Deep-learning-based contrast synthesis in knee magnetic resonance imaging: performance across multiple target contrasts

INTRODUCTION: Quantitative MRI has been held back, because it increases scan time on top of clinical protocols. In this cross-sectional study, we explore how accurately unmeasured contrasts can be synthesized from magnetic resonance fingerprinting (MRF) data. MATERIALS AND METHODS: MRF, T1-weighted (T1w), gadolinium-enhanced T1w fat-saturated (Gd T1w fs), and double-echo steady-state (DESS) data were acquired at 3 T from volunteers having knee joint symptoms (n=88). Ten subjects were excluded due to missing MRF-data. U-Nets were trained to synthesize the other contrasts out of the MRF data, and the image quality was evaluated for both the synthetic and conventional by two radiologists. Presence of synovitis was evaluated from Gd T1w fs images. Wilcoxon signed-rank test was utilized to test differences in the image quality scores. RESULTS: Synthetic images overall had only slightly lower image quality in radiologists' evaluation for T1w (3.54 vs 3.88, p<0.01), DESS (3.22 vs 3.90, p<0.01), and Gd T1w fs images (3.40 vs 3.76, p<0.01). Synovitis could not be reliably observed from synthetic Gd T1w fs images. CONCLUSIONS: The image quality gap observed earlier persisted in this study and manifested itself steadily across all the joint tissues. Synovitis assessment was not possible from the synthetic contrast-enhanced images.