Post-contrast 3D T1-weighted MRI is a time consuming component of cancer neuroimaing protocols. The goal of this study is to accelerate the acquisition of 3D MRI using deep learning reconstruction of undersampled k-space data beyond clinically-available acceleration techniques. A two-stage deep learning framework called Accelerated Deep Modular Image REconstruction (ADMIRE) was developed to first remove aliasing artifacts in the accelerated coronal orientation (unaliasing network) and subsequently enhance magnitude images in axial orientation (enhancement network). The two networks were trained independently using data from two 3D T1-weighted sequences on GE Healthcare scanners: BRAVO and MPRAGE (Ntotal = 136; NBRAVO = 47; NMPRAGE = 89). 3D MPRAGE k-space data with eightfold acceleration (4 × 2 ky-kz undersampling) were acquired on 15 patients with enhancing brain lesions and reconstructed with ADMIRE to test feasibility. Validation was performed quantitatively against data-driven and unrolled deep learning methods and against the clinical acquisition that uses a lower acceleration factor using a qualitative reader study performed by three radiologists. Composite and individual one-sided Wilcoxon signed-rank tests were utilized to assess noninferiority of the proposed deep learning approach. Quantitative and qualitative evaluation, scoring and statistical testing revealed that ADMIRE displays better image quality than both data-driven and unrolled methods and was noninferior compared to the clinical standard, despite 32%-46% reductions in acquisition time, in terms of overall quality (summing all metric scores and using a composite noninferiority margin of 8, p-value < 0.001) and diagnostic confidence specifically (using an individual noninferiority margin of 2, p-value = 0.004). ADMIRE further increases the current clinical acceleration of T1-weighted 3D MRI by enabling higher k-space undersampling factors. The use of a clinical sequence and the fast computation speed facilitate clinical translation of ADMIRE.
Mekhanik et al. (Wed,) studied this question.