2019;40:92–98. * PubMed * CrossRef 17. 1. Fazekas F, Chawluk JB, Alavi A, Hurtig HI, Zimmerman RA. MR signal abnormalities at 1.5 T in Alzheimer’s dementia and normal aging. AJR Am J Roentgenol 1987;149:351–356. * PubMed * CrossRef 18. 1. Ohlmann-Knafo S, Morlo M, Tarnoki DL, Tarnoki AD, Grabowski B, Kaspar M, et al. Comparison of image quality characteristics on silent MR versus conventional MR imaging of brain lesions at 3 tesla. Br J Radiol 2016;89:20150801 * PubMed * CrossRef 19. 1. Brewer JB. Fully-automated volumetric MRI with normative ranges: translation to clinical practice. Behav Neurol 2009;21:21–28. * PubMed * CrossRef 20. 1. Han K, Ryu L. Statistical methods for the analysis of inter-reader agreement among three or more readers. Korean J Radiol 2024;25:325–327. * PubMed * CrossRef 21. 1. Rudie JD, Gleason T, Barkovich MJ, Wilson DM, Shankaranarayanan A, Zhang T, et al. Clinical assessment of deep learning-based super-resolution for 3D volumetric brain MRI. Radiol Artif Intell 2022;4:e210059 * PubMed * CrossRef 22. 1. Apostolova LG, Dutton RA, Dinov ID, Hayashi KM, Toga AW, Cummings JL, et al. Conversion of mild cognitive impairment to Alzheimer disease predicted by hippocampal atrophy maps. Arch Neurol 2006;63:693–699. * PubMed * CrossRef 23. 1. Jaspan ON, Fleysher R, Lipton ML. Compressed sensing MRI: a review of the clinical literature. Br J Radiol 2015;88:20150487 * PubMed * CrossRef 24. 1. Lin DJ, Johnson PM, Knoll F, Lui YW. Artificial intelligence for MR image reconstruction: an overview for clinicians. J Magn Reson Imaging 2021;53:1015–1028. * PubMed * CrossRef 25. 1. Larsen ABL, Sønderby SK, Larochelle H, Winther O. Autoencoding beyond pixels using a learned similarity metric. [accessed on December 2, 2023]. Available at: https://proceedings.mlr.press/v48/larsen16. 26. 1. Pourjabbar S, Singh S, Singh AK, Johnston RP, Shenoy-Bhangle AS, Do S, et al. Preliminary results: prospective clinical study to assess image-based iterative reconstruction for abdominal computed tomography acquired at 2 radiation dose levels. J Comput Assist Tomogr 2014;38:117–122. * PubMed * CrossRef 27. 1. Kalra MK, Woisetschläger M, Dahlström N, Singh S, Digumarthy S, Do S, et al. Sinogram-affirmed iterative reconstruction of low-dose chest CT: effect on image quality and radiation dose. AJR Am J Roentgenol 2013;201:W235–W244. * PubMed * CrossRef 28. 1. Tanenbaum LN, Tsiouris AJ, Johnson AN, Naidich TP, DeLano MC, Melhem ER, et al. Synthetic MRI for clinical neuroimaging: results of the magnetic resonance image compilation (MAGiC) prospective, multicenter, multireader trial. AJNR Am J Neuroradiol 2017;38:1103–1110. * PubMed * CrossRef