3D and 4D Free-Breathing Abdominal T1-Weighted MRI in Clinical Practice Using Deep Learning Auto-Navigation and...

To develop and evaluate an automated clinical prototype for a 1-min free-breathing T1-weighted 3D MRI and a 2.25-min 4D MRI utilizing radial k-space acquisition and deep learning (DL) auto-navigation and reconstruction.

The clinical prototype was deployed on 3 T GE Healthcare scanners, using the GE DISCO-Star (radial golden-angle stack-of-stars) pulse sequence, DL auto-navigation (RANGR), DL reconstruction (Movienet), and vendor-specific image processing to efficiently generate DICOM files. The system automatically collects raw data, transmits the data to an external high-performance computer, performs image reconstruction, and generates DICOM files. A customized version of the Movienet network was trained, using compressed sensing references, to achieve 2.25-fold and 2-fold acquisition acceleration for 3D and 4D MRI, respectively. The prototype was evaluated on 50 patients (ages: 8-87) with TE = 1.46-1.6 ms, TR = 3.2-3.4 ms, flip angle = 12°, in-plane resolution = 1.17-1.64 mm, and slice thickness = 4 mm. Image quality was assessed qualitatively by three expert radiologists, who compared Movienet to conventional vendor methods, followed by a statistical analysis using Wilcoxon signed-rank tests.

The Movienet prototype demonstrated remarkable efficiency, requiring only 90 s of GPU and 4 min on a CPU computation for 3D reconstruction, while exhibiting better performance, characterized by reduced streaking artifacts and about one-point improvement in image quality on a five-point scale compared to vendor technology. For 4D reconstructions, reconstruction time increased by 20 s (GPU, +1 min CPU) while maintaining comparable quality metrics. For all metrics, the differences were statistically significant (all p-values < 0.0001).

The Movienet prototype significantly enhances motion robustness and acquisition speed in abdominal MRI, offering a transformative approach for clinical applications.