Purpose: To investigate the impact of a vendor deep learning (DL) reconstruction image enhancement algorithm on basic image quality parameters for routine brain MRI.
Methods: All experiments were performed on a 3T GE Premier scanner using a head coil. The ISMRM phantom was imaged with three pulse sequences that are performed as part of a routine brain MRI protocol. The three pulse sequences included T1-weighted spin echo, T2-weighted spin echo, and fluid attenuated inversion recovery (FLAIR). Four images were acquired for each pulse sequence: (1) no filtering, (2) low strength DL reconstruction, (3) medium strength DL reconstruction, and (4) high strength DL reconstruction. Spatial resolution, signal-to-noise ratio (SNR), and image contrast were quantified and compared for all images. Image contrast was determined by characterizing 14 ROIs in each of the T1, T2, and proton density (PD) arrays of the ISMRM phantom. SNR was quantified by taking the ratio of the mean of a signal ROI measured within a uniform region of the phantom and the standard deviation of a noise ROI measured in a void region outside the phantom. Image artifacts were also assessed.
Results: DL reconstruction reduced Gibbs ringing artifacts for all reconstruction strength settings. DL reconstruction did not produce a significant difference in MRI signal of the T1, T2, and PD arrays for the T1, T2, or FLAIR acquisitions. SNR increased with increasing DL reconstruction strength. Spatial resolution also improved with DL reconstruction. However, DL reconstruction introduced artifacts along the frequency encoding direction for some high contrast objects.
Conclusion: DL reconstruction simultaneously improved SNR and spatial resolution, while also reducing Gibbs ringing artifacts. In-vivo experiments are necessary to determine if these findings hold in patient imaging.