Deep learning reconstruction for accelerating synthetic magnetic resonance imaging of the breast: a comparative analysis of accelerated and standard protocols

Introduction

Breast cancer is one of the most common malignant tumors among women worldwide, and its incidence has been increasing year by year, posing a significant challenge to global health (1,2). Triple-negative breast cancer (TNBC), the most aggressive and biologically heterogeneous subtype of breast cancer, has attracted particular attention (3). Its early detection and accurate assessment are crucial for improved prognosis. Magnetic resonance imaging (MRI) is widely employed for the screening, diagnosis, and postoperative evaluation of breast cancer (4,5). Clinical breast imaging typically requires scanning multiple sequences, which can prolong the data acquisition times. This not only increases the risk of patient discomfort but also makes the occurrence of motion artifacts more likely, compromising the image quality and reducing the efficiency of clinical workflows (6). Therefore, techniques that can both shorten breast MRI scanning times and maintain image quality are urgently needed (7).

Synthetic MRI (SyMRI) is an emerging MRI technology that synthesizes multiple contrast-weighted images using quantitative relaxometry parameters measured from a single acquisition. This allows for the simultaneous generation of quantitative maps—such as T1, T2, and proton density (PD)—and contrast-weighted imaging, including T1-weighted imaging (T1WI), T2-weighted imaging (T2WI), and PD-weighted imaging (PDWI) (8,9). Quantitative parameters may provide important information regarding the microstructural and biological characteristics of tissues (10). Compared with conventional MRI, SyMRI offers improved quantitative mapping and contrast-weighted image quality, facilitating more efficient clinical diagnosis (8,11,12). Owing to its versatility, this technique has demonstrated considerable potential for applications across various fields of medical imaging (8,13,14). Quantitative maps generated by SyMRI have demonstrated the ability to accurately differentiate between benign and malignant breast lesions and have exhibited strong diagnostic performance in breast cancer subtype classification and prognosis prediction (15-17). Notably, the image quality of synthesized T1WI and T2WI has been reported to be comparable to that of conventionally acquired images, supporting their use in diagnostic settings (11). Despite these advantages, one major limitation of SyMRI is its relatively long scanning time, which may hinder its widespread clinical application. Although a higher in-plane parallel imaging factor can be used to accelerate SyMRI, it often reduces the signal-to-noise ratio (SNR) and may compromise image detail, potentially affecting diagnostic accuracy (18-21).

Recently, a novel deep learning reconstruction (DLR) method employing a deep convolutional neural network (CNN) has been introduced (22). DLR reconstructs images from fully sampled k-space data by leveraging its powerful data-processing capabilities (23-26) and can reduce image noise and enhance sharpness (27-32). However, its application in accelerating SyMRI for breast imaging has not been fully evaluated. Previous studies have indicated that the T2 signal from synthetic imaging, along with the presence of cystic degeneration or necrosis within breast tumors, may serve as potential imaging biomarkers for distinguishing TNBC from non-TNBC prior to surgery (3). Building upon this foundation, we sought to determine whether such diagnostic differentiation can be preserved under accelerated imaging conditions and with the aid of DLR. Specifically, we assessed the ability of DLR to accelerate SyMRI by comparing the quantitative parameters, SNR, and subjective image quality between DLR accelerated SyMRI and standard SyMRI. We present this article in accordance with the STARD reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2024-2972/rc).

Methods

Participants

This prospective single-center study was conducted as per the principles outlined in the Declaration of Helsinki and its subsequent amendments, and was approved by the Institutional Review Board of West China Hospital, Sichuan University (IRB 2023-02-15 No. 187). Informed consent was obtained from all patients. For this study, female patients who had undergone breast MRI from February to August 2024 were recruited. The patients were selected if they were (I) 18 years or above and (II) had mass-type breast lesions scheduled for a pathological biopsy. The exclusion criteria were as follows: (I) previous administration of antitumor radiotherapy, chemotherapy, or surgical treatment prior to MRI; (II) a history of other tumors; (III) breast implants, pregnancy, or breastfeeding; and (IV) poor image quality, defined as the presence of severe motion artifacts in any SyMRI or dynamic contrast-enhanced (DCE) MRI sequences that compromised diagnostic interpretation or regions of interest (ROIs) placement. Finally, 58 female patients with clinically suspected breast cancer were included in the study.

Imaging protocol

MRI examinations were performed on a 3T MRI scanner (SIGNA Premier, GE Healthcare, Milwaukee, WI, USA) with a dedicated eight-channel HD flat GEM table breast array coil (GE Healthcare). Each participant was positioned in a prone position, which ensured that both breasts naturally hung in the center of the coil. Routine clinical MRI sequences for breast imaging included axial fat-saturated T2WI, T1WI, and DCE MRI. Prior to the DCE MRI, GE’s SyMRI sequences were scanned under magnetic resonance image compilation (MAGiC) through use of multidynamic multiecho (MDME) acquisition (33). Two sets of SyMRI scans were implemented: a manufacture-recommended standard protocol [2× SyMRI: repetition time (TR) =4,277 ms, echo time (TE) =21.6 ms, effective echo time 2 =108.3, flip angle =90°, slice thickness =4 mm, spacing =1.0 mm, field of view (FOV) =320×288 mm², matrix =320×256, bandwidth =25, slices =20, echo train length =16, in-plane acceleration factor =2, and scan time =4 min and 51 s] (11) and an accelerated protocol (3× SyMRI). The accelerated 3× SyMRI protocol employed the same parameters as those of the standard 2× SyMRI protocol, except it involved a parallel imaging acceleration factor of 3 and a reduced scan time of 3 min and 2 s.

In addition to the conventional reconstruction, the k-space data for the 3× SyMRI were reconstructed with the vendor-provided DLR algorithm (AIR Recon DL, GE Healthcare, Milwaukee, WI, USA) to generate images from 3× SyMRI with DLR (SyMRI-DLR). This DLR algorithm employs a deep CNN that directly operates on the raw k-space data, producing images with a higher SNR, reduced truncation artifacts, and enhanced spatial resolution. The DLR algorithm provides three levels of noise reduction (low, medium, and high) through different intensities of CNN application. In this study, a high level of denoising was selected to ensure optimal image quality under the accelerated acquisition protocol (22).

Data analysis and processing

SyMRI software (SyMRI; version 8.0.4, SyntheticMR AB, Linköping, Sweden) was used to automatically generate synthetic T1WI, T2WI, PDWI, and T2 short-time inversion recovery (STIR) for 2× SyMRI, 3× SyMRI, and 3× SyMRI-DLR, along with the corresponding quantitative tissue maps (T1, T2, and PD maps) (Figure 1). We employed the default settings for TR, TE, and TI to create the synthetic contrast-weighted images. Additionally, dcm2niix software (version v1.0.20230411) (RRID: SCR_017672) was used to convert the Digital Imaging and Communications in Medicine (DICOM) data into the Neuroimaging Informatics Technology Initiative (NIfTI) format.

Figure 1 Quantitative tissue maps (T1, T2, and PD maps) for 2× SyMRI, 3× SyMRI, and 3× SyMRI-DLR. T1, T2, and PD represent longitudinal relaxation time, transverse relaxation time, and proton density, respectively. DLR, deep learning reconstruction; PD, proton density; SyMRI, synthetic magnetic resonance imaging.

Measurement of the quantitative parameters of tissue

Data analysis was conducted via ITK-SNAP software version 3.8.0 (RRID: SCR_002010). Two radiologists, one with 8 years of experience (F.Y.) and the other with 5 years of experience (Y.J.), delineated the ROIs for the breast gland and lesion areas on the quantitative tissue maps (T1, T2, and PD maps) generated from the three groups of SyMRI data. For the breast gland area, the maximum glandular layer on the normal side of the patient’s breast was selected, and three circular ROIs were uniformly placed. In the breast lesion area, contours were drawn based on the location and morphology of the lesion as seen in the DCE-MRI images, with the maximum diameter of the lesion being outlined. The position and area of the ROIs were kept consistent across different sequences for the same patient during each measurement. The software automatically generated the mean T1, T2, and PD map values for these regions.

SNR measurement

The SNRs for the synthetic T1- and T2-weighted images generated by the three SyMRI groups were measured and compared. The contour of the lesion was delineated at the plane of the maximum diameter to obtain the average signal intensity (SI) of the lesion tissue. Concurrently, three circular ROIs were placed in an area with uniform background at the same plane to obtain the average background noise SI and its standard deviation (SD). The SNR of the lesion tissue was calculated with the following formula:

SNR=SIlesion/SDbackground[1]

Qualitative assessment

A systematic evaluation was conducted by the aforementioned two radiologists on the synthetic T1- and T2-weighted images obtained from the three SyMRI groups. Images of three SyMRI protocols were reviewed in separate sessions, with a 1-week washout period between each session being used to reduce recall bias. During each session, the cases were presented in a randomized order. The T1- and T2-weighted images were evaluated independently rather than as fused sets. The assessment criteria included the overall image quality, anatomical clarity, diagnostic information, tissue contrast, image uniformity, and artifact status. A 5-point Likert scale was used for scoring (Table S1) under the following scheme: 1= poor, 2= fair, 3= average, 4= good, and 5= excellent (11,21). Both radiologists were blinded to patient identity, acquisition protocol, and reconstruction method.

Diagnostic value of quantitative parameters from SyMRI

Patients with breast cancer were categorized into a TNBC group and non-TNBC group based on immunohistochemical results (34,35). We analyzed the quantitative parameters (T1, T2, and PD values) obtained from the three SyMRI groups to discern the differences between the breast glandular tissue and breast cancer lesions and those between TNBC and non-TNBC.

Statistical analysis

Statistical analysis was performed with SPSS version 27.0 (IBM Corp., Armonk, NY, USA; RRID: SCR_019096) and MedCalc version 23.0.2 (MedCalc Software, Ostend, Belgium; RRID: SCR_015044) software. Quantitative data following a normal distribution after normality testing are expressed as the mean ± SD, while nonnormally distributed data are expressed as the median and interquartile range. The qualitative data are expressed as the frequency and percentage. For datasets that met the normality assumption, one-way analysis of variance (ANOVA) was used to compare differences between the three protocols. When the data did not meet the normality assumption, the Friedman test was applied as a nonparametric alternative. If the global test showed statistical significance, post hoc pairwise comparisons were conducted via Bonferroni-corrected t-tests (for ANOVA) or Wilcoxon signed-rank tests with Bonferroni correction (for Friedman test) to identify specific group differences. Spearman correlation coefficients (r) were calculated to assess the correlation of quantitative data between the different SyMRI group. Intraclass correlation coefficients (ICCs) were used for consistency testing. The Bland-Altman method was employed to evaluate the consistency of the quantitative values for breast glandular tissue and lesions across different sequences. The differences in T1, T2, and PD quantitative parameter values between breast glandular tissue and breast cancer lesions, as well as those between the TNBC and non-TNBC groups, were evaluated with t-tests or Mann-Whitney tests. Receiver operating characteristic (ROC) curves were employed to assess the classification ability of T1, T2, and PD values obtained from the three SyMRI groups. The area under the curve (AUC) was employed to quantify the performance. A P value <0.05 was considered statistically significant.

Results

Baseline characteristics

A total of 58 female patients with clinically suspected breast cancer, with an average age of 50.38±1.34 years, were included in this study. The majority of patients exhibited sparse breast tissue (79.3%) and heterogeneous breast tissue (15.5%). Pathological biopsies following MRI examinations confirmed that all 58 patients were diagnosed with breast cancer. The pathological data of one of the patients were incomplete. Among these patients, three had carcinoma in situ and the remaining had invasive carcinoma. In addition, 10 (17.2%) patients were diagnosed with TNBC, while 47 (81.0%) were classified as non-TNBC.

Impact of DLR on the quantitative tissue values of SyMRI

The acquisition time for both the 3× SyMRI and 3× SyMRI-DLR protocols was 37.46% shorter than that for the 2× SyMRI. The analysis of the quantitative parameters T1, T2, and PD values obtained from the three SyMRI groups revealed no significant differences in the T1, T2, or PD values of breast lesions across the different SyMRI data (all P values >0.05) (Table 1, Figure 2A). The Bland-Altman plots demonstrated that the mean difference lines for the quantitative values measured from the 3× SyMRI and 2× SyMRI images were close to the zero line, with the 95% confidence intervals (CIs) for the differences being zero and within the limits of agreement, indicating minimal systematic error. Furthermore, 3× SyMRI-DLR did not amplify this bias (Figure 2B). As shown in Table 2 and Figure 3, the Spearman correlation analysis showed excellent correlation in the quantitative T1, T2, and PD values of both breast lesions and glands between the 2× SyMRI and 3× SyMRI and between the 2× SyMRI and 3× SyMRI-DLR protocols (r_s>0.89; P<0.001). The correlation was optimal between 3× SyMRI and 3× SyMRI-DLR (r_s>0.99; P<0.001). The quantitative accuracy of the T1, T2, and PD and tissue parameters was not significantly reduced due to the acceleration.

Figure 2 Box plot (A) and Bland-Altman plot (B) results. Box plots show the T1, PD, and T2 values of the gland and lesion. Bland-Altman plots show the percentage mean difference (solid blue line) and ±1.96 SDs from the mean difference (dotted red line). The solid yellow line shows the regression line of the differences. T1, T2, and PD represent longitudinal relaxation time, transverse relaxation time, and proton density, respectively. DLR, deep learning reconstruction; PD, proton density; SD, standard deviation; SyMRI, synthetic magnetic resonance imaging.

Figure 3 Scatter plots showing the correlation of breast lesion T1, T2, and PD values between 3× SyMRI and 2× SyMRI, between 3× SyMRI-DLR and 2× SyMRI, and between 3× SyMRI-DLR and 3× SyMRI. Spearman rank correlation coefficient r (two-tailed) and P values are also displayed. P values <0.05 were considered statistically significant. T1, T2, and PD represent longitudinal relaxation time, transverse relaxation time, and proton density, respectively. DLR, deep learning reconstruction; PD, proton density; SyMRI, synthetic magnetic resonance imaging.

SNR measurement results

Significant differences were observed in the SNR values among the three SyMRI protocols for both T1WI and T2WI sequences (T1WI: P<0.001; T2WI: P<0.001). Pairwise comparisons were subsequently performed with Bonferroni correction. For the T1WI sequence, the SNR of 3× SyMRI-DLR was significantly higher than that of 3× SyMRI (P<0.001) but not significantly different from that of standard 2× SyMRI (P=0.343). In the T2WI sequence, 3× SyMRI-DLR demonstrated a significantly higher SNR compared to both 3× SyMRI (P<0.001) and 2× SyMRI (P<0.001). The detailed SNR results are presented in Table 3.

Impact of DLR on the qualitative indicators of SyMRI

Subjective evaluation of the synthesized T1WI and T2WI resulted in statistically significant differences among the three SyMRI groups in terms of anatomical clarity (T1WI: P<0.001; T2WI: P<0.001), tissue contrast (T1WI: P<0.001; T2WI: P=0.009), image uniformity (T1WI: P<0.001; T2WI: P<0.001), and artifact status (T1WI: P<0.001; T2WI: P<0.001) (Table 4 and Figure 4). In contrast, the differences in the overall image quality and diagnostic information regarding lesions were not statistically significant (P>0.05). The synthesized T1- and T2-weighted images from the 3× SyMRI-DLR protocol demonstrated clearer anatomical structures and better image uniformity as compared to those from 3× SyMRI. However, both 3× SyMRI and 3× SyMRI-DLR produced more artifacts as compared to 2× SyMRI, with the most common artifact being localized structural duplication at the margins of glandular tissue.

Figure 4 Axial contrast-weighted images of a 45-year-old woman who underwent breast MR imaging. With the application of DLR, the anatomical clarity from 3× SyMRI-DLR significantly improved the tissue contrast and image uniformity of tumor (big arrow) and mammary gland (small arrow) as compared to 3× SyMRI. The image quality of 3× SyMRI-DLR was comparable to that of 2× SyMRI. DLR, deep learning reconstruction; MR, magnetic resonance; SyMRI, synthetic magnetic resonance imaging; T1WI, T1-weighted imaging; T2-STIR, T2-weighted short-time inversion recovery; T2WI, T2-weighted imaging.

Diagnostic value of quantitative parameters from SyMRI

A comparison of the average T1, T2, and PD values measured in both breast tissue and lesion areas between the three SyMRI groups was conducted (Table 1). Among the quantitative parameters obtained from 2× SyMRI, the T1 and T2 values of the breast cancer lesion areas were significantly higher than those of the normal breast tissue (P<0.05), with AUC values of 0.857 (95% CI: 0.779–0.915) and 0.632 (95% CI: 0.537–0.720), respectively. Similarly, among the quantitative parameters measured from 3× SyMRI, the T1 and T2 values of the breast cancer lesion areas were significantly higher than those of the normal breast tissue (P<0.05), with AUC values of 0.858 (95% CI: 0.781–0.916) and 0.627 (95% CI: 0.532–0.715), respectively. For the quantitative parameters obtained from 3× SyMRI-DLR, the T1 and T2 values of the breast cancer lesion areas were significantly higher than those of the normal breast tissue (P<0.05), with AUC values of 0.857 (95% CI: 0.779–0.915) and 0.625 (95% CI: 0.530–0.714), respectively.

Further comparisons of the average T1 and T2 values of the lesions in the three groups of SyMRI images revealed that among the quantitative parameters measured from 2× SyMRI, the T2 value for TNBC was significantly higher than that for non-TNBC (P<0.05), with an AUC of 0.687 (95% CI: 0.551–0.804). Similarly, among the quantitative parameters from 3× SyMRI, the T2 value for TNBC was significantly higher than that for non-TNBC (P<0.05), with an AUC of 0.721 (95% CI: 0.587–0.832). A similar pattern was observed among the quantitative parameters from 3× SyMRI-DLR, where the T2 values for TNBC were significantly higher than those for non-TNBC (P<0.05), with an identical AUC of 0.721 (95% CI: 0.587–0.832) (Table 5 and Figure 5).

Figure 5 Receiver operating characteristic curves showing the diagnostic performance of the T1 and T2 values from 2× SyMRI, 3× SyMRI, and 3× SyMRI-DLR in differentiating breast cancer lesions from normal glandular tissue (left) and in distinguishing TNBC from non-TNBC (right). T1 and T2 represent longitudinal relaxation time and transverse relaxation time, respectively. DLR, deep learning reconstruction; ROC, receiver operating characteristic; SyMRI, synthetic magnetic resonance imaging; TNBC, triple-negative breast cancer.

Consistency analysis

The quantitative and qualitative analysis results reported by the two physicians demonstrated good consistency (ICC >0.75), indicating a high level of repeatability and reliability in clinical practice (Table S2).

Discussion

DLR in MRI accelerates image acquisition while maintaining high quality. This study compared DLR-accelerated SyMRI with standard SyMRI. It was found that 3× SyMRI-DLR reduced scan time significantly; preserved stable T1, T2, and PD measurements; and provided a greater image quality (SNR, clarity, uniformity, and contrast) compared to 3× SyMRI and a similar quality compared to 2× SyMRI, with minor artifacts.

The use of 3× SyMRI with higher acceleration can further reduce the total scan time. However, this reduction in the scan time comes at the expense of SNR and image quality. This issue can be alleviated by employing DLR technology. The 3× SyMRI-DLR protocol not only addresses the issue but also maintains quantitative accuracy comparable to that of 2× SyMRI. Our results showed that although the AUC values of the 3× SyMRI and 3× SyMRI-DLR groups were slightly higher than that of the standard 2× SyMRI group, the differences did not reach statistical significance. This may be due to the fact that the tissue-based T1, T2, and PD values remained stable across all sequences, thus leading to comparable classification performance. From a clinical perspective, this finding is still highly valuable: DLR-enabled 3× acceleration reduced scan time by approximately 37.5% without compromising diagnostic efficacy. This result aligns with those of previous studies on the application of DLR technology in MRI (36,37). By learning the features of high-quality images from large datasets, DLR can compensate for the SNR loss caused by accelerated scanning, thereby maintaining high image quality and quantitative accuracy while improving the imaging speed (21,37-40). Johnson et al. compared accelerated DLR knee MRI with conventional accelerated MRI and reported that DLR nearly halved the scan time while maintaining the diagnostic equivalence (37). Kim et al. demonstrated that the use of DLR technology in pediatric neuroimaging could decrease the scan time by 42% while preserving image quality and lesion detectability (21). Moreover, the SNR of 3× SyMRI-DLR was significantly higher than that of 3× SyMRI and was comparable to or even exceeded that of standard 2× SyMRI. The high SNR level in 3× SyMRI-DLR reduces the impact of background noise on the lesion signal, thereby enhancing the lesion visibility (41). The lower artifact scores observed under the 3× SyMRI and 3× SyMRI-DLR protocols were primarily localized structural aliasing related to parallel imaging reconstruction.

DLR enables the acquisition of high-quality images in shorter scan times, which helps in improving patient tolerance and reducing the likelihood of motion artifacts (42,43). However, DLR has certain limitations, such as a heavy reliance on the training dataset, which may restrict its generalization ability to certain cases; moreover, it requires a large amount of high-quality training data, which can be challenging to obtain in certain clinical settings (42,44). Our DLR algorithm includes a deep CNN, and its main advantage is its extensive training dataset, which helps ensure robust stability and generalization capabilities. Our study found that the tissue T1 and T2 values acquired from the three SyMRI protocols effectively identified breast cancer lesions and distinguished them from normal glandular tissue. In addition, the tissue T2 values obtained by SyMRI showed good diagnostic performance in distinguishing TNBC from non-TNBC, which is in line with previous studies (3,45-47). TNBC is the most aggressive and heterogeneous subtype of breast cancer, with poor prognosis and high treatment challenges. Therefore, the ability to distinguish TNBC from non-TNBC before surgery holds considerable clinical significance (35,48).

This study involved certain limitations which should be noted. First, only 58 patients from a single institution were included, which may not be representative of the broader breast cancer population. Future research should consist of large-scale, multicenter prospective studies to enhance the statistical power. Second, the TE, TR, and inversion recovery values for SyMRI were set as fixed parameters in this study. However, there is currently a lack of standardized and consensus protocols for breast SyMRI scanning parameters. In addition, the reliability of quantitative imaging remains to be validated with physical phantoms, while high acceleration factors can cause parallel imaging-related artifacts to appear in the image. Importantly, our results demonstrated that these artifacts do not significantly affect the quantification of mean T1, T2, and PD values within the lesions or glandular tissue. Although their impact was minimal in breast imaging, further validation may be necessary for other anatomical regions or broader clinical applications. Finally, only ROI-based mean values were used for quantitative analysis, and we did not include advanced approaches such as histogram analysis, radiomic analysis, or deep learning. Our aim was to evaluate whether accelerated SyMRI protocols could retain quantification consistency and diagnostic performance, but we acknowledge that future studies could benefit from more sophisticated metrics to further characterize lesion heterogeneity.

Conclusions

The incorporation of DLR into accelerated synthetic MRI shortens scan time while maintaining quantitative accuracy and image quality. The quantitative parameters of this approach can diagnose breast cancer and differentiate TNBC from non-TNBC, enhancing the clinical application of synthetic MRI in breast diseases.

Acknowledgments

The authors would like to thank the research staff for their help.

Footnote

Reporting Checklist: The authors have completed the STARD reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2024-2972/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2024-2972/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-2024-2972/coif). B.Z. and H.L. are from GE HealthCare MR Research. 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 Institutional Review Board of West China Hospital, Sichuan University (IRB 2023-02-15 No. 187) and informed consent was obtained from all patients.

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. Harbeck N, Gnant M. Breast cancer. Lancet 2017;389:1134-50. [Crossref] [PubMed]
2. Veronesi U, Boyle P, Goldhirsch A, Orecchia R, Viale G. Breast cancer. Lancet 2005;365:1727-41. [Crossref] [PubMed]
3. Zhang Q, Zhao Y, Nie J, Long Q, Wang X, Wang X, Gong G, Liao L, Yi X, Chen BT. Pretreatment synthetic MRI features for triple-negative breast cancer. Clin Radiol 2024;79:e219-26. [Crossref] [PubMed]
4. Mann RM, Cho N, Moy L. Breast MRI: State of the Art. Radiology 2019;292:520-36. [Crossref] [PubMed]
5. Wekking D, Porcu M, De Silva P, Saba L, Scartozzi M, Solinas C. Breast MRI: Clinical Indications, Recommendations, and Future Applications in Breast Cancer Diagnosis. Curr Oncol Rep 2023;25:257-67. [Crossref] [PubMed]
6. Cott Chubiz JE, Lee JM, Gilmore ME, Kong CY, Lowry KP, Halpern EF, McMahon PM, Ryan PD, Gazelle GS. Cost-effectiveness of alternating magnetic resonance imaging and digital mammography screening in BRCA1 and BRCA2 gene mutation carriers. Cancer 2013;119:1266-76. [Crossref] [PubMed]
7. Patel S, Heacock L, Gao Y, Elias K, Moy L, Heller S. Advances in Abbreviated Breast MRI and Ultrafast Imaging. Semin Roentgenol 2022;57:145-8. [Crossref] [PubMed]
8. Ji S, Yang D, Lee J, Choi SH, Kim H, Kang KM. Synthetic MRI: Technologies and Applications in Neuroradiology. J Magn Reson Imaging 2022;55:1013-25. [Crossref] [PubMed]
9. Matsuda M, Tsuda T, Ebihara R, Toshimori W, Takeda S, Okada K, Nakasuka K, Shiraishi Y, Suekuni H, Kamei Y, Kurata M, Kitazawa R, Mochizuki T, Kido T. Enhanced Masses on Contrast-Enhanced Breast: Differentiation Using a Combination of Dynamic Contrast-Enhanced MRI and Quantitative Evaluation with Synthetic MRI. J Magn Reson Imaging 2021;53:381-91. [Crossref] [PubMed]
10. Shin DJ, Choi SH, Yoo RE, Kang KM, Yun TJ, Kim JH, Sohn CH, Jo SW, Lee EJ. Application of T1 Map Information Based on Synthetic MRI for Dynamic Contrast-Enhanced Imaging: A Comparison Study with the Fixed Baseline T1 Value Method. Korean J Radiol 2021;22:1352-68. [Crossref] [PubMed]
11. Fujioka T, Mori M, Oyama J, Kubota K, Yamaga E, Yashima Y, Katsuta L, Nomura K, Nara M, Oda G, Nakagawa T, Tateishi U. Investigating the Image Quality and Utility of Synthetic MRI in the Breast. Magn Reson Med Sci 2021;20:431-8. [Crossref] [PubMed]
12. Yang X, Lu Z, Tan X, Shao L, Shi J, Dou W, Sun Z. Evaluating the added value of synthetic magnetic resonance imaging in predicting sentinel lymph node status in breast cancer. Quant Imaging Med Surg 2024;14:3789-802. [Crossref] [PubMed]
13. Hilbert T, Omoumi P, Raudner M, Kober T. Synthetic Contrasts in Musculoskeletal MRI: A Review. Invest Radiol 2023;58:111-9. [Crossref] [PubMed]
14. Li X, Fan Z, Jiang H, Niu J, Bian W, Wang C, Wang Y, Zhang R, Zhang H. Synthetic MRI in breast cancer: differentiating benign from malignant lesions and predicting immunohistochemical expression status. Sci Rep 2023;13:17978. [Crossref] [PubMed]
15. Kazama T, Takahara T, Kwee TC, Nakamura N, Kumaki N, Niikura N, Niwa T, Hashimoto J. Quantitative Values from Synthetic MRI Correlate with Breast Cancer Subtypes. Life (Basel) 2022;12:1307. [Crossref] [PubMed]
16. Liu J, Xu M, Ren J, Li Z, Xi L, Chen B. Synthetic MRI, multiplexed sensitivity encoding, and BI-RADS for benign and malignant breast cancer discrimination. Front Oncol 2022;12:1080580. [Crossref] [PubMed]
17. Matsuda M, Fukuyama N, Matsuda T, Kikuchi S, Shiraishi Y, Takimoto Y, Kamei Y, Kurata M, Kitazawa R, Kido T. Utility of synthetic MRI in predicting pathological complete response of various breast cancer subtypes prior to neoadjuvant chemotherapy. Clin Radiol 2022;77:855-63. [Crossref] [PubMed]
18. Hamilton J, Franson D, Seiberlich N. Recent advances in parallel imaging for MRI. Prog Nucl Magn Reson Spectrosc 2017;101:71-95. [Crossref] [PubMed]
19. Yoshida T, Urikura A, Shirata K, Nakaya Y, Terashima S, Hosokawa Y. Image quality assessment of single-shot turbo spin echo diffusion-weighted imaging with parallel imaging technique: a phantom study. Br J Radiol 2016;89:20160512. [Crossref] [PubMed]
20. Shafieizargar B, Byanju R, Sijbers J, Klein S, den Dekker AJ, Poot DHJ. Systematic review of reconstruction techniques for accelerated quantitative MRI. Magn Reson Med 2023;90:1172-208. [Crossref] [PubMed]
21. Kim E, Cho HH, Cho SH, Park B, Hong J, Shin KM, Hwang MJ, You SK, Lee SM. Accelerated Synthetic MRI with Deep Learning-Based Reconstruction for Pediatric Neuroimaging. AJNR Am J Neuroradiol 2022;43:1653-9. [Crossref] [PubMed]
22. Lebel RM. Performance characterization of a novel deep learning-based MR image reconstruction pipeline. arXiv 2020. arXiv:2008.06559v1.
23. Kim JH, Yoon JH, Kim SW, Park J, Bae SH, Lee JM. Application of a deep learning algorithm for three-dimensional T1-weighted gradient-echo imaging of gadoxetic acid-enhanced MRI in patients at a high risk of hepatocellular carcinoma. Abdom Radiol (NY) 2024;49:738-47. [Crossref] [PubMed]
24. Wang X, Ma J, Bhosale P, Ibarra Rovira JJ, Qayyum A, Sun J, Bayram E, Szklaruk J. Novel deep learning-based noise reduction technique for prostate magnetic resonance imaging. Abdom Radiol (NY) 2021;46:3378-86. [Crossref] [PubMed]
25. Lee DH, Park JE, Nam YK, Lee J, Kim S, Kim YH, Kim HS. Deep learning-based thin-section MRI reconstruction improves tumour detection and delineation in pre- and post-treatment pituitary adenoma. Sci Rep 2021;11:21302. [Crossref] [PubMed]
26. Iwamura M, Ide S, Sato K, Kakuta A, Tatsuo S, Nozaki A, Wakayama T, Ueno T, Haga R, Kakizaki M, Yokoyama Y, Yamauchi R, Tsushima F, Shibutani K, Tomiyama M, Kakeda S. Thin-slice Two-dimensional T2-weighted Imaging with Deep Learning-based Reconstruction: Improved Lesion Detection in the Brain of Patients with Multiple Sclerosis. Magn Reson Med Sci 2024;23:184-92. [Crossref] [PubMed]
27. Yoo RE, Choi SH. Deep Learning-based Image Enhancement Techniques for Fast MRI in Neuroimaging. Magn Reson Med Sci 2024;23:341-51. [Crossref] [PubMed]
28. Brain ME, Amukotuwa S, Bammer R. Deep learning denoising reconstruction enables faster T2-weighted FLAIR sequence acquisition with satisfactory image quality. J Med Imaging Radiat Oncol 2024;68:377-84. [Crossref] [PubMed]
29. Choi H, Lee SK, Choi H, Lee Y, Lee K. Deep learning-based reconstruction for canine brain magnetic resonance imaging could improve image quality while reducing scan time. Vet Radiol Ultrasound 2023;64:873-80. [Crossref] [PubMed]
30. Estler A, Zerweck L, Brunnée M, Estler B, Richter V, Örgel A, Bürkle E, Becker H, Hurth H, Stahl S, Konrad EM, Kelbsch C, Ernemann U, Hauser TK, Gohla G. Deep learning-accelerated image reconstruction in MRI of the orbit to shorten acquisition time and enhance image quality. J Neuroimaging 2024;34:232-40. [Crossref] [PubMed]
31. Lee EJ, Hwang J, Park S, Bae SH, Lim J, Chang YW, Hong SS, Oh E, Nam BD, Jeong J, Sung JK, Nickel D. Utility of accelerated T2-weighted turbo spin-echo imaging with deep learning reconstruction in female pelvic MRI: a multi-reader study. Eur Radiol 2023;33:7697-706. [Crossref] [PubMed]
32. Zerunian M, Pucciarelli F, Caruso D, Polici M, Masci B, Guido G, De Santis D, Polverari D, Principessa D, Benvenga A, Iannicelli E, Laghi A. Artificial intelligence based image quality enhancement in liver MRI: a quantitative and qualitative evaluation. Radiol Med 2022;127:1098-105. [Crossref] [PubMed]
33. Tanenbaum LN, Tsiouris AJ, Johnson AN, Naidich TP, DeLano MC, Melhem ER, Quarterman P, Parameswaran SX, Shankaranarayanan A, Goyen M, Field AS. Synthetic MRI for Clinical Neuroimaging: Results of the Magnetic Resonance Image Compilation (MAGiC) Prospective, Multicenter, Multireader Trial. AJNR Am J Neuroradiol 2017;38:1103-10. [Crossref] [PubMed]
34. Geng P, Chi Y, Yuan Y, Yang M, Zhao X, Liu Z, Liu G, Liu Y, Zhu L, Wang S. Novel chimeric antigen receptor T cell-based immunotherapy: a perspective for triple-negative breast cancer. Front Cell Dev Biol 2023;11:1158539. [Crossref] [PubMed]
35. Obidiro O, Battogtokh G, Akala EO. Triple Negative Breast Cancer Treatment Options and Limitations: Future Outlook. Pharmaceutics 2023;15:1796. [Crossref] [PubMed]
36. Li J, Wu LH, Xu MY, Ren JL, Li Z, Liu JR, Wang AJ, Chen B. Improving Image Quality and Reducing Scan Time for Synthetic MRI of Breast by Using Deep Learning Reconstruction. Biomed Res Int 2022;2022:3125426. [Crossref] [PubMed]
37. Johnson PM, Lin DJ, Zbontar J, Zitnick CL, Sriram A, Muckley M, Babb JS, Kline M, Ciavarra G, Alaia E, Samim M, Walter WR, Calderon L, Pock T, Sodickson DK, Recht MP, Knoll F. Deep Learning Reconstruction Enables Prospectively Accelerated Clinical Knee MRI. Radiology 2023;307:e220425. [Crossref] [PubMed]
38. Bash S, Wang L, Airriess C, Zaharchuk G, Gong E, Shankaranarayanan A, Tanenbaum LN. Deep Learning Enables 60% Accelerated Volumetric Brain MRI While Preserving Quantitative Performance: A Prospective, Multicenter, Multireader Trial. AJNR Am J Neuroradiol 2021;42:2130-7. [Crossref] [PubMed]
39. Bischoff LM, Peeters JM, Weinhold L, Krausewitz P, Ellinger J, Katemann C, Isaak A, Weber OM, Kuetting D, Attenberger U, Pieper CC, Sprinkart AM, Luetkens JA. Deep Learning Super-Resolution Reconstruction for Fast and Motion-Robust T2-weighted Prostate MRI. Radiology 2023;308:e230427. [Crossref] [PubMed]
40. Xie Y, Tao H, Li X, Hu Y, Liu C, Zhou B, Cai J, Nickel D, Fu C, Xiong B, Chen S. Prospective Comparison of Standard and Deep Learning-reconstructed Turbo Spin-Echo MRI of the Shoulder. Radiology 2024;310:e231405. [Crossref] [PubMed]
41. Goerner FL, Clarke GD. Measuring signal-to-noise ratio in partially parallel imaging MRI. Med Phys 2011;38:5049-57. [Crossref] [PubMed]
42. Kiryu S, Akai H, Yasaka K, Tajima T, Kunimatsu A, Yoshioka N, Akahane M, Abe O, Ohtomo K. Clinical Impact of Deep Learning Reconstruction in MRI. Radiographics 2023;43:e220133. [Crossref] [PubMed]
43. Yi J, Hahn S, Lee HJ, Lee Y, Bang JY, Kim Y, Lee J. Thin-slice elbow MRI with deep learning reconstruction: Superior diagnostic performance of elbow ligament pathologies. Eur J Radiol 2024;175:111471. [Crossref] [PubMed]
44. Hahn S, Yi J, Lee HJ, Lee Y, Lee J, Wang X, Fung M. Comparison of deep learning-based reconstruction of PROPELLER Shoulder MRI with conventional reconstruction. Skeletal Radiol 2023;52:1545-55. [Crossref] [PubMed]
45. Amano M, Fujita S, Takei N, Sano K, Wada A, Sato K, Kikuta J, Kuwatsuru Y, Tachibana R, Sekine T, Horimoto Y, Aoki S. Feasibility of Quantitative MRI Using 3D-QALAS for Discriminating Immunohistochemical Status in Invasive Ductal Carcinoma of the Breast. J Magn Reson Imaging 2023;58:1752-9. [Crossref] [PubMed]
46. Matsuda M, Tsuda T, Ebihara R, Toshimori W, Okada K, Takeda S, Okumura A, Shiraishi Y, Suekuni H, Kamei Y, Kurata M, Kitazawa R, Mochizuki T, Kido T. Triple-negative breast cancer on contrast-enhanced MRI and synthetic MRI: A comparison with non-triple-negative breast carcinoma. Eur J Radiol 2021;142:109838. [Crossref] [PubMed]
47. Herent P, Schmauch B, Jehanno P, Dehaene O, Saillard C, Balleyguier C, Arfi-Rouche J, Jégou S. Detection and characterization of MRI breast lesions using deep learning. Diagn Interv Imaging 2019;100:219-25. [Crossref] [PubMed]
48. Keihan Shokooh M, Emami F, Jeong JH, Yook S. Bio-Inspired and Smart Nanoparticles for Triple Negative Breast Cancer Microenvironment. Pharmaceutics 2021;13:287. [Crossref] [PubMed]