Feasibility and clinical value of deep learning reconstruction in accelerated MRI of shoulder

Abstract

Objective To investigate the feasibility and clinical value of the deep learning reconstruction (DLR) algorithm in enhancing image quality and reducing scan time in shoulder MRI. Methods Fifty patients suspected of having shoulder joint lesions in the Fourth Affiliated Hospital of Nanjing Medical University from June to October 2023 were prospectively included. Routine sequence scanning images with 1.5T MRI were grouped as Fsecon, while scanning images with parallel acquisition acceleration factor 2 were grouped as Fsefast. The scanning sequences included fat-suppression proton density-weighted imaging (PDWI-FS) and T1 weighted imaging (T1WI). The Fsefast group was then transferred to Subtle MRTM dlr to obtain images in the Fsedlr group. The signal-to-noise ratio (SNR) of supraspinatus muscle, long head tendon of biceps brachii, glenoid labrum cartilage, and humerus marrow, as well as the contrast-to-noise ratio (CNR) of supraspinatus muscle/glenoid labrum cartilage, were measured and compared among the three groups. Two radiologists blindly evaluated the image clarity and artifacts of the Fsedlr group and the Fsecon group using the Likert 4-point scale, and compared the diagnostic efficacy of pathological abnormal structures between the two groups. Results Compared with the Fsecon group, the scan time of the Fsedlr group was shortened by 44%, and the image clarity and artifact scores were higher, with statistically significant differences (P＜0.05). The intraclass correlation coefficients（ICC）for subjective scores within the two radiologists' groups were 0.797 to 0.919. Among the objective evaluation indicators, the SNR and CNR of the Fsedlr group were significantly higher than those of the Fsecon group and the Fsefast group, with statistically significant differences (P＜0.05). In the evaluation of the pathological abnormal structures of the Fsecon group and the Fsedlr group by two radiologists, the diagnostic results of the two groups were consistent (Kappa value: 0.675-1.000), and also showed excellent consistency in the evaluation of the same radiologist (Kappa value: 0.771-1.000), among which the Kappa values of humerus bone marrow, joint bursa and long head tendon of biceps brachii were higher than 0.8. Conclusion The application of the DLR algorithm in shoulder MRI examination can improve image quality, shorten image acquisition time, ensure diagnostic efficacy, and improve examination efficiency, demonstrating good clinical value.

Keywords

• deep learning reconstruction， shoulder joint， magnetic resonance image， signal-to-noise ratio， contrast-to-noise ratio， image quality