Sub-5-Minute Non-Contrast MRA of the Thoracic Aorta using Deep Learning: Same Quality, Half the Time

Background: REACT (Relaxation-Enhanced Angiography without ContrasT) is a reliable non-contrast magnetic resonance angiography for imaging of the thoracic aorta but remains time-consuming. This study evaluates acceleration of image acquisition using compressed sensing and parallel imaging (Compressed SENSE, CS) combined with deep learning-based image reconstruction (CS-AI).

Methods: In this prospective single-center study, 40 volunteers underwent ECG- and navigator-triggered 3D REACT at 3T using CS acceleration factor 4 (CS4) and 8 (CS8). CS8 data were reconstructed with standard and CS-AI methods (CS8-AI). Two radiologists measured aortic diameters, rated subjective image quality and performed pairwise comparisons. Additionally, objective image quality metrics were calculated. Statistics included t-tests, non-parametric tests, and assessments of reliability and variance, with CS4 as the reference standard.

Results: Median scan time was reduced by 44% (CS4: 8:41 min; CS8/CS8-AI: 4:52 min). All techniques showed excellent agreement in aortic diameter measurements (mean differences < 0.2 mm; P>0.999). CS8-AI demonstrated reduced deviation from CS4 (P=0.003), and improved reproducibility compared to CS8 (interrater: P=0.024; intrarater: P=0.002). Measurement variance was significantly lower with CS8-AI than CS8 (interrater: P< 0.001; intrarater: P< 0.001), and comparable to CS4. CS8 showed significantly lower subjective image quality scores than CS4 (P< 0.001), while CS8-AI showed comparable or higher scores. Forced-choice comparisons favored CS4 over CS8 (P< 0.001), but no preference was observed between CS4 and CS8-AI (P>0.999). Objective metrics predominantly confirmed the subjective results.

Conclusion: Deep learning-based reconstruction enables the acquisition of REACT of the thoracic aorta in less than five minutes while preserving high image quality and maintaining excellent measurement reproducibility.

Example of a healthy 58-year-old male volunteer. Coronal source images of the three methods (the reference standard Compressed SENSE acceleration factor 4 (CS4) and acceleration factor 8 with (CS8-AI) and without (CS8) adaptive intelligence reconstruction) with axial inserts at the level of the origin of the right coronary artery (arrowhead). CS8 shows a higher image noise and poorer vessel delineation compared to CS4, which could be significantly improved in CS8-AI.
Bland-Altman plots depicting the agreement between the respective methods (CS4, CS8, CS8-AI; A) as well as the reproducibility within a method between two readers (B) and with repeated measurement of one reader (C). Each row contains one plot per measurement level with three data sets. The colored lines show the mean difference and the dotted lines the 95% limits of agreement, each averaged over all measurement levels (aortic annulus, aortic sinus, STJ, ascending aorta, and at the descending aorta) for better comparability. CS4 = Compressed SENSE acceleration factor 4; CS8 = Compressed SENSE acceleration factor 8; CS8-AI = Compressed SENSE acceleration factor 8 with adaptive intelligence reconstruction; STJ = sinotubular junction
Cumulative distribution of subjective scores for Compressed SENSE acceleration factor 4 (CS4) and acceleration factor 8 with (CS8-AI) and without (CS8) adaptive intelligence reconstruction in the categories image quality, image noise, edge sharpness and artifacts using a 5-point Likert scale.