Share：
Share this content in WeChat
X
[Chinese] [PDF] 35 1
• Technical Article •
A comparative study of AI-assisted compressed sensing combined with deep-learning reconstruction for high-resolution T2-weighted breast imaging at 5.0 T MRI
PAN Ziwei  DING Jie  MA Minjia  YANG Jie  CHEN Dazhi  ZHOU Liqiang  GAO Yongbin  ZHU Rongrong 

Cite this article as: PAN Z W, DING J, MA M J, et al. A comparative study of AI-assisted compressed sensing combined with deep-learning reconstruction for high-resolution T2-weighted breast imaging at 5.0 T MRI[J]. Chin J Magn Reson Imaging, 2026, 17(10): 153-160, 217. DOI:10.12015/issn.1674-8034.2026.10.018.


[Abstract] Objective To investigate the effects of artificial intelligence-assisted compressed sensing (ACS) with different acceleration factors combined with deep learning reconstruction (DLR) of varying strengths on scan time and image quality of high-resolution 5.0 T breast T2WI.Materials and Methods Fifty-five patients (109 breasts) were prospectively enrolled. All patients underwent T2-weighted scans with three acceleration factors (ACS2, ACS5, ACS8). Images acquired with each ACS setting were reconstructed using non-DLR (Non-DLR), moderate-strength DLR (DLR-3), and high-strength DLR (DLR-5). Scan times under different acceleration factors were recorded. Subjective and objective image quality assessments were performed across combinations of ACS and DLR parameters. Objective evaluation included signal-to-noise ratio (SNR) of glandular tissue, adipose tissue and pectoralis major muscle, lesion contrast-to-noise ratio (CNR), and background standard deviation (SD). Two radiologists performed 5-point Likert-scale subjective scoring for overall image sharpness, lesion boundary delineation, adjacent-organ clarity, image distortion, and motion artifacts. Weighted Kappa test was used to assess inter-observer agreement. The Friedman χ2 test was applied to compare differences in SNR and CNR among various ACS-DLR combinations.Results Scan time decreased significantly with increasing ACS acceleration factors: 7 min 58 s for ACS2, 4 min 53 s for ACS5, and 3 min 59 s for ACS8. Under identical ACS settings, increasing DLR strength (0, 3, 5) yielded higher signal intensity (SI) and SNR for lesions and glandular tissue as well as reduced background SD (P < 0.001), whereas no significant difference was found in CNR (P > 0.05). With the same DLR strength, lesion SI tended to decrease as the ACS acceleration factor increased; lesion SNR first increased and then decreased. No statistically significant difference in lesion SNR was observed between ACS5 and ACS8 under identical DLR conditions (P > 0.05); background SD decreased (P < 0.001), while CNR showed no significant difference (P > 0.05). The ACS8 combined with DLR-5 protocol achieved the highest subjective image scores (P < 0.05). Inter-observer agreement for the five subjective dimensions was good (Kappa: 0.783 to 0.818). The intra-class correlation coefficient (ICC) for composite subjective scores ranged from 0.831 to 0.912, with the lower bound of the 95% confidence interval (CI) as low as 0.721, indicating moderate-to-good agreement between the two radiologists and satisfying the requirements for clinical research.Conclusions ACS can markedly reduce background SD and improve SNR. The ACS8 combined with DLR-5 protocol shortens scan time by 50% compared with ACS2 while maintaining favorable image quality. High-resolution T2WI provides a technical foundation for non-contrast fast breast imaging.
[Keywords] artificial intelligence-assisted compressed sensing；deep-learning reconstruction；high-definition T2WI；5.0 T breast magnetic resonance imaging；signal-to-noise ratio

PAN Ziwei1   DING Jie1*   MA Minjia1   YANG Jie1   CHEN Dazhi1   ZHOU Liqiang2   GAO Yongbin1   ZHU Rongrong1  

1 Medical Imaging Center, Ningxia Hui Autonomous Region People's Hospital (Affiliated Autonomous Region People's Hospital of Ningxia Medical University/Third Clinical Medical College), Yinchuan 750021, China

2 Shanghai United Imaging Medical Technology Co., Ltd., Shanghai 201800, China

Corresponding author: DING J, E-mail: jie19880119@126.com

Conflicts of interest   None.

Received  2026-06-20
Accepted  2026-09-22
DOI: 10.12015/issn.1674-8034.2026.10.018
Cite this article as: PAN Z W, DING J, MA M J, et al. A comparative study of AI-assisted compressed sensing combined with deep-learning reconstruction for high-resolution T2-weighted breast imaging at 5.0 T MRI[J]. Chin J Magn Reson Imaging, 2026, 17(10): 153-160, 217. DOI:10.12015/issn.1674-8034.2026.10.018.

[1]
LUO W B, ZHENG Y, LIU X, et al. Value analysis of deep learning model based on DCE-MRI images in the differential diagnosis of benign and malignant breast tumors[J]. Chin J Magn Reson Imaging, 2024, 15(10): 22-29. DOI: 10.12015/issn.1674-8034.2024.10.005.
[2]
WEKKING D, PORCU M, DE SILVA P, et al. Breast MRI: clinical indications, recommendations, and future applications in breast cancer diagnosis[J]. Curr Oncol Rep, 2023, 25(4): 257-267. DOI: 10.1007/s11912-023-01372-x.
[3]
WU P Q, YANG Y L, ZHOU Y L, et al. Construction of MRI radiomic prediction models for the differentiation of benign and malignant lesions of breast[J]. J Mol Imag, 2021, 44(5): 764-770. DOI: 10.12122/j.issn.1674-4500.2021.05.07.
[4]
WANG W J, LI J J, WANG Z M, et al. Application of deep learning model fusion of T 2WI and DCE-MRI in breast lesion classification[J]. J Clin Radiol, 2024, 43(11): 1871-1876. DOI: 10.13437/j.cnki.jcr.2024.11.033.
[5]
ZHANG L Y, ZHANG T Z, ZHAO X. Diagnostic value of DWI and T2WI morphological assessment based on breast imaging reporting and data systems for breast lesions[J]. J South Med Univ, 2025, 45(9): 1809-1817. DOI: 10.12122/j.issn.1673-4254.2025.09.01.
[6]
CETINTAS D, TUNCER T, KILICARSLAN G, et al. Dense-MoE vs lite-MoE: a gating-weight-aware pruning framework for unpaired multimodal breast cancer diagnosis[J/OL]. J Imag Inform Med, 2026 [2026-06-19]. https://link.springer.com/article/10.1007/s10278-026-02026-7. DOI: 10.1007/s10278-026-02026-7.
[7]
MESROPYAN N, KATEMANN C, LEUTNER C, et al. Accelerated high-resolution T1- and T2-weighted breast MRI with deep learning super-resolution reconstruction[J]. Acad Radiol, 2025, 32(6): 3147-3156. DOI: 10.1016/j.acra.2024.12.055.
[8]
YANG F, JIANG Y, XIAO Y T, et al. Deep learning reconstruction for accelerating synthetic magnetic resonance imaging of the breast: a comparative analysis of accelerated and standard protocols[J]. Quant Imaging Med Surg, 2025, 15(10): 8953-8967. DOI: 10.21037/qims-2024-2972.
[9]
LIU J X, GAN J. 5.0T ultra-high-field whole-body magnetic resonance imaging: technological breakthroughs and challenges[J]. Chin J Med Imag, 2025, 33(7): 689-693. DOI: 10.3969/j.issn.1005-5185.2025.07.001.
[10]
YANG F, PAN X L, ZHU K, et al. Accelerated 3D high-resolution T2-weighted breast MRI with deep learning constrained compressed sensing, comparison with conventional T2-weighted sequence on 3.0 T[J/OL]. Eur J Radiol, 2022, 156: 110562 [2026-06-19]. https://www.ejradiology.com/article/S0720-048X(22)00412-0/abstract. DOI: 10.1016/j.ejrad.2022.110562.
[11]
LIU S W, YUAN H S. Advances and challenges of ultra-high-field 5.0 T MRI technology in non-neurological systems application[J]. Mod Instrum Med Treat, 2025, 31(6): 1-6. DOI: 10.11876/mimt202506001.
[12]
LI S, WANG X C. Clinical advancements in the application of compressed sensing techniques in magnetic resonance imaging[J]. Chin J Magn Reson Imaging, 2023, 14(12): 198-202. DOI: 10.12015/issn.1674-8034.2023.12.036.
[13]
NING K, XU H, LI X M, et al. Sequent optimization of AI-assisted compressive sensing techniques in brain 3D-TOF-MRA[J]. China Med Equip, 2025, 22(10): 15-19. DOI: 10.3969/j.issn.1672-8270.2025.10.003.
[14]
SHEN L T, XU H, LIAO Q, et al. A feasibility study of AI-assisted compressed sensing in prostate T2-weighted imaging[J]. Acad Radiol, 2024, 31(12): 5022-5033. DOI: 10.1016/j.acra.2024.06.048.
[15]
WANG Q Z, ZHAO W L, XING X Y, et al. Feasibility of AI-assisted compressed sensing protocols in knee MR imaging: a prospective multi-reader study[J]. Eur Radiol, 2023, 33(12): 8585-8596. DOI: 10.1007/s00330-023-09823-6.
[16]
JIANG H T, YUE Y L, SHEN Q L, et al. Application of 3T contrast-enhanced Dixon water-fat separation coronary magnetic resonance angiography with artificial intelligence-assisted compressed sensing on coronary artery aneurysm in children with Kawasaki disease[J/OL]. Transl Pediatr, 2026, 15(1): 6 [2026-06-19]. https://pubmed.ncbi.nlm.nih.gov/41657456/. DOI: 10.21037/tp-2025-aw-731.
[17]
YE Z, LYU X Y, ZHAO R P, et al. Accelerated MRI in temporomandibular joints using AI-assisted compressed sensing technique: a feasibility study[J]. Eur Radiol, 2025, 35(12): 8046-8057. DOI: 10.1007/s00330-025-11734-7.
[18]
ZHANG T, ZHANG L X, XU X E, et al. Application value of artificial intelligence compressed sensing technology in single hip joint MRI with different coils[J]. China Med Devices, 2025, 40(9): 123-128, 167. DOI: 10.3969/j.issn.1674-1633.20250912.
[19]
ZHOU J X, MOU Q H, SONG Y, et al. Comparative analysis and influencing factors of aneurysm wall characteristics based on artificial intelligence-assisted 3.0T MRI and 5.0T MRI high-resolution vascular wall imaging[J/OL]. BMC Med Imag, 2025, 26(1): 9 [2026-06-19]. https://link.springer.com/article/10.1186/s12880-025-01959-9. DOI: 10.1186/s12880-025-01959-9.
[20]
JIANG X T, CHENG T X, LI F F, et al. Application of artificial intelligence compressive sensing technology in MRI of the ankle joint[J]. Chin J Med Imag, 2024, 32(11): 1164-1169. DOI: 10.3969/j.issn.1005-5185.2024.11.013
[21]
LUO H D, LIU Y Y Q, ZHANG H N, et al. Optimization of multi-parameter MRI with flexible design sequences in the saddle area based on artificial intelligence-assisted compressed sensing technology[J]. Chin J Magn Reson Imaging, 2024, 15(12): 150-156. DOI: 10.12015/issn.1674-8034.2024.12.022.
[22]
YAN X H, LUO Y, RAN L P, et al. Clinical application of high resolution myocardial T2-weighted dark blood sequence based on artificial intelligence assisted compressed sensing technique in myocardial edema[J]. Chin J Magn Reson Imaging, 2022, 13(6): 76-80, 97. DOI: 10.12015/issn.1674-8034.2022.06.015.
[23]
HUANG B Y, MA Z Y, CHEN Z T, et al. Impact of deep learning reconstruction in MRI on scan time and image quality of breast fat suppression sequence[J]. J Guizhou Med Univ, 2024, 49(6): 901-905, 911. DOI: 10.19367/j.cnki.2096-8388.2024.06.016.
[24]
ZHOU S H, WANG X J, CHENG S J, et al. Diagnostic value of multi-parameter MRI of tumor and peritumoral tissue for breast cancer and its pathological factors[J]. Chin J Magn Reson Imaging, 2023, 14(3): 72-80. DOI: 10.12015/issn.1674-8034.2023.03.013.
[25]
CAO D, YU R L, HU H, et al. Study on compressed sensing and deep learning reconstruction techniques applied in orbital MRI[J]. Chin Med Equip J, 2024, 45(10): 60-65. DOI: 10.19745/j.1003-8868.2024195.
[26]
DONG X, SHI D M, QIN Z W, et al. Comparison of reduced FOV diffusion-weighted imaging of rectal cancer at 5.0T ultra-high field versus 3.0T MRI: image quality and histopathological T staging[J/OL]. BMC Med Imag, 2025, 26(1): 29 [2026-06-19]. https://link.springer.com/article/10.1186/s12880-025-02076-3. DOI: 10.1186/s12880-025-02076-3.
[27]
LÜ Y Y, YU J L, ZHANG L W, et al. Artificial intelligence-assisted compressed sensing versus parallel imaging in placental MRI: a comparative study[J]. Chin Comput Med Imag, 2026, 32(1): 116-122. DOI: 10.3969/j.issn.1006-5741.2026.01.020.
[28]
ZHAO Y J, PENG C D, WANG S F, et al. The feasibility investigation of AI-assisted compressed sensing in kidney MR imaging: an ultra-fast T2WI imaging technology[J/OL]. BMC Med Imag, 2022, 22(1): 119 [2026-06-19]. https://link.springer.com/article/10.1186/s12880-022-00842-1. DOI: 10.1186/s12880-022-00842-1.
[29]
LIU H B, DENG D L, ZENG W L, et al. AI-assisted compressed sensing and parallel imaging sequences for MRI of patients with nasopharyngeal carcinoma: comparison of their capabilities in terms of examination time and image quality[J]. Eur Radiol, 2023, 33(11): 7686-7696. DOI: 10.1007/s00330-023-09742-6.
[30]
WEN Y, MA H, XIANG S X, et al. Accelerating brain T2-weighted imaging using artificial intelligence-assisted compressed sensing combined with deep learning-based reconstruction: a feasibility study at 5.0T MRI[J/OL]. BMC Med Imag, 2025, 25(1): 224 [2026-06-19]. https://link.springer.com/article/10.1186/s12880-025-01763-5. DOI: 10.1186/s12880-025-01763-5.
[31]
SHI D, SHI Y N, WANG Q Y, et al. Predictive value of multiparametric MRI radiomics in preoperative Ki-67 expression levels in invasive breast cancer[J]. J Pract Radiol, 2026, 42(2): 236-241. DOI: 10.3969/j.issn.1002-1671.2026.02.014.
[32]
ZHANG N, DU F, CHEN C Z, et al. Feasibility of free-breathing deep learning-reconstructed single-shot cine MRI in participants with arrhythmia: comparison with conventional segmented cine MRI[J/OL]. Radiol Cardiothorac Imag, 2026, 8(2): e250298 [2026-06-19]. https://pubs.rsna.org/doi/10.1148/ryct.250298. DOI: 10.1148/ryct.250298.
[33]
YUAN C, CAO J J, BAI M, et al. Optimization of knee joint MR imaging based on artificial intelligence compressed sensing technology combined with deep learning reconstruction techniques[J]. J China Clin Med Imag, 2026, 37(2): 134-137. DOI: 10.12117/jccmi.2026.02.014.
[34]
JIANG G L, XU H, CHENG X, et al. The application of artificial intelligence-assisted compressed sensing technology in MRI of the temporomandibular joint[J]. J Pract Radiol, 2026, 42(2): 293-297. DOI: 10.3969/j.issn.1002-1671.2026.02.026.
[35]
ZHOU P H, LIU H B, LIN H, et al. Application of deep learning-based image reconstruction technology in 5.0T MRI for nasopharyngeal carcinoma[J]. Chin J Med Imaging, 2025, 33(7): 694-699. DOI: 10.3969/j.issn.1005-5185.2025.07.002.

PREV Preparation of Fe<sub>3</sub>O<sub>4</sub>-PSMA nanomolecular imaging probe for prostate cancer and its magnetic resonance imaging study in tumor-bearing mice
NEXT Intramedullary spinal subependymoma with pial enhancement: A case report
  


LINK


Tel & Fax: +8610-67113815    E-mail: editor@cjmri.cn