Arthroscopy-Validated Diagnostic Performance of a Deep Learning Reconstruction Pipeline for Rapid 7-Minute Five-Sequence 3-T Knee MRI.
Authors
Affiliations (6)
Affiliations (6)
- Department of Radiology, Grossman School of Medicine, New York University, New York, NY.
- Department of Diagnostic and Interventional Radiology, School of Medicine & Klinikum Rechts Der Isar, Technical University of Munich, Munich, Germany.
- Department of Radiology, University Hospital Basel, Basel, Switzerland.
- Department of Radiology and Biomedical Imaging, University of California, San Francisco, CA.
- Medscanlagos Radiology, Cabo Frio, Brazil.
- Department of Radiology, Hospital Do Coraçao, São Paulo, Brazil.
Abstract
<b>Background.</b> Deep learning reconstruction (DLR) methods can enhance image quality and reduce scan time of knee MRI compared with conventional approaches but require validation against independent reference standards to ensure robustness and accuracy. <b>Objective.</b> The purpose of this study was to assess the diagnostic performance of a two- to threefold parallel imaging-accelerated 7-minute five-sequence 3-T knee MRI protocol using a deep learning-based image reconstruction pipeline (AIR Recon DL, GE HealthCare), with arthroscopic surgery as the reference standard. <b>Methods.</b> A total of 117 consecutive adult patients (mean age: 44 ± 16 [SD] years; 65 men, 52 women) with painful knee conditions who underwent DLR 3-T knee MRI and arthroscopic knee surgery with a median MRI-tosurgery interval of 35 days (range: 5-88 days) between June 2021 and May 2024 were retrospectively identified and included. MRI studies were independently reviewed by seven musculoskeletal radiologists for image quality parameters using Likert scales (range: 1 = very bad to 5 = very good) and the presence of meniscus tears, cruciate and collateral ligament tears, and articular cartilage defects. Statistical analyses included interreader agreements and diagnostic performance testing. <b>Results.</b> Overall image quality of DLR knee MRI scans was good (median: 4 [IQR, 4-5]), with minimal image noise (4 [4-4]), good edge sharpness (4 [4-5]), absence of reconstruction artifacts (5 [4-5]), and high interreader agreement for all quality metrics (κ ≥ 0.83). Diagnostic performance for detecting arthroscopy-validated structural abnormalities was very good (AUC ≥ 0.81) with good to very good interreader agreement (κ ≥ 0.62). The sensitivity, specificity, accuracy, and AUC values were 89%, 87%, 88%, and 0.87 for medial meniscus tears (prevalence at arthroscopy: 78/117; 67%), 72%, 91%, 83%, and 0.82 for lateral meniscus tears (50/117; 43%), 97%, 98%, 97%, and 0.97 for anterior cruciate ligament tears (30/117; 26%), and 73%, 89%, 86%, and 0.81 for articular cartilage defects (182/702; 26%). <b>Conclusion.</b> Clinical 7-minute five-sequence 3-T knee MRI with deep learning reconstruction provides good to excellent diagnostic performance for detecting arthroscopy-validated internal derangement of the knee. <b>Clinical Impact.</b> Deep learning reconstruction enables rapid high-quality clinical 3-T knee MRI with high diagnostic performance for arthroscopy-validated abnormalities.