Clinical utility of deep learning motion correction for T1 weighted MPRAGE MR images

Purpose: To evaluate the clinical utility of the application of a deep learning motion correction technique on 3D MPRAGE magnetic resonance images acquired in routine clinical practice. Methods: An encoder-decoder deep learning network inspired by InceptionResnet was trained on public datasets. The clinical utility of the trained network was evaluated retrospectively on 27 3D MPRAGE T1 weighted motion degraded MR images identified by radiologists during reporting. The assessment of image quality was performed by one board-certified radiologist and one senior radiology trainee for nine neuroanatomical regions of the brain using a five-point visual grading scale. Results: The deep learning motion correction technique resulted in reduced ghosting, ringing and blurring for all the brain regions investigated. The larger regions of interest such as ventricles improved the least (1.81 to 1.16, p-value: < 0.0001) while the smaller but complex regions such as the hippocampus improved most (3.0 to 1.67, p-value: < 0.0001). The Wilcox rank tests of image quality differences for the nine neuroanatomical regions were all statistically significant (p < 0.001). Overall, 60 % of the neuroanatomical regions were improved, 39 % were unchanged and 1 % were degraded. Out of the unchanged cases, 28 % were already scored at the highest image quality before motion correction. It was found that approximately 13 % of repeated scans could be avoided using the DL motion correction approach. Conclusion: The deep learning motion correction technique improved the overall visual perception of the 3D T1 weighted MPRAGE brain images. This would improve the clinical utility of otherwise motion degraded images and allow visualisation of normal anatomy and even subtle pathology.