High-Resolution Bone Image from shoulder MRI Using Deep Neural Network on 3-D Accelerated Dixon GRE (CAIPIRINHA Dixon)

Authors

    Sheen Woo Lee1, Seungwook Yang2, Jooyeon Kim1


     
     

    1The Catholic University of Korea, Eunpyeong St. Mary’s Hospital, Korea, Republic of
    2AIRS Medical, Korea, Republic of
    [email protected]

    KCR (2023)

    PURPOSE

    To test the feasibility of generating high-resolution bone image resembling CT from shoulder 3-D CAIPIRINHA Dixon MRI data using deep neural network.

    MATERIALS AND METHOD

    In this IRB-approved retrospective study, patients with 3.0T MRI for shoulder pain were enrolled. Those without CT or 3D VIBE CAIPI-DIXON were excluded. The 3D VIBE CAIPI-DIXON protocol was ; TR/TE 12.7 /2.5,+3.6ms; FA 10°; FOV 159 x 159 mm; spatial resolution 0.4×0.4 mm3; acquisition time 3.02 min (Magnetom Vida, Siemens Healthineers, Germany). The dicom files were reconstructed using a commercially available DNN-based denoising and resolution enhancement algorithm (SwiftMR, AIRS Medical, Korea). The in-phase and opposed phase CAIPI-Dixon original PACS image (“CAIPI-i-ori”, “CAIPI-o- ori”), and the corresponding CAIPI-Dixon after DNN processing images (“CAIPI-i-dn”, “CAIPI-o-dn”) were viewed after gray-scale inversion, and quantitatively
    and qualitatively evaluated by two radiologists. The images were first orthogonally rotated to measure the diameters of the glenoid. Studies were scored from 1 to 5 for the clarity of the cortical outline and trabecular bone of the humeral head and glenoid, any pseudolesion, visibility of fractures. The humeral head,
    glenoid neck, deltoid, and infraspinatus muscles were marked with regions of interest. ROIs were drawn on vacant quadrants for noise. CT was used to compare the glenoid dimension measurements, ROI values, SNR, and CNR. For statistical analysis, Kruskall-Wallis and Spearman tests were used (p<0.05).

    RESULTS

    Final group included 10 patients (4 females, mean age 60.6 years). The glenoid diameters measured from CAIPI-i-dn, o-dn, CAIPI i-ori, o-ori, and CT
    did not differ significantly (p>0.05). CAIPI-i-dn, o-dn, and CAIPI-o-ori had higher humerus cortical scores than CAIPI-i-ori (4.94±0.236, 3.611±0.195, 4.39±0.916, vs 3.00±.0.29). CAIPI-i-o-dn showed higher glenoid cortical scores than CAIPI-i-o-ori (4.17±1.043 than 3.56±1.247, p<0.05). SNR of humerus and glenoid
    significantly improved with CAIPI-dn compared to CAIPI-ori (4 from 1.07, p<0.05). CT density negatively correlated with CAIPI-o-dn ROI in the humerus and
    glenoid (correlation coefficient -53.3, and -60.3, p<0.05). ICC value of the semiquantitative scores of CAIPI-i- o-dn (eg. cortical and trabecular outline) was 79.5 among the two readers. ICC value of quantitative scores (eg. ROI) was 96.4 (p<0.05).

    CONCLUSION

    Combining 3D VIBE CAIPI-DIXON MRI and the DNN algorithm enabled high resolution cortical and trabecular bone imaging.