Deep Learning reconstruction for fast cardiac MRI protocol: A Comparative Study with Conventional cardiac MR.
Authors
Affiliations (4)
Affiliations (4)
- Department of Radiology, Zhongshan Hospital, Fudan University and Shanghai Institute of Medical Imaging, Shanghai, China.
- Philips Healthcare, Beijing, China.
- Department of Radiology, Zhongshan Hospital, Fudan University and Shanghai Institute of Medical Imaging, Shanghai, China; Department of Radiology, Zhongshan Hospital Minhang Campus, Fudan University and Shanghai Geriatric Medical Center, Shanghai, China. Electronic address: [email protected].
- Department of Radiology, Zhongshan Hospital, Fudan University and Shanghai Institute of Medical Imaging, Shanghai, China; Department of Radiology, Zhongshan Hospital Minhang Campus, Fudan University and Shanghai Geriatric Medical Center, Shanghai, China. Electronic address: [email protected].
Abstract
Cardiac magnetic resonance (CMR) is a reference-standard modality for heart diseases, although its clinical application is restricted by prolonged acquisition times. Recently, artificial intelligence (AI), particularly deep learning (DL), has exhibited the potential to accelerate the CMR acquisition through technological advances. Prospective validation of its diagnostic performance across multiple clinical sequences remains underexplored. This research aims to assess the functions of the compressed sensing artificial intelligence (CSAI) algorithm in accelerating CMR acquisition, enhancing image quality, and maintaining diagnostic accuracy versus conventional sensitivity encoding (SENSE) reconstruction. A total of 105 participants scheduled for clinical CMR between February and August 2024 underwent both SENSE and CSAI-accelerated sequences containing Cine, T2 short TI inversion recovery (STIR), and late gadolinium enhancement (LGE) during a single session. The subjective image quality was assessed by a 5-point Likert scale. Quantitative image quality metrics were evaluated including SNR, CNR, edge sharpness, ventricular function, T2 signal intensity (SI) ratio, and LGE percentage. Using standard resolution, the acquisition time of CSAI-CMR was 57.4% lower than that of SENSE CMR (159.2 ± 22.4seconds vs 277.1 ± 30.4seconds; P <.001). Higher subjective scores could be found in CSAI-Cine and CSAI-T2 STIR compared with SENSE CMR (P <.001). Quantitative analyses demonstrated superior SNR and CNR across CSAI sequences in comparison with SENSE-CMR including Cine (SNR: 109.16 ± 20.36 vs. 102.52 ± 21.93, CNR: 58.98 ± 24.43 vs. 51.08 ± 23.37; both P<.001), T2 STIR (SNR: 98.17 ± 6.20 vs. 81.77 ± 11.15; P<.001), and LGE (SNR: 38.02 ± 7.90 vs. 32.54 ± 7.72, CNR: 22.24 ± 5.15 vs. 19.09 ± 4.22; both P<.001). Edge sharpness was significantly improved by using CSAI-CMR (0.141 ± 0.06 pixel⁻¹ vs 0.105 ± 0.04 pixel⁻¹; P <.01). Functional parameters, both T2 SI ratio and LGE percentage were comparable relatively (all P >.05). DL reconstruction of CMR sequences reduced acquisition times by 57% and enhanced image quality compared to conventional SENSE reconstruction, while maintaining consistent quantitative parameters and diagnostic accuracy across all sequences. These results advocate integrating DL-accelerated workflows into clinical practice.