Deep-learning reconstructed 3D MRI for comprehensive knee assessment: Comparison with a multisequence 2D protocol at 1.5 T.

OBJECTIVE

To compare a single isotropic 3D PD-weighted fat-saturated (3D-PDFS) acquisition with a standard 2D multisequence protocol in knee MRI using deep learning reconstruction (DLR) for comprehensive assessment of knee structures.

MATERIALS AND

METHODS

We retrospectively collected consecutive adult patients undergoing 1.5 T knee MRI between May 2023 and July 2024. The MRI protocol comprised one 3D-PDFS and five multiplanar 2D sequences with DLR. Two radiologists independently evaluated the 3D and 2D sequences, assessing abnormalities for medial (MC), lateral (LC), and patellofemoral (PF) compartments, overall image quality, diagnostic confidence, and artifacts. Likert-scale and binary grades were compared via Wilcoxon signed-rank and McNemar tests, respectively. Inter-reader agreement was assessed via Cohen's kappa.

RESULTS

A total of 95 patients (females, 39 %; age, 52.7 ± 14.5 y) were included. The 3D and 2D protocols did not significantly differ regarding the assessment of menisci, MC and PF cartilage, collateral ligaments, cruciate ligaments, MC and PF bone marrow edema (BME), and fractures (p > 0.057). Significant differences were shown in assessment of LC cartilage (p = 0.002), LC BME (p = 0.040) and synovitis (p = 0.002). Image quality and artifacts were not significantly different (p > 0.082). Diagnostic confidence of one reader was significantly higher for the 2D protocol (p = 0.023). Average inter-reader agreement was substantial for the 3D (k = 0.67) and 2D (k = 0.66) protocols.

CONCLUSION

Our results suggest mostly comparable performance between a single 3D-PDFS and a multisequence 2D protocol using DLR for knee assessment, with the exception of few structures.