Deep learning reconstruction for temporomandibular joint MRI: diagnostic interchangeability, image quality, and scan time reduction

Abstract

Objectives To evaluate the diagnostic interchangeability, image quality, and scan time of deep learning (DL)-reconstructed magnetic resonance imaging (MRI) compared with conventional MRI for the temporomandibular joint (TMJ).<br /> Materials and methods Patients with suspected TMJ disorder underwent sagittal proton density-weighted (PDW) and T2-weighted fat-suppressed (T2W FS) MRI using both conventional and DL reconstruction protocols in a single session. Three oral radiologists independently assessed disc shape, disc position, and joint effusion. Diagnostic interchangeability for these findings was evaluated by comparing interobserver agreement, with equivalence defined as a 95% confidence interval (CI) within +/- 5%. Qualitative image quality (sharpness, noise, artifacts, overall) was rated on a 5-point scale. Quantitative image quality was assessed by measuring the signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) in the condyle, disc, and background air. Image quality scores were compared using the Wilcoxon signed-rank test, and SNR/CNR using paired t-tests. Scan times were directly compared.<br /> Results A total of 176 TMJs from 88 patients (mean age, 37 +/- 16 years; 43 men) were analyzed. DL-reconstructed MRI demonstrated diagnostic equivalence to conventional MRI for disc shape, position, and effusion (equivalence indices < 3%; 95% CIs within +/- 5%). DL reconstruction significantly reduced noise in PDW and T2W FS sequences (p < 0.05) while maintaining sharpness and artifact levels. SNR and CNR were significantly improved (p < 0.05), except for disc SNR in PDW (p = 0.189). Scan time was reduced by 49.2%.<br /> Conclusion DL-reconstructed TMJ MRI is diagnostically interchangeable with conventional MRI, offering improved image quality with a shorter scan time