Stability of Deep Learning-based Image Quality Improvement in MRI of the Knee: Correlation with Arthroscopy


    Seung Hoon Choi1, Hee Rin Lee1, Joon-Yong Jung1, Seung Eun Lee1, So Hyun Kim2, Geunu Jeong2


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


    KCR (2023)


    To validate whether deep learning-based image quality improvement (DL-IQI) changes the diagnostic performance and visibility of core features for meniscal, cartilage, and ligament knee lesions.


    This retrospective study include 106 patients, who underwent knee MRI and arthroscopy within a 3-month interval. Fat-suppressed 2D fast spin-echo (FSE) with three orthogonal planes were improved using pre-developed DL-IQI. Two musculoskeletal radiologists independently reviewed both the original and DL-IQI-processed MRI for medial meniscus (MM), lateral meniscus (LM), articular cartilage (AC) and cruciate ligaments (CL). Sensitivity and specificity were compared with arthroscopic results used as the reference standard, and interobserver agreements were also compared. Additionally, two radiologists reviewed target lesions on consensus, which were pre-defined based on arthroscopic images and MRI. On each target lesions, the MRI was scored with respect to core features: tear plane, flap, fraying and degeneration for meniscal tear; depth and surface contour for cartilage lesions; degeneration and tear for cruciate ligament. The agreement score on the target lesions were compared among original (O) and DL-IQI processed MRI with low (L), medium (M), very high (VH) de-noising levels.


    In reader 1, the sensitivity and specificity (%) of original vs. DL-IQI-processed MRI were 91.0 vs. 92.5 (P>0.99) and 92.3 vs. 97.4 (P=0.50) for MM, 88.9 vs. 93.3 (P=0.50) and 90.1 vs. 91.8 (P>0.99) for LM, 87.2 vs. 94.8 (*P<0.05) and 92.8 vs. 85.7 (P=0.625) for AC, and 91.3 vs. 91.3 (Non-Applicable, NA) and 98.8 vs. 98.8 (NA) for CL. In reader 2, the sensitivity and specificity of original vs. DL-IQI-processed MRI were 86.6 vs. 91.4 (P=0.25) and 97.4 vs. 97.4 (NA) for MM, 86.7 vs. 86.7 (P>0.99) and 88.5 vs. 88.5 (P>0.99) for LM, 78.2 vs. 82.5 (P=0.37) and 96.4 vs. 89.3 (P=0.5) for AC, and 78.3 vs. 78.3 (NA) and 98.8 vs. 98.8 (NA) for CL. The inter-observer agreements (κ) for MM, LM, AC, and CL were 0.74, 0.77, 0.76, and 0.90 in the original MRI and 0.90, 0.87, 0.72, and 0.90 in the DL-IQI-processed MRI. The mean agreement scores between arthroscopy and MRI for target lesions were 7.18(O), 7.33(L), 7.36(M) and 7.34(VH) for meniscal tears (n=130), and 2.53(O), 3.23(L), 3.24(M) and 3.24(VH) for cartilage lesions (n=84).


    DL-based image improvement of knee MRIs can enhance the identification of cartilage lesions, without affecting overall diagnostic performances as correlated with arthroscopic results.