Validation and feasibility of a deep learning-based reconstruction technology in 5.0 tesla knee joint MR imaging

This study aimed to evaluate the feasibility of a deep learning-based reconstruction (DLR) algorithm for optimizing conventional 5.0 Tesla knee joint MR protocols.

This prospective study enrolled 69 patients who underwent both knee arthroscopy and 5.0 Tesla knee joint MR examinations using the conventional protocols before and after a DLR process with different levels. The DLR technique was applied to original images to denoise and improve their quality. Two radiologists independently measured the signal-to-noise ratio (SNRs) in cartilage, meniscus, bone, ligament, and muscle, and graded image quality from the dimensions of different tissues' delineation clarity, global artifact severity, and overall image quality using a 5-point Likert scale. Moreover, the diagnostic performance was evaluated with different types of images, compared to the results of knee arthroscopy. Cohen's kappa test was employed to assess the agreement of image quality scoring and diagnosis.

Compared to conventional images, those DLR ones demonstrated significant improvement in SNRs, with the increasement of 12.61% to 350.63% across various sequences. Two radiologists showed good-to-excellent agreement in image quality assessment, with kappa values ranging from 0.72 to 0.82. Regarding diagnostic performance, the DLR images moderately outperformed the non-DLR ones, as evidenced by a bit higher diagnostic agreement with the results of knee arthroscopy (DLR: kappa = 0.908-1; non-DLR: kappa = 0.882-0.963).

The DLR technique could improve 5.0 Tesla knee MR images' quality and obtain as least equal diagnostic efficiency without extra scan time, demonstrating its potential clinical applicability.